Attribution Models for Assistant-Mediated Conversions

A user asks ChatGPT for a recommendation on project management software. Asynchronous return compounds the problem: a ChatGPT conversation stays open indefinitely, so a user who first encountered a recommendation on mobile during a commute may convert on a work laptop three days later after IT approval, with no mechanism connecting the two events. None of that architecture applies to what just happened, because the entire sequence, awareness, comparison, objection handling, and the intent signal itself, occurred inside a single thread that no external analytics tool can see.

Why the multi-touch journey attribution was built to track disappears in AI conversations

AdventureMedia's 2026 analysis gives this phenomenon a name, citing what it calls "industry experts" terming it "conversational continuity": a user engages with a brand mention at message three of a fifteen-message conversation, and the actual decision to convert takes shape gradually across the exchanges that follow, yet the attribution system logs exactly one event, a single "ChatGPT referral".

The asynchronous nature of these conversations makes the problem worse, not better. A ChatGPT thread doesn't expire the way a browser session does.

The deeper issue here isn't just technical. A "touchpoint" was always a proxy, a stand-in for a moment when a brand exerted some measurable influence on a person's thinking. Inside a continuous conversation, that influence doesn't arrive in discrete packets with clean boundaries between them. Multiple moments of persuasion happen back to back with no natural seam the analytics layer can detect. The old proxy doesn't just lose accuracy. It loses the thing it was supposed to stand for.

How the ChatGPT ads system works

Understanding why this gap exists requires understanding how ads actually enter a ChatGPT conversation, because the opacity is a structural feature of how the system is built.

ChatGPT generates its response first, drawing on training data and the context of the conversation so far. Only once that response is finalized does a separate ad-matching process scan the conversation content to look for a relevant promotional opportunity. Advertisers can bid for placement based on topical relevance, but they have no mechanism to influence what the model actually says in its answer.

Targeting itself works differently from search. Advertisers instead write what OpenAI calls "context hints," short prose descriptions of the kinds of conversations where a given ad might make sense. OpenAI states that these hints are not exact-match keywords and come with no guarantee of delivery in any specific conversation. The matching process draws on the current conversation's context and intent, on the advertiser's landing page and ad copy, and, only when a user has opted into ads personalization, on select signals pulled from that user's broader ChatGPT history.

For an advertiser, this means the actual pairing of ad to conversation happens entirely on OpenAI's servers, in a process the advertiser cannot observe or instrument. The advertiser sees the resulting click. The conversation that produced it, the specific exchange that made the ad relevant in that moment, stays invisible.

The audience itself is also narrower than it first appears. Plus, Pro, Business, Enterprise, and Education subscribers see no ads at all. That means B2B buyers, who disproportionately sit on paid enterprise tiers, are structurally missing from the addressable audience, a gap that keyword search advertising, which reaches buyers regardless of subscription tier, simply doesn't have.

The measurement infrastructure to address any of this has been arriving in pieces, after the ad product itself was already live. OpenAI opened a beta for U.S. Free and Go users on February 9, 2026, and followed with self-serve access through Ads Manager on May 5, 2026, dropping the minimum spend commitments that had applied earlier. A Conversions API and pixel-based tracking arrived in May 2026, letting advertisers track landing page views, product views, add-to-cart actions, and completed purchases after the click. Closed-loop attribution through CRM systems was described as coming later in 2026, and HubSpot built the first such integration in September 2026, but other CRM platforms are still catching up. So marketers running campaigns today are building on infrastructure that is still being assembled underneath them.

Conversational Intent Signals

The frustrating part of this gap is that what's hidden inside these conversations is more valuable than almost anything attribution has ever had access to. A conversation is paragraphs of stated reasoning, and the signal quality available inside it is richer than anything a keyword search has ever produced. That's what makes the measurement gap so costly: marketers are losing visibility exactly where buyer intent is most fully expressed, not where it's thinnest.

A prompt can contain a budget ceiling, a timeline, a statement of prior experience with competing products, and a list of alternatives already ruled out, all in a single message, in language a keyword can never carry. A meaningful share of these prompts are transactional on their face, and even the ones that read as purely informational tend to carry comparative framing and explicit constraints that mark out exactly where a buyer sits in the purchase process.

The shape of the journey itself matters for how any new attribution approach should be built. The conversation itself isn't the conversion event. It functions as the high-intent window that precedes conversion, and the click out to the open web is best read as a proxy for a decision that was substantially made before that click ever happened.

The most consequential part of all this may be the simplest. The specific words people use in these conversations function as a leading indicator, surfacing a consumer segment while it's still forming in conversation data, ahead of purchase data, ahead of survey panels, ahead of brand trackers even having a name for the group in question. The conversion path itself may stay untraceable, but AI-conversation presence moves measurable activity in adjacent channels, and that is now documented, not speculative.

Why standard attribution models fail on assistant-mediated conversions

None of the inherited attribution models survive contact with this environment, and each one breaks for a different reason, which matters because any replacement has to be designed against the specific failure, not just a general sense that "attribution is hard" now.

Last-click attribution gives full credit to the click that exits the conversation, treating it as the moment of decision. The Princeton study on commercial persuasion in AI-mediated conversations found that LLM-driven persuasion nearly tripled the rate at which users selected sponsored products compared with traditional search placement. The mechanism doing the persuading is conversational, built across turns, not something that happens at the moment of a click, which makes last-click attribution blind to the exact thing it claims to measure.

