Field Manual · 15 min read

The Multi Touch AI Attribution Model: Crediting a Citation That Closes Months Later

First touch attribution credits the AI citation that opened a deal and nothing that follows it. Here is the anchor and decay model that keeps that citation solvent through the months of search, email and direct traffic that usually come after it, plus the CRM fields it needs.

Why crediting only the first AI citation stops working

First touch credits the citation that opened the deal and nothing else, which was fine while the only question was whether AI referral traffic existed.

Our AI Search Attribution Handbook deliberately scoped itself to first touch. That was the right call at the time: the open question in September 2026 was whether an AI citation could be traced to a CRM record at all, and first touch is the simplest version of that proof. Get a field onto the record, populate it once, and the AI influenced share of pipeline becomes a number a forecast review can use.

First touch has a cost, and the handbook named it without solving it: once a deal does not close on the day of the citation, which is the normal case rather than the exception, every touch that happens after the citation gets zero credit. The AI citation gets all of it. That is not a measurement of how much the citation actually did. It is an artefact of which touch happened to be logged first.

According to 6sense's 2025 B2B Buyer Experience Report, published November 2025, the global average B2B buying cycle ran 10.1 months in 2025, down from 11.3 months in 2024. A citation that opens a deal in month one and a deal that closes in month ten or eleven are not the same touch doing all the work. Search visits, email opens, a demo, a proposal and a renegotiated price all happen in between, and a first touch model has no way to say any of that mattered.

This is also where the rollout playbook for an AI influenced CRM field stops short deliberately: it covers how the field survives sales ops ownership and a CRM cleanup, not how to credit a second, third or fourth touch once the first one is safely recorded. That is the gap this guide exists to close.

The standard multi touch models, and why none of them fit an AI citation cleanly

Linear, time decay, U shaped and W shaped models all exist and all solve the all or nothing problem, but each one either buries an early AI touch or ignores it.

Multi touch attribution is not a new idea. HubSpot's marketing blog and Dreamdata's attribution documentation both set out the same small family of models, built long before an AI citation was a touch anyone needed to credit:

U shaped and W shaped get closer, because both reserve a fixed share for the first touch rather than letting it decay away. The problem is the word fixed. A flat 40% or 30% was set for marketing journeys in general, not calibrated to a channel that, according to our piece on AI referral conversion rates, already converts at a meaningfully higher rate than a standard search visit before any attribution model touches it. A model built for an average channel is the wrong starting point for one that is not average.

None of the four models above were built with AI referral traffic in mind, and that is not a criticism of HubSpot or Dreamdata. Both are naming the standard toolkit accurately. The gap is that nobody has yet adapted that toolkit to a touch with this shape: rare, usually first, unusually high intent, and increasingly likely to be the reason a buyer started looking at all. As far as we could establish, no analyst firm or attribution vendor has published a model built specifically around that shape as of October 2026. That absence is the reason this guide exists rather than a link to someone else's framework.

The anchor and decay model: a floor credit for the touch that started it

The AI citation keeps a fixed floor of credit no matter how many touches follow it, the closing touch keeps a second floor, and everything in between decays normally.

The model we use internally borrows the best idea from U shaped attribution, a protected floor for the touch that matters most, and fixes the part that breaks: a single flat percentage applied to every deal regardless of how many touches actually happened in between.

Four rules, in order:

1. The anchor touch, the first logged touch of the deal, keeps a floor credit that nothing downstream can reduce. We use 20%. It can move for a specific business, but it has to be a number someone decided on, not zero by default the way time decay leaves it.
2. The closing touch keeps its own floor, separate from the anchor's. We use 25%. The rep call, demo or proposal that actually closes the deal did real work too, and a model that only protects the opening touch just moves the all or nothing problem to the other end of the journey.
3. Everything between the anchor and the close splits the remaining credit by standard time decay. More recent middle touches get a larger share of what is left than older ones, exactly as time decay already does, because that part of the standard model was never the problem.
4. If the anchor touch and the closing touch are the same touch, that single touch takes both floors. A deal that closes on the session the AI citation produced keeps the AI citation's full credit rather than being penalised for happening in one step.

