Every Attribution Model Breaks in the AI Era

Attribution models are arguments about credit dressed up as arithmetic. Last-touch is not a measurement; it is the assertion that the final interaction deserves the sale. Linear asserts that every interaction contributed equally. Neither claim is derived from data — the data cannot tell you what caused a decision, only what preceded it.

That was tolerable while all four standard models shared an assumption that mostly held: that the touchpoints you can see are the touchpoints that mattered. AI assistants broke that assumption, and each model breaks in its own specific way.

The journey the models were built for

Every standard model was designed around a specific mental picture: a buyer moving through a funnel, encountering your marketing several times, each encounter nudging them along, ending in a purchase. The models differ only in how they distribute credit across that sequence.

That picture assumes three things:

1. The touchpoints are yours. Ads, emails, content, campaigns — things you ran.
2. They are observable. Each leaves a trace in your analytics.
3. Influence roughly tracks the sequence. Earlier means awareness, later means intent.

An AI-mediated purchase violates all three. The decisive event happens inside a conversation you cannot see, was not a campaign you ran, and produces at most a single click — after which the buyer continues researching through channels that are observable, and converts through one of those.

Model by model

Last-touch: erases the channel entirely

The default in most tools, and catastrophically wrong here.

Consider the actual sequence. Someone asks an assistant for a recommendation on a Tuesday, clicks through, reads your pricing page, and leaves. Over the next week they check a review site, ask a colleague, and read a comparison article. On the following Monday they type your brand name into a search box — because that is how people navigate to a site they have decided on — and buy.

The last touch is branded search. Last-touch attribution credits branded search with 100% of the sale.

Branded search did not persuade anyone of anything. It is a navigation event, the modern equivalent of typing the URL. It is the consequence of a decision, not its cause. Yet in most attribution reports it will be the best-performing channel in the business, because it sits at the end of every successful journey by construction.

Last non-direct click: worse than erasure

This is last-touch's more sophisticated cousin, and the default conversion attribution in GA4. Rather than crediting Direct, it walks back to the most recent identified channel.

The intent is sound — a direct visit is often a return by someone acquired elsewhere. The effect on AI traffic is that a referrer-less AI visit does not merely fail to receive credit; the credit is actively transferred to whatever unrelated channel touched the visitor earlier.

This is the mechanic examined at length in why GA4 cannot see your AI traffic, and it deserves restating because of its direction: it does not report AI as unknown. It reports it as organic search. Then you fund more organic search.

First-touch: right instinct, wrong mechanism

If the AI touch is early, surely first-touch attribution captures it?

Sometimes. But the assistant is frequently not the first contact either. The buyer may have seen a conference talk, read a newsletter mention, or landed on your blog through a search months earlier. The assistant is where the decision formed, not necessarily where the relationship began.

First-touch also has an unrelated flaw that becomes severe with long cycles: it credits an interaction that may have happened before your data retention window even opened. Credit assigned to a touchpoint you can no longer inspect is not measurement.

Linear: dilution as a policy

Linear splits credit evenly across every touchpoint. Its appeal is that it refuses to make a claim it cannot support.

The problem is that touchpoint count varies enormously by channel, and email is the pathological case. A buyer on your list might receive fifteen emails during a consideration period, generating fifteen touchpoints. The single AI referral that actually caused the evaluation gets one-sixteenth of the credit.

Linear does not weight by influence. It weights by frequency, and frequency is a property of your own sending cadence rather than of the buyer's decision.

Time-decay: precisely backwards

Time-decay weights touchpoints by recency, on the theory that recent interactions are more influential.

For a channel whose defining characteristic is that it operates at the discovery end of the journey, this is exactly inverted. Time-decay systematically penalises AI referral for the very property that makes it valuable — being the moment someone first learned you were an option.

The better question

The models all try to answer "which touchpoint deserves the credit?" That question has no correct answer, because credit is not a property that exists in the world — it is a convention you adopt in order to make decisions.

A more tractable question: which touchpoint introduced this buyer to the category, or to you as an option within it?

That question has a defensible answer more often than the credit question does, and it maps directly onto a decision you actually have to make: where to invest in being discovered. It also happens to be the question AI referral is uniquely well-positioned to answer, because an assistant recommendation is, almost by definition, an introduction. Nobody asks an assistant for a recommendation and receives a name they had already decided on.

Practically, this suggests a few things:

Report introduction and conversion separately. Two columns, not one blended score. "This channel introduced 40 buyers who eventually spent £120,000" and "this channel was the final click for 90 purchases" are both true, both useful, and describe different jobs.

Treat branded search and direct as navigation, not acquisition. They are the visible surface of a decision that was made elsewhere. Any model that lets them accumulate credit is measuring the buyer's route to the checkout rather than their reason for going.

Keep the raw touch history and apply models at read time. The single most valuable engineering property here is being able to change your mind. A pipeline that collapses history into a single attributed source at write time has permanently destroyed the ability to ask a different question — and given that nobody has settled on the right model for AI-mediated journeys, permanently committing to one now is a bad bet.

What about data-driven attribution?

The obvious rejoinder is that algorithmic models sidestep the whole argument by learning credit distribution from converting and non-converting paths rather than imposing a rule.

They are better, and they do not solve this.

A data-driven model can only distribute credit across the touchpoints it can see. If the AI touch is absent from the data entirely — sitting in Direct, unlabelled — the model does not underweight it. The model has no idea it exists, and will confidently attribute its influence to whichever observable touchpoints happen to correlate with it.

This is the point that matters most in the whole discussion: detection is upstream of modelling. Every argument about which model to use is downstream of whether the touchpoint is in your dataset at all. Choosing a sophisticated model on top of incomplete detection produces sophisticated, precise, confidently wrong answers.

Fix detection first. Then argue about models.