The Attribution Problem in AI Ads: Crediting a Sponsored Recommendation

TL;DR

→  The hard part of AI-ad attribution is not that conversions are difficult to track. It is that the decisive touch — the model naming you inside the answer — happens in a private conversation and leaves no click to tag. It is untrackable by design.

→  Because only the ad click is instrumented, every measurement error runs one way: it over-credits the paid slot and under-credits the earned citation sitting in the same answer. Better tracking of the click widens the bias rather than closing it.

→  That turns attribution into misallocation: you fund what you can measure, which makes earned look weak, which defunds the thing actually driving the conversion. The error is systematic, not random.

→  Stop assigning credit to touches and start measuring cause. Reconcile three numbers — platform-credited, declared, and incrementality-tested — and act on the lift, not the last click.

The new tooling measures the click better, not the problem

When ChatGPT’s ad system launched in February 2026 it reported almost nothing — impressions and clicks. By late May it had shipped a conversion pixel, a Conversions API, and view-through reporting, and the industry treated the gap as closed: now you can measure AI ads like any other channel. That relief is misplaced, because every one of those additions instruments the same thing — the click on the sponsored card — with more precision. None of them can see the touch that actually did the persuading.

The distinction between precision and accuracy is the whole subject. A more precise pixel measures the click to two decimal places; it does not make the click the cause. If the conversion was really driven by the model recommending you in the answer above the ad, then measuring the click perfectly just gives you a perfectly precise wrong number. In a channel where the decisive influence is structurally invisible, improving the visible instrument does not reduce the error — it sharpens your confidence in a figure that is biased in a known direction, which is worse than honest uncertainty.

So the benchmarks that quote a tidy cost-per-acquisition from the ad dashboard are quoting the output of a measurement system that can only see one of the two touches present at the moment of conversion. Reading that number as “what the ad did” is the central error of AI-ad attribution, and the rest of this piece is about why the error is unavoidable with touch-based tools and what to use instead.

It is worth being concrete about what shipped, because the additions sound comprehensive until you notice they all point at the same object. The pixel records a conversion after a click on the card. The Conversions API sends server-side conversions back to be matched against those clicks. View-through reporting extends credit to users who merely saw the card and later converted — which, far from helping, widens the bias, because it hands the paid slot credit for conversions it did not even earn a click on. Every lever tightens the platform’s claim on outcomes while the citation’s claim stays structurally unrecordable. The toolkit grew; the blind spot did not shrink, and view-through quietly made it larger.

Why the decisive touch is untrackable by design

In a conventional display or search campaign, the ad is usually the only branded touch on the path, so crediting the click is a defensible simplification. An AI answer breaks that assumption completely. The user sees two things in the same viewport: the answer, which may name you as the recommended option, and — sometimes — a sponsored card beneath it. One of those is instrumented and one is not, and it is the uninstrumented one that carries the engine’s borrowed authority.

The citation is untrackable for a reason that will not be engineered away: the recommendation happens inside a conversation the advertiser is never allowed to see. Conversation privacy is one of the platform’s stated operating principles, and it means the single most persuasive event in the funnel — the model telling a user, in private, that you are the answer — cannot be logged, tagged, or passed to your analytics even in principle. This is not a temporary gap in the tooling; it is a designed property of the surface. You can instrument the click on the ad forever and never instrument the sentence that made the click unnecessary.

The scale of the invisible pipe is not small. Muck Rack’s May 2026 analysis of more than 25 million cited links found earned media accounts for 84% of AI citations against 0.3% for paid, so the channel that converts best — the AI product recommendation itself — is also the one your dashboard cannot see. Attribution here is not measuring a river and missing a stream; it is measuring the stream and missing the river.

The obvious patch — just ask people how they found you — fails for a reason specific to this channel. Users do not experience the earned citation as your marketing; they experience it as the assistant’s own answer. Asked how they heard of you, they say “ChatGPT told me,” crediting the surface, not the coverage that put you in it. The most causally important input — the third-party corroboration that made the model name you — is invisible even to the user who acted on it. So the decisive touch is unseen by the platform, unrecorded by your analytics, and un-recalled by the customer: three independent measurement channels sharing one blind spot in exactly the same place.