First-touch attribution assumes a clean, identifiable first exposure, the way a cold keyword search represents a clear starting point. There's no equivalent of a neutral starting point to anchor the model to.

Linear and time-decay models both split credit across touchpoints that are assumed to be discrete events on separate pages or in separate sessions. Ad impressions, brand mentions, and follow-up questions all sit on one continuous thread, without the event separators either model needs to function.

Position-based, or U-shaped, attribution weights the first and last exposures most heavily. But the model was built for journeys that span multiple sessions, not for a single thread where the first and last relevant moments might be minutes apart and neither one maps to a trackable URL event.

Every one of these models was built to assign credit to observable events happening on properties the advertiser controls and can instrument. The interior of a conversation is not instrumented, and the advertiser has no way to instrument it. Every existing model is therefore working from a dataset that is incomplete by design, not by neglect.

The adapted and emerging frameworks practitioners are using

None of this has left practitioners waiting around for a perfect framework to arrive. Teams are adapting the models they already have and layering new measurement primitives on top, and incrementality testing has emerged as the most defensible near-term answer among them.

One workaround, described in AdventureMedia's analysis, treats conversation-aware UTM parameters as a first-party data layer. Making this work requires real changes to CRM architecture: most existing Salesforce and HubSpot setups don't have the custom field structure needed to store this kind of attribution data today. Even when it's built out, this approach only captures context at the moment of the click. It doesn't reconstruct the full arc of persuasion that happened earlier in the thread.

A second adaptation treats conversation depth itself as a proxy for intent. Some teams now weight position-based attribution so that an ad impression appearing after five follow-up questions counts for more than one surfacing in response to a broad, first-message query. This is a heuristic, not a measured signal, since the advertiser has no direct access to how deep a given conversation actually went. So teams are left inferring that depth indirectly, from UTM data or from post-click surveys.

A third direction moves the problem upstream, into bidding itself, rather than trying to solve it after the fact. Concepts like Cost Per Engaged Conversation, which charges an advertiser only when a user interacts with an ad and keeps the dialogue going, and Conversation Value Optimization, which adjusts bids based on the predicted lifetime value of users who engage in particular conversational patterns, treat attribution as a real-time bidding signal rather than something measured after the sale. OpenAI's conversion-optimized cost per click option is the first live version of this logic in production.

The most credible framework in use today doesn't try to reconstruct the individual path at all: incrementality testing, run through geo-holdouts and matched-market comparisons, accepts that the path from conversation to purchase is often invisible, so it models the data probabilistically instead of tracing it event by event. The logic is straightforward: if a market exposed to conversational AI ads converts at a meaningfully higher rate than a matched control market that wasn't exposed, that gap is attributable to the channel, even without ever reconstructing a single user's path. That causal credibility is something no multi-touch model built for this environment can currently match. The tradeoff: incrementality tests run slowly, cost money to execute well, and can't optimize an individual campaign's creative or targeting in real time. They answer whether the channel works, not which ad or audience within it is working.

A related approach works at the segment level rather than the individual level. Verve built what it describes as the first integration of intent signals from AI-native conversations into an intelligence layer alongside search data and zero-party polling data, drawing on users who have opted in to share AI chat activity through apps they already use, which Verve then pseudonymizes. That structure accepts that tracing any one person's journey isn't solvable right now and instead builds for aggregate signal quality that advertisers can act on across channels.

What remains unsolved in conversational measurement infrastructure

The honest conclusion is that better pixels and smarter UTMs, bolted onto an attribution stack built for the search-and-click era, won't close this gap. You need measurement infrastructure built around the conversation itself, treating the thread as the unit of analysis instead of the click.

Several parts of this problem remain unresolved today, a matter of missing infrastructure rather than missing effort. The path from an AI conversation to an eventual purchase is frequently untraceable. Fairing's 2025 analysis found that only about 15% of brands using its post-purchase survey tool have recorded even a single customer who voluntarily named an LLM as part of their purchase journey. Most AI-influenced conversions leave no self-reported trace at all. The data marketers do get access to is aggregated and modeled in probabilities, a different kind of evidence than the deterministic event logs that search attribution was built around. CRM integration for closed-loop attribution inside ChatGPT Ads isn't available yet; OpenAI has described it as coming later in 2026.

There's a second layer to this that complicates even the workarounds built on self-reported data. The Princeton research on commercial persuasion found that explicit "Sponsored" labels don't meaningfully reduce how persuasive an ad turns out to be, and that most users fail to notice any promotional steering at all. If people generally can't tell when they've been influenced, then survey data and post-purchase questionnaires asking them to recall that influence are unreliable by the same logic, not just incomplete.

The academic framework describing how LLM advertising actually functions names the structural cause: the standard search advertising sequence of bidding, output generation, click-through prediction, and auction assumes an order of operations that conversational systems don't follow. Search ads are selected before a result is shown. ChatGPT generates its answer first and only afterward checks for an ad opportunity. That reversal means every assumption embedded in decades of search-attribution tooling, about when a signal becomes available and in what sequence, has to be rebuilt for an interface where the conversation, not the click, is where the real decision gets made.