The marketing measurement industry has increasingly used a term: causal attribution. Attribution is the assignment of credit for an outcome to inputs. Causal is a claim about counterfactual reality. Causal attribution, as a phrase, joins them.

The industry has a long history of borrowing vocabulary from adjacent disciplines: scientific, statistical, financial. Some borrowings are merited. The field has extended real methodological work from causal inference, econometrics, and financial attribution into marketing measurement, in places. Other borrowings are unearned: terms adopted for their credibility without the underlying warrant. The four questions below are one way to tell them apart.

C3 Metrics does not use the term causal attribution. What follows is the four questions, C3’s answers to them, and a note on what the answers show.

The four questions.

The first three are attribution questions. The fourth is the causal question, and it is only meaningful if the first three have been answered.

1. At what unit of analysis is credit distributed?

2. What does the word attribution mean in the methodology’s use of it?

3. With what confidence interval, or with what disclosed assumptions in lieu of one?

4. If a causal claim is being made: under what identification strategy is that claim established, at what unit of analysis?

None of these is a gotcha. Any competent practitioner asks them of themselves before adopting a technical term. A methodology that answers all four consistently is doing what the marketing says it does. A methodology that answers three and stops at the fourth has answered the causal question in the same breath, by declining to make the claim.

The standard question three asks is disclosure. Assumptions are owed to the party relying on the number, in a form that party or its auditor can review. That is the bar. It is not that the framework must be public. It is not that the framework must be free. It is that the reliant party can review what the number rests on. Every provider is free to specify the framework. The standard is the disclosure to the reliant party.

C3 Metrics, held to the test.

C3’s four answers.

1. Unit of analysis: the individual conversion. Credit is distributed across touchpoints.

2. What attribution means, in C3’s use: the assignment of credit under a stated convention. Not a counterfactual claim.

3. Confidence interval and disclosed assumptions: the question offers two forks, and C3 takes the second, and says so plainly. No confidence interval is offered at the individual conversion, because no sampling frame exists at that unit. Disclosed assumptions stand in lieu. Those disclosures are documented in a proprietary framework, available to clients. The framework specifies the data sources used, the inferences drawn from them, and the assumptions behind those inferences. A C3 client, or an auditor authorized by the client, can review it. That is the form in which C3 meets the standard. Other providers can and do have other frameworks; the standard is not the framework but the disclosure to the reliant party in reviewable form.

4. If a causal claim were being made, at what unit? None. C3 does not claim causal identification at the individual conversion, and no counterfactual is available at that unit that would let us. We therefore do not use the word causal.

C3 answers all four. The answer to the fourth is what tells you why C3 does not use the term causal attribution.

The self-executing test.

The fourth question can only be checked against the first. If the unit in the fourth answer is coarser than the unit in the first — identification established at the geo, credit reported at the ad; identification at the channel, credit at the tactic; identification at the segment, credit at the conversion — the methodology owes a fifth disclosure: the rule by which the estimate is transported downward, and the name of that rule, which is allocation. The four questions do not force this comparison. But they enable it. That is what makes them a test.

The fifth disclosure is not exempt at this address, either. C3 distributes credit across touchpoints under a stated convention. That convention is an allocation rule, and we name it one.

This is a choice.

If C3 could use the term consistent with our own methodology, we would. It would be a stronger claim, a more impressive one, and the market might value it, at this moment. Choosing to use the term is a valid positioning choice. So is choosing not to. We have chosen not to, because we cannot use the term without knowing what it means, and without being able to test ourselves against that meaning. Until we can, we do not.

That is the same choice, in a different domain, as the distinction between measuring and confirming: to hold our language to what the underlying method can support, and to not adopt a modifier we cannot warrant.

The technical constraint.

Identification lives where the counterfactual was constructed. A randomized experiment across geos identifies at the geo. A media mix model calibrated by that experiment identifies at the channel. A conditional average treatment effect estimated for a stratum identifies at the stratum. Propagating the estimate downward, from geo to channel to ad, from stratum to individual, from population average to conversion, is an allocation rule, not an identification.

A weight can travel. A causal estimate cannot travel with it. Identification does not follow the weight down. Nothing transports it.

The frontier of causal inference over the last fifteen years has extended these tools to heterogeneous effects: conditional average treatment effects, causal forests, uplift models, and Pearl’s third rung of counterfactual reasoning. Each stratifies the population and yields an expected effect for a class. None, absent heroic assumptions the practitioner is obliged to disclose, delivers a causal effect for a specific input on a specific individual. The finest available estimate stops at the stratum. The stratum is not the person, and the person is where attribution asks its question.

Consider the smallest possible case. Two ads sit side by side, for exactly the same product, from the same company. Four out of five users choose the one on top. We can see the split. We cannot know why the fifth chose the one on the right, and no stratum, however finely cut, closes that gap. The field named this constraint before the industry existed: Holland called it the fundamental problem of causal inference, in 1986. The four questions are the test; this is the reason the test matters.

Shapley-value attribution, one common approach to principled attribution in marketing, decomposes a modeled outcome additively across features. This is a mathematically coherent operation. It is not, in its standard form, a causal claim about the underlying reality. Causal effects do not decompose additively in the presence of interaction, and interaction is the ordinary state of marketing exposures.

Causal Shapley variants exist. They respect an assumed causal graph and require the identifying assumptions to be stated explicitly. That specification does not work around the fourth question; it returns to it in different clothes: what is the identification strategy, and does the graph, treated as an assumption, survive review by someone who did not build it? Same test, same disclosure requirement.

The generous reading, granted in full.

A media mix model calibrated by properly designed geo experiments makes an explicit, defensible identification argument at its unit of analysis. Advisory committees composed of serious empiricists in causal inference review models, propose designs, and hold their principal investigators to a standard. This is real work, done at real institutions, and nothing in this piece diminishes it.

Advisory committees review models. They do not review headlines. Where the phrase causal attribution parts company with the discipline the four questions demand is at the transport: from the unit at which the model was identified, to the unit at which its outputs are marketed to end users. The failure is not in the mathematics. It is in the marketing.

The invitation.

The four questions are not proprietary. They are the questions any practitioner should ask themselves before adopting a technical term. C3 has answered them, publicly, above. Any vendor claiming to sell causal attribution can do the same. Point to the disclosed unit of analysis. Point to what attribution means in the methodology’s use. Point to the confidence interval or disclosed assumptions, in a form the reliant party can review. And, if a causal claim is being made, point to the identification strategy that establishes it at the reporting unit. If the reporting unit is finer than the unit of identification, disclose the transport rule and name it allocation.

If those answers are consistent, the phrase is doing what the marketing says. If they are not, the phrase is doing work the methodology does not authorize.

The difference is simple. Is this confirmation, or is it measurement? Only a full, consistent set of answers to the questions above can tell us. Absent that consistency, the phrase causal attribution is neither.

C3 does not use the term. The four questions are available to anyone who does.

Companion piece: Confirmation Is Not Measurement.

Brown’s example is a Lancet paper from last year claiming USAID prevented 91,839,663 deaths worldwide between 2001 and 2021. The number is precise. It is published in a journal that brags about selecting only the best research. The total decline in world deaths over those twenty years was only 79 million — meaning the paper claims USAID alone saved more lives than all causes combined, which implies everything else in global health backfired to the tune of killing thirteen million people. The number doesn’t pass the consistency check.

That check is Brown’s discipline. Take the number, find the assumption inside it, test the assumption against what it must be consistent with. If the assumption breaks the math, the number isn’t doing the work it claims. His prescription for researchers is pre-registration: state the hypothesis and method in advance, release the data and tests. His prescription for readers is skepticism: “don’t disbelieve things, be skeptical. That person believes it, maybe they have good arguments, but don’t be too sure.”

The same pattern runs through advertising measurement.

A vendor publishes an estimate — incremental revenue lift, return on ad spend, conversion attribution. The number is precise. It carries an implied claim about what would have happened if the media hadn’t run. The implied claim rests on assumptions about how the underlying system behaves: that geographies are exchangeable, that the test window is representative, that no unobserved factor confounds the comparison, that the effect generalizes. The buyer rarely strains the assumptions. The number propagates.

It helps to make a distinction the marketing tends to collapse.

Confirmation answers a question you have already decided to ask, in a controlled window, with a directional answer about whether the thing tested is working. The buyer asks “is this channel incremental?” and the methodology returns “yes” or “no” with a magnitude attached. Done well, it gives the buyer validation of a specific decision already framed.

Measurement answers a different question. It asks what is actually happening across the system, with bounded inference about magnitudes and uncertainties. The buyer doesn’t yet know which channels matter, in what combination, with what confidence. The methodology returns a picture of the whole, with the confidence intervals on each piece named explicitly. Done well, it gives the buyer understanding.

Both have legitimate use. They are different services dressed in similar vocabulary. The marketing collapses the distinction because confirmation is what most buyers think they want — a clear answer to a question already asked feels more useful than a calibrated picture of the system they’re trying to understand. The cost of the collapse is that the buyer comes away thinking they have measurement when what they have is confirmation.

Here the math gets pointed. Any methodology that produces a confidence interval is admitting it works in probability space, not in certainty space. The interval may be narrow — even very narrow — but it is not zero. A claim presented without a confidence interval is not a claim free of one; it is an implicit assertion that the interval is zero. The methodology requires this assumption to function. It is honest only if you believe it.