A further asymmetry shrinks the trackable pipe to a sliver. Sponsored cards draw click-through around 0.68% — well under a hundredth of exposures — so even on the paid side, the overwhelming majority of ad impressions leave no click and therefore no record, converting later, if at all, through paths the pixel never sees. The instrumented event is not merely one of two touches; it is a rare event on one of the two touches. Build an allocation on it and you are building on the visible corner of the visible half, then calling the result the campaign’s performance.

The bias has a direction, and that is what makes it dangerous

If the measurement error were random — sometimes over-crediting the ad, sometimes under — you could average it away. It is not random. Because only the trackable touch can receive credit, and the trackable touch is the paid one, the error runs in a single direction every time: paid is systematically over-credited and the earned citation systematically under-credited. A bias you can predict the sign of is not noise; it is a lever quietly tilting every decision the same way.

That tilt compounds through the budget cycle. You fund what you can measure, so the well-instrumented ad gets credit and gets more budget; the un-instrumented citation shows up as roughly zero in the attribution report and loses the argument for investment — even though the entity authority behind it is what made the ad convert at all. Starve the citation and the ad’s own performance quietly degrades, because the ad was harvesting demand the citation created. The measurement system thus recommends defunding the thing that makes the measured channel work, and it does so with a straight face and a confident CPA.

This is the same trap this journal has described from the spending side — paying to reach an audience your earned presence already won — seen now from the measurement side. The counter-bid looks rational precisely because the attribution system credits the harvest and hides the source. Fixing the spending mistake is impossible while the measurement mistake stands, because the numbers will keep pointing the wrong way. You cannot out-allocate a biased scoreboard; you have to change the scoreboard.

Watch the ratchet turn over a few cycles. Quarter one, the ad is credited and earns more budget while the citation reads as zero and loses its case for investment. Quarter two, with less behind earned coverage, your citation share slips, so the ad — now firing on fewer queries where you are the recommended answer — quietly converts worse, yet still beats the citation it is measured against, which is being scored at zero. Quarter three, the “data-driven” response is to cut earned further and lean harder on paid. Each step is locally rational given the numbers, and the sequence marches you toward spending more to harvest a shrinking pool of demand you are no longer creating. A directional bias does not mislead once; it sets a trajectory.

The three numbers that actually matter

The way out is to stop asking “which touch gets the credit?” — a question with no honest answer here — and start reconciling three different measurements of the same outcomes. Each sees something the others miss, and the gaps between them are the actual finding.

The first number is platform-credited: what the ad dashboard claims, on last-click plus whatever view-through it now offers. The second is declared or modeled: what users say when asked how they found you, or what a media-mix model attributes across channels without relying on clicks. The third, and the only one that answers the causal question, is incrementality-tested: the lift measured by withholding the ad from a matched group and watching what actually changes. Put the three side by side for each touch and the pattern is stark.

Read the table as a single claim: last-click credits the harvesters and starves the creator. The sponsored card and branded search both score well on platform credit because they tend to be the last step of a journey the citation began — but their incremental lift is a fraction of their credit. The citation scores near zero on credit and highest on cause. Any system that allocates budget by the first column will defund the third, which is the one paying the bills. The reconciliation does not require perfect numbers; even rough versions of all three, disagreeing in this shape, tell you your dashboard is mis-ranking your channels.

Each number has a characteristic blindness worth naming. Platform credit sees clicks and nothing upstream of them, so it flatters whatever touch comes last. Declared and modeled contribution catches touches with no click — a media-mix model infers a channel’s effect from spend-and-outcome patterns without tracing a single user — but it is coarse and slow to update. Only the incrementality test answers the causal question directly, and only for one campaign at a time. The reconciliation works precisely because the three fail differently: where the click-blind method and the click-based method disagree most, you have found a touch the dashboard cannot price. You are not hoping the three agree; you are mining their disagreement.

It is worth adding that the platform-credited number is not merely biased but unaudited. The reporting still carries no third-party verification, so the figure teams lean on hardest is self-scored by the party selling the inventory — a conflict every other mature ad channel eventually resolved with independent measurement, and one AI ads have not yet. That is not an accusation of bad faith; it is a reason to treat the dashboard as a claim to be checked rather than a reading to be trusted, which is exactly what reconciling it against declared and tested numbers does.

