Blog · Measurement

Attribution when there is no click

Published 3 February 2026 · Updated 12 October 2026 · 8 min read

Attribution when there is no click
Photograph: Felipe Correia / unsplash

The short answer: You cannot attribute a no-click answer with analytics, because no request ever reaches your server. What you can do is stop treating the referral number as the channel, measure the influence with three instruments that do not depend on a visit, and accept an estimate with a stated error instead of a false precision. Ask buyers where they first heard of you. Watch branded demand. Read the logs.

The gap is not a tracking bug somebody forgot to fix. It is a property of how the surface works, and knowing the mechanism tells you which parts are recoverable.

What the referrer carries, and what it never carried

MDN documents strict-origin-when-cross-origin as the default referrer policy, and under that default a cross-origin navigation sends the origin only. So when an assistant page links out under the default, your server gets https://chatgpt.com/ and nothing else: no path, no query string. The question that produced the citation was never in the header, so nothing downstream can recover from the header what the browser did not send. A site can set a different policy for its own outgoing links, and we have no way to audit what policy each assistant sets, so treat this as the default case rather than a rule we have confirmed engine by engine.

That has a practical edge. Anyone promising to tell you which prompt drove a session is inferring from the landing page, which is a guess dressed in a dashboard. Treat it as a guess and the rest of their numbers get easier to read.

Two further leaks sit underneath that one. A native desktop or mobile assistant app is not a browser tab and is under no obligation to send a Referer header, and a visit that arrives without one cannot be told apart from a typed URL. GA4's channel definitions say Direct is source exactly (direct) with medium (none) or (not set), so a header-less visit is filed there by rule. The same page now also defines an AI Assistant channel for arrivals from sources like ChatGPT, Gemini, Copilot and Grok, which catches the referrer-bearing ones and leaves out AI Overviews and AI Mode. That helps with the clicks you can see and does nothing for the ones that carry no header. On Google's side, Search Central says an appearance in AI Overviews or in AI Mode counts inside your ordinary Search Console totals, filed under the Web search type in the Performance report, which means the one report most teams trust for search demand does not split out the surface you are asking about.

The failure mode: mistaking the referral count for the channel

Here is the shape of the mistake. A team segments assistant hostnames in analytics, sees a small number, concludes the channel is small, and defunds the work. The segment is real. The conclusion does not follow, because the segment only ever contained the minority of answers that ended in a click.

One engine's version of the scale is measurable. Pew Research Center analysed the browsing of 900 US adults who agreed to share it, a dataset of 68,879 Google searches from March 2025, and reported that users who met an AI summary clicked a traditional result on 8 per cent of visits, against 15 per cent for those who did not meet one. That is Google AI summaries, in one month, among US adults in a volunteer panel. It is not a measurement of ChatGPT, Perplexity, Gemini or Copilot, and the two groups are not the same set of queries, so read it as roughly half the click rate between two populations rather than a before and after on identical searches. What the finding does support is the direction: attention that stays inside the answer stops leaving a row in your analytics.

So the referral count is a floor on influence, never a measure of it. Reporting it as the channel size is the quiet version of the error in How to measure AI visibility without fooling yourself: a number that is easy to collect standing in for one that is hard.

Three instruments that do not need a visit

None of these is precise. Together they bracket the thing.

Self-report at the moment of conversion. One free-text field on the signup or demo form, asked once, worded as how they first heard of you rather than what brought them here today. Free text beats a dropdown, because a dropdown that omits assistants collects none and you learn nothing. Code the responses monthly. This is the only instrument that sees a buyer who read your name in an answer, closed the tab, and came back by brand search four days later.

Branded demand as a leading indicator. Impressions on your exact brand name in Search Console, plus direct sessions that land on a deep page. Those two move when an answer names you and nobody clicks the citation. Pair them with your mention rate from the prompt set so the two series sit on one chart with the same dates.

Server logs for the retrieval side. A request from a documented assistant fetcher user agent proves your page was pulled for a live answer, whether or not a human ever followed. Fetches cluster by URL, and that distribution tells you which pages the systems actually reach for. It is the one signal in this list that is first-party, unsampled and immune to consent banners.