Causal inference, as a discipline, is meticulous about distinguishing association from causation, naming the assumptions required to claim either, and bounding the inferential uncertainty around its estimates. A methodology that presents a central estimate as ground truth is using rigorous vocabulary in a way the discipline itself wouldn’t recognize. The math forbids posterior certainty in stochastic systems, and advertising is a stochastic system: the same exposure produces different outcomes across users, contexts, and time. A methodology that returns certainty about a probabilistic system is not measurement. It is confirmation that the buyer’s prior belief was correct.

Which brings us back to Brown. His prescription is the discipline any honest measurement methodology already practices. State what you expect to be true. State the conditions under which you would consider yourself wrong. Publish the test that would distinguish the two. Let the test run. Accept the result. The methodology that skips these steps is the methodology that confirms its own belief.

A number that confirms its own certainty is almost certainly wrong, because the certainty is the input, not the output. That is Brown’s conclusion in Wrong Number. It applies to the Lancet’s USAID figure. It applies to published research that propagates as fact. It applies to the measurement claims a buyer evaluates when deciding which vendor to work with.

Brown’s counsel: be skeptical, but not nihilistic. Measurement claims deserve scrutiny without dismissing measurement itself.

A belief can be confirmed or changed. A system can only be believed if it knows it cannot be 100% certain of itself. The methodology that can be wrong is the methodology you can trust. The methodology that can’t be wrong is the one that confirms its own belief.

The buyer who asks “how would you know if you’re wrong” is asking the question the methodology should already have asked itself.

—

Peter Coy’s interview with Aaron Brown. “Wrong Number” is available from major retailers.

Three numbers landed in the AI-and-marketing newsletter cycle this past week. Each from a different source. Each pointing at the same gap.

Gartner’s 2026 CMO Spend Survey: CMOs are now allocating 15.3% of marketing budgets to AI. Only 30% of organizations report mature readiness to deploy at scale. The gap between spending on AI and winning with AI is widening. The AI-ready 30% are also the ones investing more, 21.3% of budget against the 15.3% average. They are operating from a higher overall marketing investment base, 8.9% of revenue against the 7.8% average. The pattern is consistent. The capable are pulling ahead.

Canva’s third annual State of Marketing and AI report: 97% of marketing leaders use AI in everyday creative work; 99% plan to increase AI investment in 2026. Consumer reception runs the other direction. 70% of consumers say AI-generated ads are missing their soul. 78% say they would rather see human-made ads even if AI could produce something better. 87% believe the best advertising still requires a human touch.

HubSpot’s Spring Spotlight launch of AEO, Answer Engine Optimization: organic traffic for HubSpot customers is down 27% year-over-year as users move from search results to AI answers. The product reframes the marketing surface entirely. Instead of optimizing for a Google position, marketers now optimize for how AI models characterize and cite their brands.

Three independent reads, all pointing the same way. Marketers are spending on AI faster than their organizations are getting ready to know what AI is actually doing, faster than they are protecting consumer trust as they lean into it, faster than they are tracking where the attention is moving.

The Gartner number is the cleanest articulation. 15.3% of marketing budget, 30% mature. Two thirds of CMOs are spending real money on capabilities their organizations cannot yet deploy at scale.

First, identify the gap.

If AI is going to work, you have to decide what you want it to do. For AI in marketing, that means a measurement layer that can tell you, channel by channel, what AI is contributing, where the consumer-trust cost is showing up, and which AI surfaces the traffic and attention are moving toward. Independent measurement is where readiness meets the spending decision, and it provides an unbiased signal, a true tare weight for future refinements.

My company and I have been writing about this since AI advertising surfaces started landing at scale. The AI-in-Advertising Readiness Checklist asks five testable questions:

• Is AI a separable channel in your model?

• Is the measurement window matched to the journey length?

• Is the input layer disciplined?

• Are the outputs calibrated and specific?

• Are AI-referred outcomes observed independently of platform self-reports?

The framework is the readiness picture in question form. Most stacks are less-prepared than the spending acceleration assumes.

The data

Earlier this year, the IAB published its State of Data 2026 report. The headline finding: 75% of buy-side marketers say their core measurement approaches — attribution, incrementality testing, marketing mix modeling — aren’t delivering the speed, accuracy, or trust they need. Three out of four. The industry’s own trade body, surveying the people who actually run the budgets, found that the systems they’re running don’t work.

The headline traveled. What didn’t travel as cleanly is the IAB’s prescription: AI. The report projects $26 billion in unlocked media value from AI-driven measurement, recommends always-on incrementality, faster model updates, AI-democratized multi-touch attribution.

The industry has responded in kind. Publicis acquired LiveRamp for $2.2 billion in May — vertically integrating identity resolution into the buying stack, embedding the measurement layer inside the system that buys the media, and structurally precluding the kind of independent validation that would surface the failure mode. The agentic-AI story that surrounded the deal was the wrong thing to focus on. The structural move was closing the loop — making the asserted confidence operational, not just claimed.

The response to measurement is broken is, at scale, more confident measurement. Better models. Faster cycles. More signal. That’s the wrong instruction.

Current measurement systems assert confidence about claims they cannot support. Adding AI to systems that already over-assert accelerates the failure. The systems reach wrong answers faster, wrap them in more authority, and embed them deeper in decision-making. Speed and certainty applied to a structurally flawed methodology compound the flaw.

The question

When a measurement vendor tells you they can identify what drove a conversion, attribute revenue to a media chain, or measure the incremental impact of a channel — ask this:

How would you know if you’re wrong?

Watch what happens.

The honest measurement vendor has an answer. They’ll talk about confidence intervals, holdout testing, where their model fails, when they refuse to make a claim. They’ll describe the assumptions baked into their methodology and tell you where those assumptions might break. They might disagree with you about something. But they can engage with the question.

The vendor selling certainty cannot engage with it. The question doesn’t fit inside the pitch. They’ll redirect — to features, to accuracy numbers, to client logos. They’ll reframe — let me tell you about our methodology — without actually answering. That moment, when the question doesn’t compute, is the most important data point you will get from any measurement vendor evaluation.

Why the question dissolves the pitch

Marketing causality cannot be inferred from observational data. Even controlled experiments — randomized tests, holdout studies, lift tests — don’t prove causality in the strict sense; they measure differences that we then interpret as causal. Every observational measurement methodology — multi-touch attribution, media mix modeling, identity-resolved attribution — sits at greater remove still. They all infer. None confirm.

A vendor that understands this builds it into the methodology. They produce ranges. They surface assumptions. They tell you when the model can’t say anything useful. They treat we could be wrong as the starting condition of every measurement output.

A vendor that doesn’t — or, more commonly, that understands it perfectly well but cannot sell it — builds a system that produces confident outputs regardless of underlying uncertainty. The system’s job becomes generating a number. The number’s defensibility becomes secondary to its existence. Eventually, every methodology inside the system is selected for its ability to produce numbers, and the parts that produce humility are stripped out as commercially inconvenient. The output is asserted confidence — certainty performed where certainty cannot exist.

The commercial pressure on the category selects for systems that cannot answer the procurement question. The selection pressure does the work without requiring individual vendor dishonesty. The systems are closed — input to model to output to dashboard, with no validation layer anywhere in the chain. Closed systems guarantee confident wrong answers, given enough time and enough inputs.

What the 75% tells us, and what it hides

Three out of four marketers know their measurement is broken. That is the published finding. The harder problem sits in the other quarter.

The marketers in the 75% run their decisions with awareness that the dashboard cannot be fully trusted. They hedge with judgment, treat outputs as one input rather than the input, and watch for the gap between what the model says and what the business shows. Their measurement is wasteful — billions of dollars allocated against data they do not fully believe — but the wastefulness is bounded by their own skepticism.

The interesting question is what is happening inside the 25% who say their measurement is working. Almost none of them are operating with the kind of methodological discipline that produces approximately right answers — that level of rigor is rare in any field, rarer in marketing measurement, rarest in the AI-driven measurement programs sold over the past three years. The vendor’s pitch tells them their measurement works; the dashboard tells them their measurement works; the AI layer tells them their measurement works; every layer in the system reassures them that the system is functioning. The wrongness compounds at the speed of the system, which gets faster every quarter.

The structural condition that produces this outcome lives in the architecture of what they have bought. They have no validation layer. Every reassuring signal originates inside the same closed loop that produced the original measurement. The asserted confidence is total because nothing inside the system contradicts it.

The procurement question — how would you know if you’re wrong — is the test that distinguishes the systems with a validation layer from the systems without one. It is also the test that the 75% can use on the next vendor, before they replace a broken system with a more confidently-broken one.

The terminology that should make you pause

Certain phrases in measurement marketing are doing more work than they appear to. They smuggle asserted confidence into claims the underlying methodology cannot deliver.

Deterministic attribution. Deterministic means certain. No observational measurement methodology is deterministic. The word is borrowed from physics and engineering, where it has a precise meaning, and deployed in marketing measurement where it does not. If a vendor describes their attribution as deterministic, ask what they mean.

Causal AI. Causality cannot be inferred from data alone. It requires assumptions about how the world works, expressed as a model, that the data then either supports or refutes. AI doesn’t change this. A system can be very sophisticated about pattern recognition without being any closer to causal inference.

Identity-driven measurement. Identity resolution is a real technical capability with real value in specific contexts. But the phrase implies that more identity matching produces more accurate attribution. The methodology bounds the accuracy of measurement; the quantity of identity-resolved signals does not change that bound. A closed system fed more inputs produces more confident wrong answers.

Real-time causal measurement. Speed is not virtue when the direction is wrong. Real-time confidently wrong measurement is worse than monthly approximately right measurement, because it produces more decisions per unit time.