Reading the gap between credit and cause

The direction of the gap is a diagnosis, not just a discrepancy. When platform credit greatly exceeds measured lift, the channel is harvesting: it is being handed conversions that would have happened without it, usually because it sits at the end of a path some earned touch created. A sponsored slot firing on a query where you are already the cited answer is the textbook case — the ad collects the click, the citation did the work, and the dashboard thanks the ad.

When measured lift exceeds credit, the channel is under-credited, and in this environment that is almost always the earned citation and the coverage feeding it — the digital PR and third-party corroboration that put you in the answer in the first place. These touches leave no click, so they cannot win a last-click argument, yet a presence holdout will show them carrying most of the causal load. The gap tells you where to move budget: away from the harvesters you are over-paying and toward the under-credited sources actually generating demand.

There is a genuinely incremental role for the ad, and the gap finds it too. On queries where you are not cited, the sponsored card is not harvesting anything — there is no earned touch to ride — so its credited conversions and its incremental lift converge. That convergence is the signal that a placement is doing real work rather than collecting a toll. Pricing and defence both pointed to the same conclusion from other angles; measurement confirms it: buy the slot where the citation is absent, not where it is already firing.

A rough magnitude turns the direction into a decision. If a channel’s incremental lift comes in below, say, half its platform-credited conversions, it is doing more harvesting than creating and its real cost per outcome is at least double what the dashboard shows — treat the reported figure as fiction and re-underwrite the spend. If lift lands within striking distance of credit, the channel is largely doing genuine work and the dashboard is roughly trustworthy for it. The exact threshold is yours to set; the habit is the point. Never renew a co-present, cited-query placement without a lift number beside its credit, because that is exactly the placement the credit figure is most likely to overstate.

When to stop attributing and start testing

Not every path needs an experiment. Plenty of conversions still travel simple, well-instrumented routes where last-click is a fair approximation. The skill is telling those apart from the paths where deterministic attribution is structurally impossible, so you spend your testing effort where credit and cause actually diverge. Four questions settle it quickly.

1. Is the decisive touch inside a private conversation? If the thing likely to have persuaded the user was the model’s recommendation, it is unloggable by design — no attribution tool will ever capture it, so measure lift instead.
2. Is the trackable touch merely the last step of a longer journey? A branded search or a sponsored click that ends a path some earned touch began will be over-credited by any last-click model. Suspect harvesting.
3. Does the conversion window capture the real lag? By the platform’s own accounting a large share of conversions land outside the click window; if your buying decisions assume they are captured, the reported CPA is flattering a channel whose true payback you have not seen.
4. Is the ad co-present with your citation on this query? Co-presence guarantees an untrackable rival touch. Where you are cited and advertised at once, credit and cause cannot be separated by tooling — only by a holdout.

Two or more “yes” answers mean the path cannot be attributed deterministically and must be measured causally. One or none, and last-click is probably good enough to act on. The test keeps you from the two opposite failures: treating every click as truth, and running expensive experiments on paths that never needed them.

Run a real case through it. A user asks a category question, the answer names you, a rival’s sponsored card sits beneath it, and a sale lands two days later after a branded search. Was the decisive touch in a private conversation? Yes — the recommendation. Is the trackable touch the last step of a longer journey? Yes — the branded search ends a path the citation began. Two “yes” answers already, so last-click will confidently credit the branded search, the truth sits in the citation, and the only way to size either is a holdout. The test took ten seconds and overruled the dashboard.

That is the test’s real value: it is fast enough to run in your head during a budget meeting, and it points you at the handful of paths worth the cost of a real experiment instead of spreading scarce measurement effort evenly across paths that never needed it.

Running a holdout when the platform gives you almost nothing

Incrementality sounds heavy, but it does not require the platform’s cooperation — which is fortunate, because you will not get much. The workhorse is a geographic holdout: split comparable markets, keep the campaign running in one set and suppress it in a matched set, and read the difference in conversions. Because it compares outcomes between groups rather than tracing individuals, a geo test needs no pixel on the untrackable touch and is immune to the click-window problem that distorts the dashboard — it simply measures what changed when the ad was absent.

