AI Traffic Attribution Is Now A Bidding Problem

Every analytics team has now had the same meeting. Direct traffic is up, organic is down, somebody suggests that AI assistants are behind both, and the conversation ends with a plan to build a better AI channel grouping. That plan is fine as far as it goes. What it misses is that the mislabeled rows are not sitting quietly in a report waiting to be recategorised. They are being read, every day, by the bidding systems deciding where your paid budget goes.

AI Traffic Attribution Fails In One Direction

Start with the shape of the error, because the shape is the whole argument. Measurement noise that scatters evenly is survivable. You get a fuzzier number, you widen your confidence interval, you carry on. We have made that case before when arguing for honest error bars around AI search measurement, and it holds for most of what goes wrong in analytics. This is not that.

Referrer loss from AI assistants is concentrated rather than spread. It is worse on mobile than desktop, worse in native apps than browsers, and it varies sharply by assistant. Loamly, analysing 446,405 visits in February 2026, found that 70.6% of the AI traffic it could identify arrived in GA4 labelled as Direct: 14,413 sessions of the 20,428 it detected, against only 6,015 that carried a usable referrer. Worth naming the obvious incentive, since Loamly sells detection tooling and a large gap is the argument for buying it. Treat the exact figure as a vendor estimate. The direction is corroborated elsewhere and the direction is what matters.

Corroboration comes from a source with no such incentive. Orbit Media's study of 97 B2B and lead-generation websites, covering 28.9 million sessions between July 2025 and June 2026, put AI sources at roughly 0.5% of all traffic, about 140,000 sessions. Orbit's own write-up flags that the real figure is almost certainly higher because GA4 cannot classify what never arrives labelled. When the people publishing the number tell you their number is a floor, believe them.

So the error has a direction. Traffic leaves AI assistants, sheds its referrer on the way, and piles into Direct. Nothing pushes in the opposite direction. There is no mechanism by which organic search traffic gets accidentally credited to ChatGPT. A one-way error is not noise. It is a bias, and biases compound in whatever you build on top of them.

The Mislabeled Sessions Are Your Best Sessions

If the lost traffic converted like everything else, this would be an accounting curiosity. It does not. Orbit Media tiered key events by intent, separating contact forms, demo requests and phone calls from newsletter signups and downloads, then compared high-intent conversion rates by source. AI-referred sessions converted at 1.91%. Direct came in at 0.71%. Organic search managed 0.55%. ChatGPT, which accounted for 82.3% of the AI traffic in the sample across 116,011 sessions, ran slightly ahead of the AI average at 2.08%.

Put the two findings side by side and the problem states itself. The traffic most likely to lose its label is the traffic most likely to convert. Roughly seven in ten of those sessions land in Direct, which drags the Direct conversion rate upward for reasons that have nothing to do with people typing your URL from memory. Direct stops being a channel and becomes a holding pen, and it is a holding pen with an unusually good conversion rate glued to it.

This is a different claim from the familiar one about AI referrals converting well, which we have covered when comparing LLM referral traffic against paid search on conversion quality. The familiar claim is about a channel deserving more credit. This one is about contamination. Your Direct number is a blend of two populations with wildly different behaviour, and no amount of staring at the blended figure separates them.

How AI Traffic Attribution Error Reaches Your Bidding

Here is the step most teams skip. Ask what reads your conversion data besides a human, and the answer is: nearly everything expensive. Smart Bidding consumes conversion events with their attributed sources. Value-based bidding weights them. Customer match and similar-audience modelling build seed lists from converters. Every one of those systems treats the channel label as ground truth, because from inside the platform there is nothing else to treat as ground truth.

Feed that machinery a table in which your highest-intent visitors are stamped Direct, and three things follow mechanically. Paid search gets benchmarked against a Direct cohort that looks better than it is, so the relative case for paid weakens on paper. Conversions that AI discovery genuinely originated get credited to a bucket no campaign can bid toward, so the budget that should chase that demand has nowhere to go. And audience modelling trains on a converter population whose defining shared trait, arrival via an AI assistant, is invisible in the data used to describe them.

None of those rows is exotic. They are the ordinary symptoms of a measurement system whose categories no longer match reality, which is the same failure mode we described when reporting redundancy stops covering for missing analytics data. The difference is that a wrong report costs you a wrong opinion, while a wrong bidding signal costs you money continuously and quietly, in amounts nobody reviews because the line item looks normal.

AI Max Raised The Cost Of Being Wrong

Timing is what turns this from a slow leak into something worth handling this quarter. Google began automatically upgrading eligible Search campaigns to AI Max in September 2026, ended new Dynamic Search Ads campaign creation in the same month, and pushed the DSA sunset out to February 2027. Google's own framing, in its announcement of the upgrade, cites an average of 7% more conversions or conversion value at similar CPA or ROAS when the full feature set is used.

Take that claim at face value, because the argument does not need it to be false. AI Max widens query matching, customises ad text and expands final URLs, which means more of the decision about which searches you buy moves from your keyword list into the model. That is a reasonable trade when the model is well fed. The problem is purely one of sequencing. Discretion moved toward the optimiser in the same twelve months that the quality of the conversion data underneath it degraded. Those two trends are unrelated in origin and compounding in effect.

A system with more latitude and worse inputs does not fail loudly. It fails by confidently pursuing the wrong thing, and it reports good numbers while doing it, because the numbers it reports are computed from the same mislabelled data that caused the error. This is why we push clients on measurement hygiene before expanding automation inside paid search programmes rather than after. Handing more control to an optimiser is a bet on your data quality, whether or not anyone frames it that way at the time.

The sceptical response deserves an answer: at 0.5% of sessions, is this large enough to matter? On volume, no. On influence, yes, and volume is the wrong lens. Bidding systems weight by conversions, not sessions. A source converting at roughly three times the site average contributes to the converter population far out of proportion to its traffic share, and it is precisely that population that seeds audience modelling and shapes value-based bidding. A small, dense, high-intent segment in the wrong bucket distorts more than a large, inert one.

Fix The Signal Before You Fix The Dashboard

The instinct is to build the AI channel grouping first, because it is visible and it demos well. Do the unglamorous thing instead. Establish how much of your Direct traffic is not Direct, then decide what to do about the systems reading it.

One honest caveat about all of the figures above. Orbit Media's sample is B2B and lead-generation sites with at least 100 AI-referred sessions, which selects for organisations already visible to assistants. Loamly's sample is its own detection traffic. Neither describes the whole web, and if you sell into B2B SaaS buyers your numbers will run ahead of both, while a local services business may see almost nothing. Use these as a shape to test against, not a benchmark to grade yourself on. We make the same argument about engagement metrics when the mean and median of AI traffic disagree, and the discipline is identical: the pattern travels, the magnitude does not.

Do this next. Open your analytics, filter Direct to new users only, and pull the top twenty landing pages for that segment over the last ninety days. If those pages are deep, specific and the sort of thing an assistant would cite rather than pages a person would bookmark, you have found the contamination and you can size it in an hour. Then look at the conversion rate of that segment against the rest of Direct. If it is materially higher, stop treating this as a reporting backlog item. The gap you just measured is the size of the lie your bidding systems are currently being told, and it has been compounding since long before anyone put it on a dashboard.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.