None of these phrases is automatically dishonest. Each can be used responsibly by a vendor who knows what they mean and is precise about the scope of the claim. Each is also the first place to test the procurement question. The vendor who can answer how would you know if you’re wrong can also explain exactly what they mean by deterministic, causal, identity-driven, or real-time. The vendor who can’t will use the phrases as if they were self-evident.

A second-order test

There’s a complementary question worth asking when a vendor invokes a partner or customer authority. Where is the corroboration from the people you are invoking?

When a vendor announces a partnership with a major platform, ask whether the partner has a reciprocal announcement that uses the same language. When a vendor names a flagship customer, ask whether the customer’s quote endorses the outcome of the work, or just the visibility into it. When a vendor claims a special technical relationship — log-level access, exclusive integration, first-party data sharing — ask to see the partner’s documentation of that relationship.

The pattern in the wild is that the strongest version of a claim lives in the vendor’s press release, the weakest version lives wherever a real person has to put their name on it, and the partner’s own language is silent on the strong version entirely. That gradient is itself evidence. When the corroboration runs the other direction from the claim, the claim is unilateral marketing language, not a substantiated relationship.

Both questions deflate the same architecture. How would you know if you’re wrong tests the methodology. Where is the corroboration tests the surrounding claims.

A note on what kind of argument this is

The argument here is not a causal claim. It does not prove that AI-driven measurement produces confident wrong answers, that vendors who cannot answer the procurement question will eventually deliver bad outputs, or that the 25% who report their measurement is working are mostly mistaken. None of those claims is provable in the strict sense.

What the argument does instead is reason from observed evidence — the IAB findings, the structural properties of closed measurement systems, the patterns in vendor marketing language, the architecture of walled-garden data access — to inferred conclusions. The conclusions are stronger for being inferred rather than asserted. They do not presume something the underlying data structure cannot support, and they carry the possibility of being wrong inside them. A reader who finds evidence that runs the other way can update against this argument. That is how careful argument is supposed to work.

This is also the methodological posture the piece is recommending for measurement vendors. The vendors worth trusting reason from data to inferred conclusions, with assumptions surfaced and failure modes named. The vendors who assert causal claims from observational data and refuse to acknowledge the gap are operating in asserted confidence — the same failure mode the piece has been describing.

Inferred results, in this domain, are stronger than causal ones. They do not presume the impossible, and they include their own fallibility. Causal claims about marketing outcomes from observational data presume what the data cannot deliver, and the systems that produce them are structurally incapable of telling you when they are wrong.

The piece itself operates by the standard it is recommending. That is deliberate.

The procurement frame

Every vendor evaluation should include both questions — in writing, in the RFP, in the procurement-team checklist, as an honest filter. The vendors who can answer them are worth working with. The vendors who cannot will eventually deliver confidently wrong answers at scale, and you will be the one accountable for the decisions you made off their dashboards.

There are vendors who can answer. They are not always the largest, the loudest, or the best-funded. They are the ones who treat the methodology as the product, who carry epistemic humility into the room, who would rather lose a deal than oversell the work. The market doesn’t reward them in the short run, because confident certainty wins more deals than honest uncertainty. But the deals they do win compound. Their customers stay. Their methodology survives the disillusionment cycle that will eventually come for the confidently wrong half of the category.

The question costs nothing to ask. It surfaces more signal than any features comparison, any case study, any client logo wall. It is the simplest procurement test in the category, and almost no one runs it.

Ask it.

Note the direction of the data flow. The data goes into Meta’s optimization. It doesn’t come out as measurement. Meta is being honest about what the product does.

Triple Whale, named as one of the launch partners, describes their side of the integration plainly on their own help-center page. Their Sonar product “enriches your first-party data and sends it back to Meta through Facebook CAPI, creating a smarter feedback loop for targeting, attribution, and ROAS.” That description is accurate. Your first-party data goes to Meta. Meta uses it to optimize delivery. The optimization produces conversions. Your model credits the conversions. Your model feeds Meta. The loop holds steady.

This is not a Meta story

The first read on Custom Attribution is that Meta has done something new. That read is wrong. CAPI passback isn’t novel; it’s the latest instance of a pattern that’s now visible across all three major ad platforms simultaneously.

Google moved through 2025 and into 2026 consolidating its ingestion paths into the Data Manager API. The older UploadClickConversions endpoint deprecates June 15. Enhanced Conversions for web and Enhanced Conversions for leads collapsed into a single setting in April. Google’s framing of the destination state is direct: the API “gives advertisers direct control over which conversions Google learns from and optimizes toward.”

Amazon expanded Amazon Marketing Cloud eligibility to sponsored-ads advertisers, ran new partner-certification programs through the Amazon Ads Academy, and rolled out expanded measurement capabilities for AMC’s certified-partner ecosystem. Measurement runs inside Amazon’s clean room. Certified partners operate inside Amazon’s environment. The data stays where the platform can see it.

The three platforms describe the products in different vocabulary — Custom Attribution at Meta, Data Manager API and Enhanced Conversions at Google, AMC and the certified-partner network at Amazon. The structural mechanic underneath is identical:

1. The advertiser shares first-party data with the platform.

2. The platform uses the data to optimize its own delivery.

3. A certified measurement partner produces attribution that’s consistent with the platform’s claims.

4. The attribution feeds back into the platform’s optimization.

5. The loop runs.

Three different brands, one move. The category isn’t “Meta is doing this thing.” The category is “this is what platform-aligned measurement looks like in 2026.”

And the cost on the buyer’s side is real. Each of the three platforms justifies the integration evolution partly through “improved match rates” — and the match-rate improvement, in every case, comes from the buyer providing more identifying data to the platform than the prior integration shape required. Meta’s CAPI passes hashed user-provided data (email, phone, name) alongside the conversion event. Google’s Enhanced Conversions does the same. Amazon’s Clean Room positioning explicitly names IP address and hashed PII (email, name, phone) as the identifiers the new integration supports. The platform’s measurement improvement is paid for in the buyer’s data surface. Whatever else the trajectory is, it’s a steady expansion of the identifying data the platform receives from the buyer — and a steady contraction of the integration shapes (S2S, server-to-server with no PII exchange) that don’t require it.

The historical arc

The pattern didn’t appear in May. It’s been compounding for seven years.

2018 — Conversions APIs arrive as the response to iOS 14’s signal loss. Server-to-server replaces pixel where pixel is failing. The framing is data resiliency.

2021–2022 — Data clean rooms become the consolidation story. Amazon Marketing Cloud expands; Google Ads Data Hub; Meta Advanced Analytics. Each platform builds a privacy-safe environment where partners can operate, but only inside platform terms. The framing is privacy.

2022–2023 — Modeled conversions become the default. Platforms infer conversions where signal is missing, using their own models, with no external audit. The framing is accuracy.

2024–2025 — Certified partner programs replace open APIs as the preferred integration shape. To get the “good” data flow, you have to be inside the platform’s certification. The framing is quality control.

2026 — Custom Attribution, Data Manager API, expanded AMC measurement. Passback becomes the dominant integration pattern, and the platforms choose which measurement partners participate. The framing is AI readiness.

Each step looked, in its moment, like a technical accommodation. In aggregate, they describe a destination: the advertiser’s data flows into the platform’s optimization, the platform’s optimization defines the conversion universe, the platform’s certified partners measure inside the platform’s environment, and the entire stack is self-consistent. The measurement doesn’t grade the platform; it calibrates the platform.

The product is real, for the right buyer

None of this means CAPI passback is bad. It means it’s an optimization product, and it’s well-suited to a specific kind of advertiser.

If your business is single-channel-dominant — most spend in Meta, most journey inside Meta’s ecosystem, most decisions about how to grow involve optimizing within Meta — passback is a sharper Meta optimization layer. The numbers will stay internally consistent. The campaigns will optimize toward what you’re measuring. If you’re willing to believe the platform’s feedback, and willing to provide the platform your data, you get a smarter version of what you were already doing.

That’s a real buyer. Social-driven D2C is the archetype: brands where Meta represents the majority of spend, the customer acquisition path is genuinely Meta-mediated, the conversion event is close in time to the ad exposure, and the finance team is comfortable using Meta’s reported numbers as the source of truth.

The same logic applies on the other platforms. If you’re an Amazon seller whose whole business is Amazon, AMC + certified-partner measurement is fit-for-purpose. If you’re a Google-search-dominant lead-gen business, Data Manager API + Enhanced Conversions does the job. The category these products belong to is platform optimization, and platform optimization is a legitimate product when the buyer’s business shape matches the platform’s terms.

Where this stops being measurement

The category problem isn’t with the product. It’s with what the product is called.

Measurement, as a discipline, is the test of whether the platform’s contribution is what the platform says it is. That test requires a view that’s independent of the platform’s optimization — across channels, across windows, across the funnel between exposure and outcome, against a source of truth that doesn’t live inside the platform’s ecosystem.

The moment your attribution model is feeding the platform’s optimization, your model can’t perform that test anymore. The model and the platform are now tuning each other. They’re not grading each other. The loop produces internal consistency, not external truth.

Three specific failure modes are visible the moment you look for them.

Cross-channel comparisons stop working. The Meta-passback model credits Meta against Meta’s optimization. The Google-passback model credits Google against Google’s optimization. The Amazon-passback model credits Amazon against Amazon’s optimization. Each is self-consistent. None of them are comparable, because none of them sit at the same edge. The buyer with three passback feeds has three internally coherent attributions and no way to reconcile them against each other.