Matching the markets is where the rigour lives. Choose regions with similar baseline demand, comparable citation presence for your brand, and similar seasonality, then hold the suppression long enough to clear the lag. Where clean geographic splits are impossible, an audience holdout or a staggered on/off schedule across time can approximate the same logic, and a media-mix model can triangulate when experiments are genuinely out of reach. None of these is exotic; they are the standard causal toolkit, and they are the same methods you should already be using to value earned channels whose effects also refuse to show up in last-click, from link velocity to the slow compounding of newsworthy coverage.

The output is a single honest number: incremental conversions, and from them an incremental cost per outcome that you can put beside the earned figure on equal terms. It is less precise than the dashboard’s CPA and far more accurate, which is exactly the trade the whole subject demands. Instrument your own owned destinations well — a citable asset such as an interactive calculator makes a clean, measurable arrival point, and the technical measurement foundations let you see a branded or direct arrival cleanly — and the holdout gains power, because the clearer your view of the trackable half, the sharper the inference about the half you cannot see.

A few pitfalls decide whether a first holdout is worth running. Contamination is the main one: a national campaign, a podcast read, or word of mouth that crosses your market boundary muddies the comparison, so isolate the channel under test and keep other big moves out of the window. Small markets read noisily, so give the test enough volume to detect the lift you care about, an easier bar in some international markets than others. And the click-window problem cuts in your favour here: because a geo test reads totals rather than click-timed events, it captures the large share of outcomes that land outside the window and that the dashboard silently drops — but only if you hold the suppression long enough for those laggards to arrive, which usually means a quarter, not a fortnight. A holdout ended early understates earned’s slow build and flatters the fast-clicking ad.

A worked example

The company is invented and the figures illustrative; the method is exactly the one above. A B2B security vendor — call it Novareef — runs sponsored cards across its category and reports a healthy quarter: the ad dashboard credits the campaign with about 300 conversions at roughly £120 each, £36,000 of spend that looks like a clear win on last-click. On that number the team plans to double the budget.

Before doubling, they run a geo holdout: matched markets, the campaign suppressed in half of them for a full quarter to clear the lag. Conversions in the suppressed markets barely fall — scaled up, only about 90 of the 300 credited conversions disappear when the ad is switched off. The other ~210 converted anyway, because Novareef was already the cited answer on most of those queries and the ad was collecting clicks the citation had earned. The true incremental cost per outcome is not £120 but £36,000 ÷ 90 ≈ £400 — more than three times the dashboard figure, and a very different budgeting decision.

The reallocation follows the gap. Novareef keeps the sliver of spend that is genuinely incremental — the placements on queries where it is not cited — and moves the harvesting budget into the earned coverage a presence holdout showed was carrying the real load. It also stops judging the citation by a dashboard that scores it at zero, and starts valuing it by the lift it produces. Doubling the ad budget would have poured money into re-buying demand Novareef already owned, on the confident recommendation of a measurement system that could only see the cheaper half of the truth. The holdout cost a quarter of patience and changed the entire allocation.

The non-cited queries told the other half of the story. On placements where Novareef was not the recommended answer, suppression bit hard — conversions there fell close to one-for-one with the ad switched off, because there was no earned touch to harvest and the card was doing real acquisition. That convergence is the mirror image of the cited-query result, and it is what let Novareef keep a confident, smaller ad budget rather than cutting paid to zero in a panic. One caution earned its place in the write-up: a single quarter is one observation, seasonality can distort it, and the honest move is to re-run the test before treating £400 as a constant rather than this quarter’s estimate. The method is a discipline, not a one-off verdict.

The arithmetic of the reallocation is worth stating plainly, because it is the point of the exercise. Novareef had been about to double a £36,000 budget on the strength of a £120 CPA; the holdout put the incremental CPA nearer £400, so doubling would have meant paying a premium to re-buy demand it already owned. Redirecting the harvesting portion into the earned coverage that a presence holdout showed was carrying the load did not just cut waste — it funded the actual cause of the conversions the ad had been claiming. Same total budget, opposite trajectory, and the only thing that changed was refusing to let the dashboard rank the channels.

The objection this argument has to survive

The strongest challenge is practical, not theoretical: incrementality testing is a luxury most teams cannot afford. Clean geo holdouts need scale, matched markets, and the nerve to switch off spend for a quarter; plenty of advertisers have none of those. For them the platform’s attribution is not one option among three — it is the only data that exists, and directionally-biased data, the objection runs, still beats no data at all. Telling a small team to run experiments it cannot run is telling it to fly blind on principle.

