94.8% of AI Search Sessions Never Click. Your Attribution Model Has Not Adjusted.

A Bocconi University study using Comscore U.S. desktop clickstream data found that ChatGPT produces outbound clicks in only 5.2% of conversation sessions. 94.8% of the time, the user gets what they need and never visits your site. Your attribution stack was built to trace what happens after a click. The click is disappearing. And most measurement teams have not updated a single instrument.

The 5.2% Finding and Why It Matters More Than the Headline

The study, published July 2026 by researchers Qiaoni Shi, Kai Zhu, and Kai Gu, did something most attribution commentary does not: it used actual clickstream data. Not surveys. Not self-reported estimates. Comscore desktop panel data comparing ChatGPT and Google information-seeking sessions side by side.

Three findings matter for anyone running a brand.

First, the click rate. ChatGPT produces outbound clicks in 5.2% of sessions. Google's referral ratio is dramatically higher. This is not a marginal difference. It is a different category of behavior.

Second, the clicks that do happen are not a scaled-down version of Google traffic. They skew toward specialized destinations and away from ad-supported sites. The referral stream is not shrinking evenly. It is restructuring.

Third, wider ChatGPT Search access cuts traditional search use by 9.4%. That traffic does not vanish. It moves inside the chat. The information need gets resolved without a referral. The researchers call this "satisfying information needs inside the intermediary." I call it the end of attribution as we practiced it.

Everyone Agrees Attribution Is Broken. Nobody Agrees on the Fix.

I have watched this debate accelerate for the last three months. The convergence is striking. The disagreement on what to do about it is more striking.

SegmentStream published data showing last-click attribution undercounts AI search revenue by 5 to 8 times. For B2B and high-value products, up to 15 times. Their fix: identity graphs that stitch cross-device journeys, first-click attribution, and self-reported re-attribution at sign-up.

Search Engine Journal found that 80 to 90% of leads arriving through AI platforms get incorrectly tagged as "organic" or "direct" in HubSpot. Their recommendation: track prompts, monitor cross-model visibility, and add a "How did you hear about us?" field to your lead form. One of their clients discovered 5% of registrations came from ChatGPT despite doing zero AI visibility work.

eMarketer documented the same problem from the retailer side. Global Gravity quantified a 15x gap between Adobe's Q1 2026 finding that AI referral traffic accounts for less than 1% of U.S. retail site traffic and BrightEdge's same-period finding that AI agent activity approaches 15% of total website traffic. Same quarter. Same category. Two numbers that differ by a factor of 15.

Every one of these companies diagnosed the same disease. Every one prescribed a different treatment. That pattern tells you something: they are all looking at the problem from inside the wrong frame.

The Frame Is Wrong. There Are Three Questions, Not One.

The attribution debate will not resolve because everyone is trying to answer three different questions with one metric.

Question one: Am I visible in AI answers? This is the upstream question. Before clicks, before conversions, before revenue attribution, a machine either mentioned your brand or it did not. This is measurable right now. Citation monitoring across ChatGPT, Claude, Perplexity, Google AI Mode, and Gemini tells you whether you are in the answer. The Machine Relations Index measures citation rates by source segment across six AI answer engines. You do not need to guess. You can count.

Question two: Is AI driving consideration? This is the dark-funnel question. Someone read about your brand in a ChatGPT response, then googled your name directly, then filled out a demo form. Your CRM tagged that lead as "organic search." Self-reported attribution closes this gap. SegmentStream and Search Engine Journal both converge on the same prescription here: ask the buyer how they found you. One text field. It is not perfect. But SegmentStream's data shows that without it, the majority of AI-originated leads disappear into the "direct" bucket.

Question three: Is AI search converting to revenue? This is the only question traditional attribution was built to answer. And it still works for this layer, provided you expand the attribution window to 60 or 90 days and stitch cross-device journeys. The problem is not that the conversion measurement is broken. The problem is that two entire layers of measurement are missing above it.

Three layers. Three instruments. The industry keeps debating which single instrument to use. That is why the debate never resolves.

What to Build This Week

Stop waiting for the attribution vendors to figure this out. Here is the measurement stack that works right now.

Layer one: citation monitoring. Set up tracking for your brand across ChatGPT, Claude, Perplexity, Google AI Mode, and Gemini. You need to know whether your brand is in the answer before you can measure whether it drove a conversion. If you are not in the answer, no attribution model on earth will find you, because there is nothing to attribute. Tools like Microsoft Clarity's AI Citations dashboard, Cloudflare's AEO dashboard, and the Machine Relations Index serve this layer.

Layer two: self-reported attribution. Add "How did you hear about us?" to every lead form. Free text, not a dropdown. The buyer's own words reveal channels that no click tracker can see. Route the responses into your CRM so you can track them to pipeline and closed revenue. This is the fastest single move you can make, and both SegmentStream and Search Engine Journal recommend it independently.

Layer three: cross-device attribution. Extend your attribution window to at least 60 days. Stitch mobile research sessions to desktop conversions. Use first-click credit on identity-resolved journeys so the upstream AI touchpoint gets recognized, not the brand search that followed it. This is the expensive layer. It also matters least until layers one and two are in place.

Build from the top. Visibility first, then consideration, then conversion. Most brands are trying to measure conversion from a channel where 94.8% of sessions never produce a click. That is not an attribution problem. That is an instrument problem.

FAQ

Is last-click attribution dead for AI search?

Last-click still works for channels where clicks are the primary user action. For AI search, it is structurally wrong. The Bocconi study shows that 94.8% of ChatGPT sessions never generate an outbound click. When the user gets the answer inside the chat, there is no click to attribute. SegmentStream's cross-customer data shows that proper measurement reveals 5 to 8 times more AI search revenue than last-click reports.

Should I stop tracking citations if they are not a revenue metric?

No. Citations measure the visibility layer. If your brand is not cited in the AI response, it cannot drive consideration or conversion. Self-reported attribution measures consideration. Revenue attribution measures conversion. You need all three. The mistake is treating any one of them as the whole picture.

How much is AI search actually contributing to revenue?

Most brands dramatically undercount it. SegmentStream reports that AI search contribution runs 5 to 8 times higher than last-click shows, and up to 15 times higher for B2B and high-value products. Search Engine Journal found that 80 to 90% of AI-originated leads were being misattributed as organic or direct. The revenue is there. The instruments are not pointed at it.

What is Machine Relations and how does it connect to attribution?

Machine Relations is the discipline of managing how AI systems perceive, evaluate, and represent your brand. The Machine Relations Index measures citation rates across AI answer engines. It serves the visibility layer of the measurement stack described above. Attribution measures what happens after the machine decides to cite you. Machine Relations measures whether the machine cites you at all.