Finance reconciliation gets harder, not easier. The historical “why don’t your numbers match Shopify” argument is the visible symptom of the underlying problem: platforms credit conversions they didn’t drive. Passback doesn’t solve the discrepancy; it dresses it up. Meta’s numbers and your model’s numbers will agree, because they’re tuned to. The numbers still won’t match Shopify, because Shopify isn’t in the loop. The advertiser now has one fewer surface that can name the discrepancy.

The long funnel disappears. Multi-stage conversion paths — the shape that financial services, education, healthcare, and B2B businesses actually have — require attribution that holds across weeks of lag between exposure and outcome. Passback is structurally biased toward conversion events close in time to the ad. The longer the funnel, the worse the fit. Passback is built for single-event conversions; it’s not built for the journeys most non-D2C businesses actually run.

The structural test

There’s a clean test for whether what you’re buying is measurement or optimization. It’s not about vendor selection or methodology choice. It’s about data flow.

Ask one question: Does the model that grades the platform sit outside the platform?

If the model receives no data from the platform other than what every advertiser sees, and produces attribution that the platform can’t adjust, you have measurement. The model can grade the platform because the platform can’t influence the model.

If the model is inside the platform’s environment, certified by the platform, fed by the platform, and producing attribution that the platform optimizes against — you have optimization. There’s nothing wrong with optimization. It’s just not measurement, and treating it as measurement removes the only check the buyer has against the platform’s own claims.

The five-criteria framework C3 published in the AI in Advertising Readiness Checklist names this as Criterion 5: independent cross-source observation. Passback fails Criterion 5 by definition. The vendor on the platform’s certified list cannot, structurally, be the independent cross-source. Whatever else the vendor does, that’s the one thing they gave up by being certified.

Why the category confusion is the actual story

Meta isn’t trying to deceive anyone. Triple Whale isn’t either. The product descriptions are accurate when you read them carefully. Custom Attribution is “an optimization model.” Sonar “creates a smarter feedback loop for targeting, attribution, and ROAS.” Both descriptions are honest about what the products do.

The deception, where it happens, is in the secondary commentary that calls these products “measurement.” When an industry voice frames Meta’s selection of certified partners as “the closest thing to category endorsement that exists in this space,” they’re conflating two products. The vendor inside the platform isn’t an endorsed measurement vendor; they’re a selected optimization partner. The category confusion benefits the platform — because the platform now gets its optimization layer described as measurement, and the buyer loses the language to ask the question that would surface the distinction.

The buyer’s defense isn’t to refuse to use passback. The defense is to know which product they’re buying, and to keep at least one structurally independent measurement view in their stack for the questions passback can’t answer. Is the platform worth what we’re paying, against finance? Across the multi-channel funnel? Against the actual business outcome that lives beyond the ad event? Those questions don’t get answered inside a feedback loop with the platform. They get answered by something outside the loop.

That’s the only category distinction that matters in 2026. Everything downstream of it — vendor selection, methodology choice, contract terms, evaluation criteria — falls out of getting that one distinction right.

Note on sources: Meta’s framing of Custom Attribution is from the Performance Marketing Summit launch materials, May 14, 2026. Triple Whale’s product description is from their Sonar × Meta integration help-center page. Google’s Data Manager API positioning is from 2025–2026 Google Ads developer documentation. Amazon Marketing Cloud expansion and certified-partner framing is from Amazon Ads Academy and AMC release notes through Q1 2026.

The morning AdExchanger published its story about The Trade Desk’s UID2 detection failure, I was already in Bing Webmaster Tools, pulling the AI Performance Report for c3metrics.com. We’ve been baselining our own AI search visibility since Bing released the report in February — 132 citations to the site so far, 81% of them to the homepage. Steady, modest, useful as a sanity check on how AI surfaces are reading what we publish.

The AdExchanger piece reported on a major CTV publisher who ran a broken UID2 implementation for three months without anyone catching it. The Trade Desk’s account reps gave positive feedback throughout. When the publisher independently discovered and fixed the error, revenue didn’t change in either direction. TTD confirmed that the information available to the platform fell short of what would have been required to catch the problem.

What got me was the structure of the failure, not the specifics. AdExchanger was describing — on the sell side — a version of something we’d documented on the buy side a couple of months back.

Earlier this year we ran a programmatic traffic quality audit on a $20M-per-month flight. One of the findings: TTD line items where viewthrough beacon counts exceeded impression counts. A physical impossibility, by definition. A beacon fires when an ad is served. When the beacon count sits above the impression count, the delivery data has a structural integrity problem.

The platform reporting passed those line items as delivered. The agency reporting passed them as delivered. The finding emerged through independent signal-level reconciliation, run from outside the buying ecosystem.

Same structural failure, two different sides of the buy.

The mechanism in each case is different. The publisher’s UID2 tokens were broken at the encryption layer; our finding was about beacon inflation at the delivery layer. The structure underneath is identical. In both cases, the party best positioned to detect the problem held a financial relationship to the impression in question. In both cases, the problem surfaced through measurement that sat outside that financial relationship.

The CTV publisher in the AdExchanger story put the operating reality plainly: “you can’t audit and you can’t track it.” That captures the practical condition of most programmatic delivery data. The technical capacity to audit exists. The willingness to publish findings sits with parties whose revenue is connected to the inventory under audit.

What’s worth noticing: the pattern points to incentive architecture rather than software capability. Both TTD and the agency in our audit have substantial engineering investment in their reporting infrastructure. Reliable auditing of data running through a platform’s own pipes asks the platform to flag findings that reduce confidence in its own inventory — a request that runs against the financial logic of the platform’s business.

Independent reconciliation is the mechanism that closes the gap. It lives at the level of posture more than methodology or technology. Measurement run by a party with no financial relationship to the impression produces a different number than measurement run by the party that priced, delivered, or managed it. That difference is where the actionable findings live.

We documented our buy-side finding more formally a couple of months ago. The new piece, The Audit Nobody Did (https://www.c3metrics.com/dl-uid2-accountability.html), ties the AdExchanger story and our finding together with the full structural argument. The original buy-side methodology lives in What Programmatic Fraud Actually Costs (https://www.c3metrics.com/dl-programmatic-fraud.html) for the deeper version.

A bit of meta: I find it useful when something happens in the industry on the same day I’m looking at a parallel case in our own data. Coincidences like that tend to be structural — they happen because the underlying conditions are the same on both sides, and the conditions show up wherever someone is looking. Independent measurement is the practice of looking from outside the financial chain. The harder you look, the more you find.

That’s the asymmetric knowledge, and it accumulates.

— Greg

Both numbers describe offline match rates. They describe different methodologies.

The 80% figure comes from online-to-CRM matching — connecting an observed device to a customer record in a CRM file. That works, within its limits. It tells you that someone who touched your ad later appeared in your customer database. It does not tell you which channels produced the conversion.

The 80% figure also comes from platform offline conversion imports — Meta, Google, TikTok ingesting a CRM file and attributing offline events to their own ads, through their own user graphs, using their own attribution windows. That also works, within its limits. It tells you what each platform claims it earned. It does not tell you what actually happened across the full media mix.

The 4% figure describes something different. It describes the rate at which an offline event — a phone-bound policy, a dealer visit, a branch account opening, a pharmacy pickup — can be connected to a fully attributed, multi-channel digital journey, through identity infrastructure no advertising platform owns. It describes the rate at which a real-world outcome can be individually tied to the digital touchpoints that actually produced it.

Same name, different methodology. A program that calls them both “match rates” without distinguishing which one it means produces attribution numbers that cannot be defended when the question gets asked — because they rest on a definition that was never made explicit.

This is the starting point of the conversion architecture question.

Every program has a conversion architecture. Most were never designed.

A conversion architecture is the set of decisions about which conversion signals enter the attribution model, in what form, at what confidence level, and how overlap between them gets resolved. These decisions determine what the attribution output actually means.

Most programs inherit their architecture by default. Digital completions come from the analytics platform. Offline conversions come from a CRM file upload. Platform conversion data comes from the reporting APIs. Everything goes into the model. The architecture was never deliberately constructed — it accumulated.

The consequence is a model that combines signals with very different levels of certainty and treats them as equivalent. A methodology problem that looks like a data output. And it produces numbers that look precise but reflect choices no one explicitly made.

Three types of conversion data. Three different certainty levels.

A well-constructed conversion architecture handles each type deliberately.

Digital online conversions are the strongest signal. Directly observed at the session level, tied to individual journeys through first-party and authenticated data mechanisms, deterministic in their connection to specific touchpoints. These form the foundation of any attribution model. A program that reports only what it can prove from digital signals is incomplete — but what it reports is real.

Independently matched offline conversions carry a structurally weaker signal. The match rate for true independent offline attribution — connecting an individual offline event to an individually attributed multi-channel digital journey, independent of the platforms being measured — runs 4% to 20% under real-world conditions. The constraints are physical: households share IP addresses and devices across networks; point-of-sale transactions carry no digital identifier; approximately 6% of US households have no digital identity to match at all.

Offline conversions should still be included in the model, at the right confidence level — matched where matching is possible, disclosed where it is not, and never inflated through probabilistic inference that manufactures a higher-looking rate. The consumers whose offline conversions cannot be matched to a digital journey appear in the model as conversions without an attributed digital journey. That accurately describes reality. Forcing a match produces a larger number at the cost of measurement quality.