The premise is fair: most teams cannot run a pristine experiment, and this piece is not pretending otherwise. But the conclusion does not follow, because biased data is not neutral data with more noise — it points a specific, predictable direction, and acting on it means over-investing in harvesting and under-investing in the source, every cycle, with compounding cost. A known bias is not a reason to trust the number; it is a reason to correct it. The practical answer is not a perfect experiment but a cheap one: a single crude geo split, one quarter, one campaign, is within reach of far more teams than believe it is — and even where no experiment is possible, applying a standing discount to last-click credit on co-present, cited queries beats taking the dashboard at face value. The bar is not scientific purity. It is refusing to treat a number you know is biased as if it were the truth. A team that cannot measure lift precisely can still decline to be led by a figure whose error has a direction — and that refusal, not a lab-grade holdout, is what separates good allocation from confidently funding the wrong half.

A more optimistic version of the objection says: give it time and the platforms, or a good media-mix model, will solve attribution the way the industry eventually did for social. They will not, and the reason is the structural fact this piece keeps returning to. Social ads became measurable because the decisive touch — the ad — was the thing being instrumented. Here the decisive touch is a private recommendation the platform has committed not to expose, so no amount of tooling maturity reaches it; a better model can infer around the blind spot but never see into it. Attribution improved for social by instrumenting the cause; in AI answers the cause is the one thing that cannot be instrumented, which is why the durable answer is measurement of lift, not attribution of touches.

Why this gets worse as the ad system matures

The gap between what you can measure and what converts is not stable; the surface is evolving in ways that widen it. Multi-advertiser layouts, already in testing, will place several sponsored options beside the answer, multiplying trackable clicks around a citation that stays untrackable — more precise credit for the harvesters, no more visibility for the source. Agentic commerce pulls the same way and harder: as purchases complete inside the assistant, on the platform’s own rails and against its roughly four-percent transaction fee, even the conversion event begins to happen somewhere you cannot fully see, so the trackable half of the journey shrinks from the far end as well as the near one.

The privacy wall around the recommendation is not loosening, and the commercial surface is moving more of the funnel behind it. The implication is not despair but sequencing: the teams that build a causal measurement habit now — while a click still exists to anchor a holdout against — will keep a working scoreboard as the trackable share erodes. The ones waiting for the dashboard to become trustworthy will find it becoming confidently less so. Measurement here is a race between your causal discipline and the surface’s growing opacity, and the discipline is the only side of that race you control.

What to do Monday

The whole discipline reduces to demoting the dashboard from truth to one biased witness, and cross-examining it against cause.

• Never read platform-credited CPA as “what the ad did.” It measures the click, not the recommendation above it. Treat it as a ceiling on the ad’s value, not an estimate of it.
• Reconcile three numbers, not one. Put platform-credit, declared/modeled contribution, and incrementality-tested lift side by side; the disagreement between them is the finding, and even rough versions expose a mis-ranked channel.
• Run the untrackable-assist test before you trust a path. Two or more “yes” answers mean the path must be measured causally, not attributed. Save the experiments for where credit and cause diverge.
• Buy the slot where the citation is absent. That is where credit and lift converge and the ad does real work; on queries where you are already cited, the placement is mostly harvesting — measure before you renew it, the way you would sanity-check any AI-visibility number against the underlying data.
• Value the citation by lift, not by clicks. A presence holdout will show the earned answer carrying load the dashboard scores at zero; fund it on that evidence, and instrument your owned surfaces so the trackable half stays clean even as AI browsers and agents reshape the path.

The reassurance underneath all of this is that the citation you cannot measure is usually the asset you most want. A sponsored click is easy to count and easy to over-value; the model naming you as the answer is impossible to tag and hard to over-value, because it is doing the persuading the ad merely collects on. Build the link-building strategy that earns that recommendation, measure it by the lift it causes rather than the clicks it cannot leave behind, and pressure-test every confident CPA against the current benchmarks, the real capabilities of the tools you rely on, and a holdout. The teams that lose here are the ones that let the one number they can see decide the budget; the ones that win treat measurement as a question about cause — and know the difference between counting a click and understanding what link building is actually doing.