How to Attribute a Free Trial Signup to an AI Assistant Recommendation

A product-led growth team can see almost everything: which page a visitor landed on, how long the trial took to activate, which feature drove the upgrade. The one thing it usually cannot see is the moment that started the whole journey, when a buyer asked ChatGPT for a tool in your category and the engine named you. That recommendation leaves no referrer, no UTM, no row in the analytics that any standard model can credit. The signup shows up as Direct, branded search takes the credit, and the AI touch that created the demand goes uncounted. This is how to attribute that trial signup properly, without pretending you have data you do not.

Why the AI touch disappears before the trial starts

The problem is structural, not a tracking bug you can patch. When a buyer reads an AI answer and then navigates to your site on their own, the browser sends no referrer, so the visit lands in Direct. Even when the buyer clicks a link inside the assistant, several platforms strip or fail to pass the referrer header. Analyses of AI-driven traffic in 2026 have found the majority of it arriving with no referrer at all; one study of more than 446,000 visits put the share of AI traffic misattributed as Direct at roughly 70 percent.

For a free-trial funnel, the timing makes it worse. The AI recommendation is almost always a first touch. A buyer asks an assistant, hears your name, reads a page or two, and leaves. Days or weeks later they come back, type your brand into Google, and sign up. A last-click model hands every point of credit to that branded search. The AI answer that put your name in the buyer's head, the actual cause of the signup, is invisible. The mechanics of why AI referrals collapse into Direct, and the GA4-level fixes that recover part of it, are covered in how to track AI referral traffic in GA4 and tie it to pipeline.

So the honest starting position is this: you will not attribute AI-sourced trials with server logs alone. You attribute them by combining a direct question to the user with corroborating signals that make the answer trustworthy.

Signal one: a self-report field built for the trial funnel

The highest-value change you can make is also the simplest. Add a "how did you hear about us" question to the signup flow. But the way most teams implement it throws away most of the signal.

Three design choices decide whether the field works:

• Ask at signup, not in a later email. The intent is freshest at the moment of the trial start. A survey sent three days after activation catches a fraction of users and a fraction of their memory.
• Make it a required, structured field with a specific AI option. A blank optional text box gets skipped or filled with "Google." Offer explicit choices, and include "ChatGPT or another AI assistant" as its own option, not buried under "other" or "search engine." Buyers do not think of ChatGPT as search, so if you make them file it there, they will not.
• Add one open follow-up. When a user picks the AI option, ask which assistant and, optionally, what they asked. That free-text answer is gold, because it tells you the exact prompt that surfaced you, in the buyer's own words.

Self-report is noisy on its own. People misremember, and the last thing they touched (branded search) often overwrites the first thing that mattered (the AI answer). That is why self-report is the anchor of the method, not the whole of it. The complete build for a self-report survey that catches AI-sourced conversions, including question wording and how to handle recall bias, is in how to capture AI-sourced deals with a self-reported attribution survey.

Signal two: corroborate the self-report with behavioral cohorts

A single self-reported answer is a guess. Three signals that point the same way is attribution. The point of corroboration is to check whether the users who told you "AI" behave differently from the rest of your Direct signups, in ways an AI-sourced buyer plausibly would.

When the self-report and the behavior agree, your confidence climbs. If users who claim they found you through an AI assistant also landed on the exact page ChatGPT tends to cite, took longer than average to return, and converted at a higher rate, you are not guessing anymore. You have four independent readings of the same event. If the self-report and behavior disagree, you have found a data-quality problem worth investigating rather than a number to report.

Signal three: read branded-search lift as an AI fingerprint

Branded search is the sneaky one, because it is both the thing that steals the credit and one of the better proxies for AI-driven demand. The logic runs like this: AI answers rarely produce a click, but they do plant a name. A buyer who hears you in ChatGPT does not click through, they later Google your brand. So a rise in branded-search volume, with no campaign, launch, or press to explain it, is a fingerprint of AI recommending you upstream.

To use it, hold a baseline. Track branded-query volume weekly, note every event that could lift it (a launch, an ad flight, a podcast), and watch for increases that have no such cause. When branded search climbs in the same window that your citation share rises in a category, and your self-report AI cohort grows, you have a coherent story: the engine started naming you, buyers started searching you, and some of them started trials. The method for decomposing branded search into baseline demand and AI-induced demand is the harder version of this, and the three-signal approach to proving pipeline when no referrer exists is laid out in how to prove AEO pipeline when the buyer leaves no referrer.

Put it together: a defensible attribution you can bring to a board

None of these signals is proof on its own. The method is to require agreement. Count a trial signup as AI-sourced when the self-report names an AI assistant and at least one behavioral signal corroborates it. Report it in tiers, so a skeptical CFO sees exactly how much confidence each number carries.

• Confirmed AI-sourced. Self-report names an assistant and the behavioral cohort matches. Report as a firm number.
• Probable AI-sourced. Self-report names an assistant but no behavioral corroboration, or strong behavioral signal without a self-report. Report as a range.
• AI-influenced. Branded-search lift and citation gains in the period, but no user-level link. Report as a directional trend, not a count.

This tiering is what keeps the method honest. You are not claiming a precision you do not have. You are showing a board that a real and growing share of trials trace back to AI recommendations, sized conservatively, with the assumptions on the table. For a PLG team specifically, the upstream work of getting the product named in those answers in the first place is the other half of the loop, covered in the PLG demand gen playbook for getting recommended by AI assistants.

The upstream fix: be the answer the engine names

Attribution measures a result. It does not create one. If you want more AI-sourced trials to attribute, the engine has to name you more often, which means the sources it reads have to say the right things about you clearly and consistently. Give crawlers a clean canonical source with a free llms.txt generator, keep your product and pricing facts current where the engines read them with the AI Feed Engine, and treat the whole cycle, from earning the citation to measuring the trial it drove, as one system, which is what how OnlyAEO works is built around. The arc of a product going from unnamed to routinely recommended, and the signup lift that followed, is documented in the FastTrackr AI case study. External practitioners cover the tracking side in depth; see TapClicks on fixing AI referral attribution and Clickport on why ChatGPT traffic shows as Direct in GA4.