The marketed “80% offline match rate” from many vendors describes either online-to-CRM matching (40-70%, connecting device IDs to customer records) or platform offline conversion imports (40-50%, the platform attributing offline events to its own ads via its own user graph). Both are real capabilities. They measure different things than independent offline attribution. Ask which one the vendor means.

Platform self-reported conversions are variable in quality and require their own handling. Meta’s default attribution window claims conversions within 7 days of a click. Google’s claims conversions within 30 days. A consumer who saw a Meta ad on Tuesday and a Google ad on Wednesday and converted on Friday is claimed by both. Neither platform applies the other’s attribution logic. Neither has visibility into the full journey. Deduplication — assigning each conversion event once, across all platforms, through a single attribution pass — has to happen before platform-reported conversions enter the model. Summing platform-reported conversions without deduplication inflates total conversion counts in ways that distort every channel efficiency calculation that follows.

The proxy insight — and why the rigorous model shows the difference

In many programs, a digital conversion serves as a strong proxy for a downstream offline conversion.

A completed online insurance quote application is a strong leading signal for a phone-completed policy bind. An online dealer inquiry — configured appointment, trade-in submission, financing request — is a strong signal for a vehicle purchase. A completed online loan pre-qualification is a strong signal for a branch-executed account opening.

When the independent offline match rate for the downstream event is low — 6%, 10% — the digital proxy often provides more complete and more reliable coverage of the same underlying consumer behavior. It captures the intent signal with higher confidence, without the noise introduced by probabilistic identity inference.

The analytically rigorous approach runs both models and shows the difference.

First model: built on digital signals and digital proxies for offline conversions. Highest confidence, most complete coverage, fully deterministic. Second model: extended to include the directly matched offline conversions at their actual match rate. The difference between the two models is itself informative.

Where the delta is small — where the digital proxy captures most of what the offline match adds — the proxy model is adequate for most optimization decisions. Where the delta is large — where the matched offline conversions show a materially different channel attribution picture than the digital signals alone — the offline matching reveals something the digital signal missed. That kind of finding changes spend allocation.

Showing both models, and explaining the difference, is what analytically honest measurement looks like.

The answer varies. One approach is always wrong.

The conversion architecture question has different answers in every industry, and for every program within an industry.

Insurance programs have agent-bound policy applications, direct-mail-triggered phone calls, and quote requests that complete over multiple sessions across days or weeks. Automotive has dealer visits, F&I transactions, and trade-in events that are identifiable in some programs and invisible in others. Financial services has branch openings, advisory relationship initiations, and loan closings — each with different data availability. Pharma DTC has prescription fills, patient enrollment, specialty pharmacy transactions, and care navigator interactions. Retail has POS transactions, loyalty redemptions, and wholesale replenishment events.

Each requires a different matching mechanism, a different confidence level, and a different deduplication approach. A program that applies a single conversion methodology across all of them — regardless of what the conversions actually are or how they complete — bypasses the architecture problem rather than solving it.

The only wrong answer is a universal one.

What the methodology disclosure makes possible

A conversion architecture that is documented — where each conversion type is labeled, its confidence level stated, its match rate recorded, its proxy relationship to digital signals explained — produces numbers the advertiser can interrogate.

The attribution output carries the methodology with it. When a CMO asks “how confident are you in these TV attribution numbers?” the answer sits in the record: here is the conversion type, here is how it was matched, here is the match rate, here is how it compares to the digital proxy model, here is what changes when you include it.

That stands up as a different kind of conversation than presenting a multi-touch rate and hoping no one asks how it was derived. It also stands up as the only conversation that holds when a sophisticated buyer, a CFO, or an auditor looks closely.

The measurement program that can explain its conversion architecture is the measurement program that means something.

Follow the Spend is a newsletter about marketing measurement — what it gets right, what it gets wrong, and what it takes to close the gap. Published by Greg Collins, CEO at

C3 Metrics.

The full methodology piece on match rates and conversion architecture → What omni-channel measurement actually requires →

There is a question worth asking any measurement vendor before you look at a single number they produce: how do you know?

Not what the number is. How they know.

The answer to that question is the methodology. And the methodology is what determines whether the number means something — or whether it is a well-constructed approximation presented as a measurement.

Start with deduplication — because it holds before anything else is resolved

Every platform in your media plan claims every conversion it touched. Google claims the conversion. Meta claims it. CarGurus claims it. AutoTrader claims it. Each platform applies its own attribution window, its own conversion logic, and reports against its own user graph with no visibility into what the other platforms reported.

Sum those reports. The total is a fiction — the same purchase, the same form submission, the same policy application, counted four times over.

Multi-touch attribution assigns each conversion exactly once. One consumer, one conversion, fractional credit distributed across the touchpoints that contributed. Before a single multi-touch journey is reconstructed, before a single upstream exposure is attributed, the math is already more honest than platform self-reporting.

This is the deduplication advantage. It does not require resolving the walled garden problem, the mobile sampling problem, or the offline matching problem to hold. It holds structurally, on every program, on every flight.

In automotive, where CarGurus, AutoTrader, and Google Search all claim attribution for the same vehicle sale as a matter of standard operating procedure, the gap between the platform total and the deduplicated MTA count is typically 50-65%. In insurance, where EverQuote, QuoteWizard, and the carrier’s own platforms claim the same bind, the pattern is similar. The overlap between endemic platforms and major platforms is widely known. What changes when you measure it independently is that you see the size of it, consistently, rather than encountering it as an occasional discrepancy that gets rationalized away.

The multi-touch rate: floor, not ceiling

After deduplication, the attribution model assigns credit to the journeys it can observe. And here the honest answer is harder.

When C3 reports a multi-touch journey, it means we have demonstrated — on an individually attributed basis — that a specific consumer was exposed to more than one media touchpoint before converting. We have the data to show it. When we cannot demonstrate that, we report a single-touch journey. Not because we assume the upstream exposure did not happen. Because we cannot prove it did.

The multi-touch rate in any C3 program is a floor. The actual rate in the real world is almost certainly higher. We do not report what we cannot prove.

This distinction matters because the alternative — and it is a widely practiced alternative — is to assume the journeys. Probabilistic models infer likely upstream exposure based on similar audiences and campaign timing. Panel-based augmentation uses aggregate behavioral patterns to estimate what the full population probably did. Definitional expansion classifies all conversions in a multi-channel program as multi-touch by definition.

None of these methods are necessarily dishonest. Some are useful for directional planning. But they are not the same as demonstrating that a specific consumer had a specific multi-touch journey — and presenting them as equivalent is where the market misleads.

Why the ceiling on provable journeys is structural — and what that means

Several real, structural constraints limit how many journeys can be individually demonstrated, regardless of how good the measurement infrastructure is.

Walled gardens do not share connecting signals. Google’s attribution data does not connect to Meta’s. A consumer who saw a Meta ad, then a YouTube pre-roll, then searched and converted represents a journey that cannot be reconstructed from platform-reported touchpoints. The chain is broken at the platform boundary. A December 2024 IAB survey found 64% of US ad buyers plan to focus significantly more on cross-platform measurement — not because the problem is solved, but because it is not.

Mobile environments sample rather than record. SDK-based measurement captures aggregate patterns, not complete individual event chains. Only 8% of companies have a fully unified view of app marketing performance across channels, per a 2025 Branch survey. The remaining 92% are working with partial data at the point where mobile enters the journey.

Offline conversion matching has a structural ceiling that vendor marketing obscures. The match rate figures commonly marketed — 70%, 80% — typically describe online-to-CRM matching or platform offline import matching, both of which solve easier problems than true independent offline attribution. Independent offline attribution, connecting a physical conversion event to an individually attributed multi-channel digital journey, runs 4-20% under real-world conditions. The constraints are structural: shared IP addresses, multi-device fragmentation, households with no digital identity to match, point-of-sale transactions that carry no digital identifier. Any vendor claiming materially higher rates for true independent offline attribution is measuring something different.

Organic touchpoints have no upstream media to attribute. Direct visits, organic search, word-of-mouth referrals — these are single-event interactions by definition. Any program with meaningful organic traffic will show a significant portion of single-touch journeys because those journeys are single-touch.

The ceiling exists. The question is whether the vendor acknowledges it or models over it.

What the Signal Manifest and Attribution Manifest make possible

Showing the work requires infrastructure for showing it.

The Signal Manifest documents chain of custody for the data entering the measurement system — what sources were used, when the data was delivered, what quality checks were applied, where gaps exist. The Attribution Manifest documents what was observed at each point in the journey, with what confidence, and for offline conversion outcomes, what match rate was achieved and which conversion methodology was applied.

Together, they make the reported number interrogable. A CMO asks: how confident are you in the TV attribution? The answer is in the record — here is what was observed, here is how the journey was constructed, here is what remains unobserved and why. The multi-touch rate is the proven rate. What remains undemonstrated is labeled as such, not averaged away.

That is a harder conversation to have than presenting a 70% rate and hoping no one asks how it was derived. It is also the only conversation that holds up when a sophisticated buyer, a CFO, or an auditor looks closely.

The structural argument

Independent multi-touch attribution is better than last-touch measurement — not because it is a perfect representation of consumer reality, but because it is a closer and more defensible one. The limitations are real. Walled gardens prevent cross-platform stitching. Mobile environments sample rather than record. Offline conversion matching has a structural ceiling well below what most vendors market. Organic touchpoints are single-touch by definition.

These are honest constraints, not weaknesses of the approach. Modeling over them — through probabilistic extension, panel augmentation, or definitional expansion — produces a number that looks better and means less.

We show you what we can prove. That is the methodology.