A check you can run this week

Pull ninety days of sessions. Filter to medium (none) and source (direct), then split by landing page into two buckets: the home page and everything else. Direct traffic to a home page is mostly memory, bookmarks and untagged email. Direct traffic landing straight on a deep comparison page, a pricing page or a specific documentation URL is not memory, because nobody types that path. Track the second bucket as its own series, week by week.

Then look at where it sits relative to your named assistant referrers. On the handful of client sites we have looked at during audits, the deep-page direct series has usually been the larger of the two. We are not publishing a figure for that, because it was read off client dashboards one site at a time rather than collected under a method we could hand to an auditor, and a handful of sites is a sample nobody should generalise from. Treat the split as a thing to run on your own data, not as a result of ours.

The same logic underpins Zero click, and what replaces the click: the unit of success moved from a session to a correct sentence with your name in it, and the instruments have to move with it. If you want the manual version of this, the self-check lists ten things you can test without buying anything.

Build the estimate, and show your working

Write the arithmetic down where a finance person can argue with it.

Start with click-path revenue, which is ordinary analytics on the assistant referrers. Then estimate the influenced part: take the self-report share for assistants, apply it to conversions in the period, subtract the ones already counted in the click path so you are not double counting, and value the remainder at your normal conversion economics. Put a range around it, because a survey with a few hundred responses and a response bias you have not characterised does not deserve a point estimate.

That last step is the one teams skip, and skipping it is why these models get dismissed the first time someone senior pokes at them. A number with a stated error survives a meeting. A confident number without one does not.

Two cautions on reading the self-report data, both of which are assumptions we have not measured. Buyers may under-report assistants relative to channels they can picture, which would bias the raw counts downward. And the people who bother to answer a free-text field are not a random sample of buyers. If both hold, they push the same way, and your estimate is likelier to be conservative than inflated. We cannot prove either from our own data, so do not lean on them harder than that.

Where this connects to visibility reporting

Attribution and visibility answer different questions, and conflating them produces the worst reports in this category. Visibility asks how often you are named and what is said. Attribution asks what that was worth. A mention-rate chart is an input to the second question, never an answer to it, which is part of why Share of voice in AI answers, and its limits is worth reading before anyone puts a percentage in a board pack. The broader framing sits in What is answer engine optimisation.

Honest limits

This post does not establish a causal model, and the bracketed estimate it describes is a convention, not a result: a floor from referrals, a ceiling from self-report, and a judgement call between them that you revisit every quarter. The Pew figures are Google AI summaries in one month among US adults, so we are using one engine's click behaviour to argue about a gap that sits on surfaces we have not measured. We have no view inside answer delivery on the assistants themselves, which means the no-click share is something we infer from an absence rather than something we have counted. Our own deep-page direct observation comes from a handful of audits read off client dashboards, with no fixed method, no sample size we can state and no date range we can defend, which is why there is no number attached to it. And an engine can change its linking behaviour at any time, which moves the referral floor without anybody on your side touching the site.

FAQ

Can I see which prompt sent a visitor? No. The browser default sends only the origin on a cross-origin navigation, so the path and query never arrive. Landing-page inference is a hypothesis, not a measurement.

Why do some assistant visits show up as direct traffic? A native app need not send a referrer header, and GA4 defines Direct as source (direct) with medium (none) or (not set), so a header-less visit lands there by rule. Visits that do carry an assistant referrer now have their own AI Assistant channel in GA4's default grouping.

Does Search Console separate AI Overviews and AI Mode? Not as their own rows. Google's documentation states that appearances in AI features are included in overall search traffic and reported within the Web search type.

Is a self-report survey good enough for a board report? It is good enough if you report it as a range and say how many responses it came from. Presented as a precise channel figure it will not survive the first informed question.

Sources

All four were opened and checked on 12 October 2026.

Related reading

How to measure AI visibility without fooling yourselfShare of voice in AI answers, and its limitsZero click, and what replaces the clickWhat is answer engine optimisationProving a change worked: before and after in AI searchBenchmarking your answers against competitors