This is a position we are stating plainly as ours, not a published standard. It is built to solve the specific failure named in the previous section: a protected floor that does not depend on the deal having exactly one middle touch or exactly the shape HubSpot's and Dreamdata's models assume. The AI Search Schema Markup Playbook and the AI Crawler Access Playbook both make the case that getting cited is the hard, structural part of this problem. The anchor and decay model is our answer to the question both of those guides leave open: once the citation worked, how do you keep it from being written out of the record it earned.

A worked example: a 10 month deal with one AI citation at the start

Plain time decay reduces a day zero AI citation to roughly 2% of a closed deal's credit. The anchor and decay model keeps it at a fixed 20%.

Take a deal that opens with a ChatGPT citation and closes 300 days later, consistent with the 10.1 month average cycle 6sense reported for 2025. Five touches get logged along the way.

Both columns sum to 100% of the deal's credit. The difference is where that credit lands. Plain time decay hands more than half the deal to the touch that happened to close it and reduces the citation that opened the deal ten months earlier to a rounding error. Report that number to a forecast review and AI search looks like it barely influenced the deal at all, which is not what happened.

The anchor and decay model still rewards the touches that did real work close to the close date, the direct visit and demo at day 210 keep the largest single share in both columns, but it stops the model from quietly deleting the touch that started the deal in the first place. That 20% floor is also the number that should show up in the AI influenced pipeline share calculation from the quarterly audit in the Attribution Handbook, replacing the all or nothing first touch figure that calculation currently produces.

What the CRM actually needs to hold this, not just one field

A single first touch field cannot run this model. The record needs a repeatable touch list, each entry carrying a channel, a timestamp and a source detail.

The Attribution Handbook's CRM step added one field: an AI influenced source flag, populated once at first touch. That field still matters and should stay. It is not enough on its own to run a model that needs to see every touch in the journey, not just the first.

What the record needs instead is a repeatable list, sometimes called a touch log or an engagement timeline depending on the CRM, where every tracked touch adds an entry rather than overwriting the last one. Each entry needs three things at minimum: the channel (AI assistant, organic search, email, direct, paid), a timestamp, and a source detail (the specific assistant domain, the campaign, the referring page). How to get cited by ChatGPT, Perplexity and Google AI and what GEO actually covers are both upstream of this step: they are about earning the citation this log is built to record.

Two implementation details matter more than the schema itself:

• Write to the log automatically, from the same server side match described in the Attribution Handbook. A touch log a rep has to fill in manually will have the same decay problem every self reported CRM field has, which why most CRMs become graveyards already covers in general. A log nobody trusts cannot feed a crediting model anybody should trust either.
• Calculate the split once, at closed won, and store the result. Recomputing the anchor and decay split on every dashboard refresh is wasted work and risks the split changing retroactively as new touches log against an already closed deal. Freeze it at close, the same way a forecast snapshot gets frozen at the end of a review cycle.

The HowTo box below is the build order for adding this without a CDP, assuming the AI influenced field from the Attribution Handbook is already in place.

Why this cannot be built on GA4's channel label alone

GA4's AI Assistant channel does not recognise Google's own AI Overviews traffic and is inconsistent on Perplexity, so the touch log has to run on a CRM tag, not the session label.

GA4 added a native AI Assistant channel to its Default Channel Group on 13 May 2026, which the Attribution Handbook covers in full as the referral layer of its three layer model. It is a genuinely useful channel, and it is not a complete record of every AI referral touch a deal will have.

eSEOspace's own breakdown of the channel, published 13 October 2026, confirms that the AI Assistant channel excludes Google's own AI Overviews and AI Mode by design. Those sessions still report as ordinary Organic Search, indistinguishable from a typed in search, because Google's AI Overviews pass the same referrer a standard search result does. Coverage of Perplexity is inconsistent for the same structural reason: a Perplexity click carries a perplexity.ai referrer, which several independent write ups of the channel describe landing in generic Referral rather than the named AI Assistant channel, because Google has not published the full list of referrers the channel recognises.