Follow the Spend is a newsletter about marketing measurement — what it gets right, what it gets wrong, and what it takes to close the gap. Published by C3 Metrics.

Read the full Data Lab piece →

The debate about TV and multi-touch attribution has settled into a comfortable consensus: broadcast media cannot get into an MTA model, so practitioners should use media mix modeling for TV and reserve MTA for digital. The logic is clean. The conclusion is wrong.

The industry has focused on whether TV can be measured. The more consequential question is what happens to your model when TV is absent from it.

The Model Is Missing Half the Picture.

When a television campaign runs, branded search volume increases. Direct traffic rises. Social engagement lifts. Retargeting conversion rates improve. These are consistent, measurable, and well-documented patterns in attribution data across categories.

Your MTA model records all of that downstream activity and attributes it to the digital channels where it shows up. Paid search receives credit for conversions it did not originate. Retargeting takes credit for demand it did not create. The model measures the performance marketing layer sitting on top of a demand-creation process it cannot see. When a television campaign runs and your paid search ROAS improves, the two events are causally related — and a TV-blind model treats them as coincidence.

This is the TV problem worth solving. For the integrity of every channel already in the model.

The BOS Signal

BOS stands for Blended Offline Signal — the behavioral footprint that offline media exposure creates in downstream digital activity. When a consumer encounters a TV spot, a radio placement, or an outdoor campaign, conversion rarely happens at the moment of exposure. It happens later: through a branded search, a direct site visit, a navigation from memory. BOS captures that behavioral response and maps it back to the offline media that generated it.

The mechanism works at the individual spot level. When a specific creative airs in a specific market, downstream digital activity in that geography responds in a short, identifiable window after the airing. That response — detached from background signal, validated against the actual flight schedule at the DMA level — becomes the basis for attribution credit. A touchpoint, placed in the model at the spot level rather than aggregated across a flight window.

Branded search is the most observable signal, and the most common starting point. A consumer who searches a brand name in the minutes after a spot airs is demonstrating something clear and measurable. But branded search is also what every TV measurement tool is already watching. The more useful question is what else moves — direct traffic, organic search behavior, secondary digital signals that respond to the same exposure and tell a richer story about the consumer’s path. The BOS signal is configurable for a reason. The obvious signal gets you in the door. The fuller signal set is where the real attribution work happens.

This is what distinguishes BOS from standalone TV measurement. Specialized TV measurement vendors do rigorous work quantifying how individual spots drive branded search lift. Their answer tells you how well TV performed as a channel. BOS asks a different question: given that this spot drove this consumer toward a correlated digital signal, what was the full journey from there to conversion, and where does TV belong in that path?

The measurement is probabilistic. The touchpoints are real.

What Changes When TV Is in the Model

The ORAC framework classifies every touchpoint by its functional role in the consumer journey: Originator, Roster, Assist, or Converter. Offline channels earn the position the data assigns them.

In most programs with significant TV investment, offline media registers as the most common Originator — the channel that creates awareness before purchase intent exists. The digital channels that follow fill the downstream roles, doing exactly what they are built to do: converting demand that was already in motion.

The channels that look strong in a TV-blind model frequently look different once BOS attribution is included. Paid search on brand terms — one of the most expensive line items in most budgets — often captures demand that TV created. That distinction matters for every budget conversation that follows.

Getting TV into the model does not just correct TV’s attribution. It corrects the model.

The technical methodology behind BOS signal integration — including DMA-level validation, unlimited attribution windows, and halo effect quantification — is covered in full at the C3 Metrics Data Lab: c3metrics.com/dl-bos-signal.html

They’re related. And together, they make the case for what MTA is actually for.

What the channel was trying to do

The ORAC framework — Originator, Roster, Assist, Converter — does one specific thing: it breaks the path to conversion into functional roles and assigns each channel the success signal that matches its role. A Converter gets credit for converting. An Originator gets credit for initiating. A Roster channel gets credit for maintaining awareness during the consideration window.

The question ORAC asks is different from standard attribution: not “which touchpoint preceded conversion?” but “what was this channel trying to accomplish, and did it succeed at that specific job?”

That distinction matters. A TV ad that creates branded search intent two weeks later looks expensive in a model measuring last-touch digital conversion. Measured against its actual job — initiating the journey — the economics look different. The credit flows to where it belongs, which changes what every channel in the model looks like, not just TV.

When the intent surfaces

The day-of-week analysis from the same dataset surfaces a different dimension of the same data.

Thursday CPC: $18.85. Friday CPC: $45.37. That’s not a rounding error — it’s a 140% differential across one quarter of actual campaign data. The floor-case argument: shift 20% of Friday’s budget to Thursday and recover at least 842 additional conversions without increasing total spend. The efficiency differential gets averaged away in aggregate reporting — it takes disaggregation to see it.

The Thursday advantage connects back to ORAC. Consumers who convert on Thursday are converting after they’ve been reached — the campaign ran, the TV flights did their work, and Thursday is when that intent surfaces in search and site behavior. Friday conversions are often captures of the same intent pool, at higher cost. The Originator activity created demand earlier in the week. The Converter budget followed a schedule that didn’t account for it.

What MTA actually does

ORAC tells you which channels create intent. Day-of-week tells you when that intent surfaces for capture. Together they answer a question that neither MMM nor incrementality testing can reach: when does Originator activity create Converter opportunity, and is the Converter budget deployed at the right moment?

That’s what omni-channel MTA does. It follows the journey — all of it, including the single-touch journeys that account for most conversions in considered-purchase categories — and reads the structure of what happened, not just the outcome.

Incrementality testing is a useful complement. It answers “did this channel work?” with a clean causal answer. It doesn’t tell you which role the channel played, when its effect surfaced in downstream behavior, or whether the deployment schedule aligned with the moment consumers were ready to act. Holdout experiments don’t produce day-of-week findings. They answer a different question.

The data here requires investment — building complete journey attribution across channels, including offline, including single-touch, requires methodology and signal discipline. The payoff doesn’t wait. The Thursday-Friday finding is actionable in the next media plan. The Originator reallocation changes the next budget conversation. The return starts with the first read-out.

Two pages from one dataset

These two findings came from a single quarter of campaign data. One analysis. But they’re not the whole list — not close.

When you run an honest attribution program against a real dataset, you come back with a working list of questions worth following: creative performance by ORAC role — which creatives work as Originators, which as Converters? BOS lift windows by DMA. Time-to-conversion by originator type. Seasonal BOS patterns versus flight schedule. Cross-device journey reconstruction. Channel interaction effects — does TV move activity in channels beyond search? The list runs well past a single meeting before you’ve gotten through one quarter.

The discipline isn’t generating the questions. It’s deciding which ones lead somewhere actionable, running the analysis honestly, and being willing to accept when the data says there’s nothing there. Questions lead to analysis lead to revised recommendations — or they don’t. Either answer is useful.

These two surfaced because we asked. They’re compelling and they’re actionable — 842 additional conversions from a budget reallocation is a number you can take to the next planning meeting. And they fund the next question, and the analysis after that.

The findings pay for the work and the next round of work. Payback is both immediate and recurring.

Data Lab: dl-orac.html and dl-dayofweek.html

Every platform-side measurement system shares the same structural feature: it has no incentive to tell you when you’re spending more than the channel can efficiently absorb. YouTube’s auction will accept spend above whatever threshold makes your buying efficient, report the delivery as performing, and give you no visibility into where the curve breaks. The platform is on the other side of that transaction.

This is what that looks like in a real program — and why the finding only surfaced through external analysis.

The Setup: A National Brand’s Quarterly YouTube Program

Analyzing a national brand’s single-quarter YouTube program, C3 identified a clear efficiency threshold at approximately $100,000 in weekly spend. Below that level, cost per click ran consistently between $12 and $20. Above it, costs escalated to $25–35 per click — roughly double to triple the rate — with cost per impression rising in the same pattern.

The inflection is not subtle. Plotted as a scatter of weekly spend against cost per click across the quarter, the break is visible: a cluster of efficiently-priced weeks below the threshold, and a distinct, higher-cost cluster above it. Cost per impression shows the same curve independently, which rules out click-quality variance as the explanation. The auction itself is becoming more expensive — not just less responsive.

What made this finding possible was the nature of the quarter’s spend pattern. Weekly YouTube investment varied significantly — some weeks below $50,000, others above $175,000 — creating enough natural variation to reveal the efficiency curve across the dataset. That variation was not planned as a test. The brand was managing spend against other priorities. But the variation created the analytical equivalent of a natural experiment: observable spend differences, observable efficiency differences, and enough data points to see the pattern clearly.

[Full scatter chart — Cost per Click vs. Weekly YouTube Spend — at c3metrics.com/dl-youtube-efficiency.html]

Why Platform Reporting Doesn’t Surface This

YouTube’s native reporting shows campaign-level cost per click, cost per view, and impression metrics. What it does not show is how those metrics change as a function of weekly spend level — because that would require the platform to present a saturation curve, which is effectively a recommendation to spend less with them.

This is not a conspiracy. It is a structural incentive. The platform’s optimization objective is to maximize relevant ad delivery within your targeting parameters and budget. When your budget exceeds the efficient reach of your targetable audience, the auction clears at progressively higher prices — you’re bidding against your own prior delivery, reaching lower-attention inventory, or paying for frequency you’ve already exhausted. The platform reports all of this as delivering. Technically it is. Efficiently it is not.

The only way to see the saturation curve is to analyze your own delivery data externally — spend and performance by week, over long enough a period to see the variance — without relying on platform-aggregated summaries that smooth the signal you’re looking for.

A second signal confirms the finding isn’t a click-quality artifact. When cost per impression — a metric with no click component — shows the same inflection at the same threshold, the explanation is the auction itself clearing at higher prices, not a shift in who’s clicking.

[Full scatter chart — Cost per Impression vs. Weekly YouTube Spend — at c3metrics.com/dl-youtube-efficiency.html]

What This Threshold Means — and Doesn’t Mean

Finding a saturation threshold is not an argument for cutting YouTube. It is an argument for knowing where to stop, and acting on that knowledge.

The brand in this analysis had weeks well above $100,000, during which it was paying $25–35 per click for the same clicks it could have obtained for $12–20 below the threshold. The dollar magnitude of above-threshold spend across the quarter was significant — expenditure that was measurable, and that generated clicks at nearly twice the unit cost of the efficient range. That is not a rounding error. It is a reallocation opportunity.

The right response is not to cut YouTube to the threshold and call it efficiency. It is to redirect the above-threshold budget to channels still operating within their efficient range, or to expand targeting parameters to reset the auction dynamics and raise the ceiling. Both responses require knowing where the threshold is. That knowledge is not available from the platform.

It is also worth noting that the threshold is specific to this program, this audience, and this quarter. A different brand with a larger targetable audience will have a different inflection point. The methodology is the same; the number will differ. What is consistent across programs is the structural dynamic: every platform auction has a point beyond which incremental spend yields diminishing — and eventually negative — marginal efficiency. Finding that point for your specific program is the work.

The Broader Pattern: Platform Incentives and Measurement Independence

YouTube is not unique in this dynamic. Every major platform auction — paid social, display, video — has the same structural feature. Spend within the efficient reach of your targetable audience and you get good value. Exceed it and you pay more for less, without the platform flagging the shift.

What is notable about YouTube at scale is that the magnitude of above-threshold spend can be substantial, and the efficiency gap between the efficient and saturated zones is pronounced. The scatter in this analysis showed cost per click more than doubling above the threshold — not a marginal degradation but a step change. That step change is invisible in platform reporting and visible only in external analysis of your own delivery data.

Independent measurement exists, in part, to surface exactly this kind of finding. Not because platforms are acting in bad faith, but because the information required to identify a saturation threshold is not in the platform’s interest to produce. The advertiser is entitled to that information. Getting it requires looking at your own data from outside the platform’s reporting interface.

The Reallocation Question

When a brand discovers it has been spending above its YouTube efficiency threshold, the immediate question is where the above-threshold budget should go. The answer depends on which other channels in the program are still operating below their own saturation points — which requires the same external analysis applied across the full media mix, not just YouTube.

The finding that YouTube has a saturation threshold at $X per week is one output. The full output is a mapped efficiency curve across every channel in the program — showing which are underinvested, which are saturated, and where reallocation generates the most marginal return. That is the measurement program the platform has no incentive to run for you.

This week's piece is the one I've been thinking about longest. AI isn't just changing how consumers find information — it's moving the most valuable part of the consumer journey somewhere measurement can't follow. Here's the framework for what that means, category by category.

The consumer research journey has always been a sequence. Someone becomes aware of a brand — through a TV ad, a social post, a display unit, a word-of-mouth recommendation. They enter a consideration phase, gathering information, comparing options, reading reviews. They develop intent and eventually convert. Attribution exists to assign credit to the touchpoints in that sequence.

Generative AI assistants are automating the middle of that journey — and doing it invisibly. When a consumer asks ChatGPT to compare SUV brands, or asks Perplexity to summarize financing options, or uses an AI assistant to research pharmaceutical options, that entire research session produces no analytics signal. There is no impression, no click, no referral. The journey happened; you just can’t see it.

This is not primarily a content strategy problem or an SEO problem, though it is both of those too. It is a marketing measurement problem — and it is coming industry by industry, at different speeds and in different forms, depending on how deeply research was already baked into the category. What follows is a framework for where it’s already arriving and what it will require of attribution infrastructure when it does.

The ORAC Lens on What AI Does

C3 Metrics classifies every touchpoint in a consumer journey by its functional role: Originator (first introduced the brand), Roster (maintained visibility during consideration), Assist (actively moved toward conversion), and Converter (present at the moment of purchase). Standard attribution models assign credit; ORAC classification assigns role. The distinction matters because role determines strategic value — a channel that originates 60% of converting journeys deserves a very different budget decision than a channel that converts 60% of journeys that were already going to convert.

What AI assistants are doing, at scale and with increasing sophistication, is absorbing the Originator and Roster functions of the consumer journey into private sessions that leave no measurable footprint. A consumer researches your brand inside a ChatGPT conversation. They form a preference. They leave the session and, days later, search your brand name on Google and convert. Your attribution model sees: brand search → conversion. It credits the Converter. The actual Originator — the LLM session that generated the preference — is invisible.

This is a structural problem, not a data quality problem. No additional tagging, no server-side collection improvement, no cookie will solve it. The data doesn’t exist to collect. The conversation happened in a closed system.

It Makes the Existing Problem Worse

The over-crediting of brand search and last-click converters is not a new issue. In every attribution program that has moved from last-click to multi-touch, the credit distribution for paid brand search drops substantially — because the journey analysis shows that brand search is almost always a Converter, not an Originator. The consumer was already going to buy; the brand search was how they returned to complete the purchase. The channel that should receive Originator credit is typically much further up the funnel: TV, upper-funnel display, a content touchpoint that first surfaced the brand.

AI-compressed journeys accelerate this distortion sharply. As a larger share of Originator activity moves into LLM sessions — private, untracked, without any referral data — the channels that remain measurable are disproportionately the Converters. The measurement environment becomes more last-click by default, not by design. Budgets optimized against that signal will flow toward Converter channels. Originator channels — the ones that created the brand preference that the AI session then reinforced — get cut. The pipeline eventually empties.

The behavioral logic is observable in what little referral data does exist. Consumers who click through from an AI citation arrive with a formed preference, not an open question. The research happened inside the AI session; the click is the Converter event. You see the end of the journey, not the journey itself. Claims about exactly how much better this traffic converts vary widely by source — and critically, most of the sources making large claims have financial interests in validating AI’s performance. The directionality is real; the specific multiples should be treated skeptically until measured independently within your own programs.

The Industry Asymmetry

The urgency of the AI attribution problem is not uniform across categories. It scales with how deeply research was already embedded in the consumer journey — and how much of that research was previously generating measurable signals.

Automotive is the clearest current case. Endemic research behavior has always been a defining feature of the category: buyers spend weeks comparing on Edmunds, KBB, manufacturer configurators, and dealer sites before they ever contact a dealer. That behavior was always generating measurable touchpoints — third-party platform impressions, clicks, time-on-site signals. AI is now compressing and absorbing that research phase into LLM sessions. Cox Automotive data from 2025 found that 25% of new-vehicle buyers used AI tools during their shopping process — the first year the study tracked the metric. The Originator signal isn’t disappearing from automotive journeys; it’s moving somewhere it can no longer be tracked. The disruption is already in process, not on the horizon.

Financial services and insurance have similar endemic research depth, but a different attribution challenge: in those categories, the conversion event is predominantly offline — a phone call, an application submission, a branch visit. Attribution was already partially broken before AI arrived. AI adds an invisible layer to a journey that was already difficult to measure end-to-end, which intensifies the need for offline conversion integration, not just better digital tracking.

Categories with shorter consideration cycles — CPG, quick-service, subscription products — face a different version of the problem. In those verticals, there wasn’t much of a digital research phase to begin with. A consumer deciding which protein powder to buy or which streaming service to try didn’t leave a rich chain of pre-conversion touchpoints. AI may actually be inserting a new Originator phase into categories that previously had almost no top-funnel digital signal — creating AI-influenced consideration where there was previously a near-direct brand-to-purchase path. The attribution challenge there isn’t about preserving signal; it’s about building the infrastructure to capture something new.

The implication: a single response to AI’s impact on attribution doesn’t fit every category. The measurement priorities differ, and the timing differs. What’s consistent is that the frameworks — ORAC role classification, BOS signal detection, new-vs.-returning segmentation — are the right analytical architecture regardless of which version of the problem a given program is facing.

The Sponsored Inclusion Question

The advertising industry has been watching AI platforms’ approaches to monetization closely, because paid inclusion in AI results creates both an opportunity and a familiar conflict. The landscape as of early 2026 is more interesting than “ads are coming.”

OpenAI launched advertising in ChatGPT in early 2026, implemented as clearly labeled sponsored results displayed below organic AI-generated answers — explicitly segregated from the organic response. Perplexity, after testing sponsored placements, explicitly abandoned the model in February 2026. Their stated reasoning was direct: if users believe ads can influence the answer, trust in the answer collapses. They chose the trust-first model and are building a merchant integration layer instead. Google is expanding advertising into AI Mode with its existing infrastructure already in place.

This divergence is meaningful. The platforms choosing organic trust over ad revenue are making a structural bet — that uncompromised recommendations are the long-term durable asset. The platforms building ad products are entering the same tension that has always defined platform self-reporting: a financial relationship with the channels being evaluated creates pressure on the evaluation. Whether the pressure bends the answer intentionally or gradually doesn’t change the structural problem.