None of this is a reason to distrust GA4's channel. It is a reason the touch log in the previous section cannot be built by reading GA4's channel field and calling it done. A deal where the opening touch was an AI Overviews citation will show Organic Search in GA4 and nothing in the native AI Assistant channel at all, which is exactly the gap the server side domain match from the Attribution Handbook was built to catch. The anchor and decay model needs the touch log fed from that match, not from the channel label, or the anchor touch itself will sometimes be invisible to the model meant to protect it.

The mistakes that undercount or over credit the AI touch

The two most common failures are applying time decay without a floor, and crediting a touch log that cannot tell an automated event from a human one.

Four failure modes show up repeatedly once a team moves from first touch to a multi touch model, and three of them are easy to avoid once named.

1. Applying plain time decay and assuming the problem is solved because it is no longer all or nothing. The worked example above shows why: time decay alone still reduces an early AI citation to a sliver of credit, it just does so gradually instead of instantly.
2. Letting the anchor floor and the closing floor overlap badly on a short deal. On a deal with only two or three touches total, a 20% anchor floor and a 25% closing floor can leave very little for a genuine middle touch to be credited with. Shrink both floors proportionally on short deals rather than applying the same fixed numbers regardless of how many touches exist.
3. Counting an automated event as a touch with the same weight as a human one. A marketing automation platform re-sending the same nurture email three times is not three times the signal of one email a prospect actually opened and clicked. Weight touches by evidence of genuine engagement, not by how many rows the automation platform logged.
4. Running the model off GA4's channel field instead of the CRM touch log. The previous section covers why this specific substitution breaks quietly rather than loudly: the model keeps running, the numbers keep looking plausible, and the AI Overviews citations that opened some of the deals are simply never in the data at all.

Why SEO is not dead in the GEO era, the content type header that makes a page invisible to AI search and why llms.txt does not earn citations all sit upstream of this guide: they are about earning the citation the touch log is built to record and credit. The GEO audit checklist and how to become the name an AI recommends cover the same ground from the visibility side. None of that work pays off in a forecast review if the crediting model throws the result away once the deal takes ten months to close.

Frequently asked questions

What is multi touch AI attribution?

It is a way of crediting an AI citation for its share of a closed deal without ignoring the search, email and direct touches that happen between that citation and the close. First touch models credit the citation once and nothing afterwards; multi touch models split credit across the whole journey.

Why does first touch attribution undercount AI citations over a long sales cycle?

It does not undercount the citation itself, it overcounts it relative to everything after. First touch gives the opening AI citation all the credit and every later touch none, which looks generous to AI search until the deal takes months to close and a rep, a demo and a proposal did real work in between.

Why not just use a standard time decay model instead?

A standard time decay model weights each touch by how recently it happened before close, which systematically shrinks the oldest touch in the journey. An AI citation that opens a deal is, by definition, the oldest touch, so plain time decay can reduce it to a sliver of credit even when it was the reason the buyer started looking in the first place.

What CRM fields does a multi touch AI attribution model actually need?

A single first touch field is not enough. The record needs a repeatable touch list, each entry carrying a channel, a timestamp and a source detail, written automatically at the point each touch happens rather than reconstructed later from memory or a single dropdown field.

Does GA4's AI Assistant channel capture everything this model needs?

No. GA4's native AI Assistant channel recognises referrer headers from assistants such as ChatGPT and Gemini, but it does not include Google's own AI Overviews traffic, which still reports as Organic Search, and coverage of Perplexity is inconsistent. The model has to be built on a server side or CRM level tag, not the GA4 channel label alone.