The measurement implication is layered. As paid inclusion scales in some AI platforms, attribution models will need to distinguish AI organic (earned citation) from AI sponsored (paid placement) — the same distinction MTA makes between organic and paid search. These are different signals with different cost structures, different intent profiles, and different credit implications. For now, that distinction isn’t being made in most attribution programs, because AI as a channel is still being established. But the architecture to make it will need to exist before the spend scales.

The Agentic Conversion Problem

The research automation case — AI handling discovery and consideration — is the present challenge. What comes next is agentic conversion: AI agents completing transactions on behalf of consumers. This is no longer entirely theoretical.

OpenAI launched in-chat purchasing with Instacart in December 2025, followed by integrations with Target and DoorDash. Perplexity’s “Buy with Pro” has been live since late 2024 and expanded to free users by late 2025, with access to 5,000+ merchants. Consumer adoption at scale remains early, but the infrastructure is shipping faster than expected.

When a consumer instructs an AI agent to “book a test drive at the nearest dealer” or “find me the best available rate and apply,” the conversion event happens without a human user journey. There is no click path. There is no site visit. There is an API call from an agent to a booking system, and a conversion that your current attribution infrastructure has no mechanism to capture. The standard MTA pipeline assumes a human consumer navigating a browser, leaving behavioral signals at each step. Agentic conversions break that assumption at the most fundamental level.

The practical response requires measurement at the agent request level: capturing the context of the AI agent interaction that produced the conversion request, not just the conversion event itself. The window to build that infrastructure before agentic commerce scales at volume is open now. It won’t stay open long.

What Shorter Funnels Mean for MTA Models

Multi-touch attribution models are trained on path data: sequences of touchpoints leading to conversion. The statistical confidence of those models improves with path length and touchpoint diversity. A model trained on paths with an average of 8–12 touchpoints produces tighter, more reliable credit assignments than one trained on paths with 2–3 touchpoints, because there is more signal to work with.

AI-compressed journeys are shorter journeys. When the Originator and Roster phases move into private LLM sessions, the visible path shrinks. Fewer touchpoints mean less data per path. At scale, this degrades model confidence — not because the attribution methodology is wrong, but because the observable signal is thinner.

Two things follow from this. First, the segmentation between new and returning customers becomes even more important. AI-influenced journeys are most likely for new customers in the awareness and consideration phase — the people who actually need to do research because they don’t yet have a relationship with the brand. Returning customers already have brand knowledge; they’re less likely to consult an AI assistant about something they’ve already decided. Mixing these populations in a single attribution model averages away the signal that matters most for understanding AI’s impact.

Second, the BOS signal methodology — which converts offline media exposures (TV, radio, OOH) into MTA touchpoints by tracking the branded search spikes they create — has a natural extension into AI attribution. When AI activity influences a consumer to search for a brand, that branded search is the measurable downstream consequence of an invisible upstream exposure. The same correlation logic that identifies TV-driven branded search spikes can identify AI-influenced branded search volume — not with certainty at the individual journey level, but statistically across program-level data. The branded search spike that follows a brand’s appearance in a widely-distributed AI response is a signal, and it’s measurable.

What Independent Measurement Protects Against

As AI platforms become advertising channels — which OpenAI’s February 2026 launch confirms they are becoming — the structural conflict of interest argument applies again with greater force. A measurement partner with commercial relationships with AI advertising platforms has an incentive toward the same favorable attribution that has long distorted platform self-reported search and social metrics. The platforms with the most to gain from being credited for AI-influenced conversions are the most motivated to offer measurement products that show those conversions. The separation between who sells the media and who measures its effectiveness doesn’t become less important as the media environment becomes more complex. It becomes more important.

There is a subtler version of the same problem worth naming. When advertisers ask for guidance on how AI will affect their measurement, the answers they receive will reflect the interests of whoever provides them. A measurement recommendation from a party with strategic ties to a major AI platform — or from an AI tool built by a company with its own advertising ambitions — is not a structurally neutral recommendation, regardless of whether the bias is intentional. Google’s position, unsurprisingly, is that advertisers should consolidate measurement inside Google’s ecosystem and trust it to report accurately on Google’s own channels. The question of whether to trust that position is exactly the question independent measurement exists to answer.

The combination of invisible AI influence, early-stage agentic commerce, and AI platforms entering the advertising market is producing an environment where the measurement problem is harder, the financial stakes of getting it wrong are higher, and the pressure on platform-reported metrics is intensifying across a new set of channels. The impartial perspective — from a party with no financial relationship to any of the channels being measured, including AI advertising — is the only structurally credible basis for answering what AI actually contributed.

C3 Metrics Approach

C3 Metrics is tracking AI-referred traffic as a distinct, first-class channel in client programs now — establishing the baseline data before AI advertising scales. We are building agentic conversion infrastructure: the measurement architecture to capture conversions completed by AI agents, not just human browser sessions. The ORAC classification framework provides the analytical lens to surface how AI is absorbing Originator-stage activity. And our industry-specific program design reflects the reality that automotive, financial services, and short-cycle consumer categories face this problem differently — with different timing, different signal gaps, and different measurement priorities. Independent measurement with no financial relationship to AI advertising channels is not a future positioning; it is the present structural fact that makes objective answers possible.

There’s a pattern worth naming in the TV and attribution measurement space right now.

Two of the most visible names in the category are both under pressure — one from a federal jury that found they built their product on data they didn’t own, the other from a data supply chain that just changed hands in ways that complicate their independence claim. The details are public. The implications are worth thinking through.

Because when a measurement company’s primary investment goes into celebrity, narrative, and market presence rather than the infrastructure underneath the number — the methodology has to come from somewhere. You can’t borrow your way to independent signal. At some point, the audit arrives.

What Actually Makes a Measurement Number Trustworthy

It’s not the brand. It’s not the famous co-founder. It’s not the chart that shows your TV spot performed well.

It’s whether the underlying signal was independently collected, verified against a baseline that isn’t owned by someone with a financial stake in the outcome, and documented in a way that survives scrutiny from finance, audit, or a federal courtroom.

Most measurement products fail that test quietly. They work well enough in normal conditions — when channels are stable, when data licensing holds, when nobody looks too closely at the methodology. The problems surface later, usually at the worst possible moment.

Where C3 Metrics Puts the Money

Every dollar C3 invests goes into the infrastructure that produces the number. No celebrity co-founders, no PR budget, no borrowed data supply chain — just Ground Signal™, the continuous verification layer that monitors every data source, flags every discrepancy, and documents every reconciliation decision in the Signal Manifest™.

The Signal Manifest™ is what we deliver alongside attribution results: a structured, auditable record of exactly how the number was produced and why it can be trusted. Not a dashboard. Not a summary. A documented proof of work — the kind that holds up when someone outside the marketing department asks hard questions.

That’s not a differentiator we invented for a pitch deck. It’s what you get when the entire budget goes into the product instead of the profile.

The Question Worth Asking

If your current measurement vendor is more famous than they are rigorous — if you’ve never seen the methodology documented, never had a data source explained, never been shown how a discrepancy was resolved — it might be worth asking what’s actually underneath the number you’re making budget decisions with.

The market is noisy right now. Some of the loudest voices have the shakiest foundations. That’s not a sales pitch. It’s just an observation from a company that has been doing this quietly, correctly, and independently for a long time.

The number is the number. We can show you how we got it.

Greg Collins is CEO of C3 Metrics®, an independent attribution platform. Follow the Spend is a newsletter about what’s actually happening in media measurement.

Subscribe now

In February, we detected 180 million suspicious views on a single campaign. Most of it arrived after midnight UTC, concentrated over roughly five hours, with no corresponding conversion activity anywhere in the funnel. Our team was up through the night managing it.

The analysis that followed identified approximately $600,000 in fraudulent or questionable spend across the campaign period. Every finding was supported by multiple independent data points: timing anomalies, device-type concentration, and traffic volume patterns inconsistent with any legitimate audience behavior.

The platform reported none of it.

That last sentence is the point of this newsletter.

Platform reporting isn’t designed to find this. It isn’t designed to find anything that reflects poorly on the platform. That’s not cynicism — it’s just the structural reality of asking the same company that sold you the media to tell you whether the media worked. The incentives don’t align with the truth. They align with the next buy.

Independent measurement exists to close that gap. Not as a supplement to platform reporting. As a replacement for it in any context where the accuracy of the answer actually matters.

The $600,000 wasn’t recovered because someone was clever. It was recovered because the tool doing the measurement had no stake in what it found. That’s the only architectural arrangement that produces honest answers consistently.

A few things that number implies for anyone managing a media budget at scale:

Your current measurement is probably not finding everything. Not because your team isn’t good — because the tools most teams rely on aren’t designed to surface findings that contradict the interests of the parties providing them. The absence of a fraud alert is not evidence of clean media. It may just be evidence of aligned incentives.

The gap between what platform reporting shows and what independent measurement finds is where media efficiency lives. In our experience, that gap is typically 15–20% of total spend. It doesn’t all look like fraud — most of it is misattribution, channel overlap, and self-reported performance that doesn’t survive independent verification. But it’s real, it’s consistent, and it’s recoverable.

Independence isn’t a feature. It’s a precondition. Any measurement methodology that involves revenue from the media being measured, equity relationships with publishers, or tools that sit inside the platforms they’re evaluating is structurally compromised — regardless of the contract language around objectivity. The conflict is architectural, not contractual.

We published the full methodology and findings in the C3 Metrics Data Lab. If your measurement comes from the same companies selling you the media, it’s worth reading.

Follow the Spend is written by Greg Collins, CEO of C3 Metrics and founder of Cape Fear Advisors. No media buying. No agency relationships. No stake in what you spend next.