TL;DR
AI referral clicks are the worst available proxy for AI visibility, because every system that counts them is blind in the same place: assistants running in native apps send no referrer. GA4 files that traffic under Direct. Cloudflare says the same gap may be inflating its published crawl ratios. Report three things instead — whether you are named in decision prompts, whether crawlers can reach your commercial pages, and what buyers say when you ask them directly.
Two measurement systems arrived in 2026 that let ordinary teams quantify AI search for the first time. Google Analytics got a native AI channel. Cloudflare put crawl-to-referral ratios on a public dashboard. Both are useful. Both are quoted constantly.
And both are broken in exactly the same way, for exactly the same reason, which almost nobody reporting these numbers mentions.
The referrer problem sits underneath everything
When you click a link on the open web, your browser usually sends a Referer header telling the destination where you came from. That header is the entire basis of referral analytics.
AI assistants running inside native mobile and desktop apps frequently do not send it.
This is not a theory. It is stated in the methodology note of Cloudflare's own announcement of the crawl-to-refer metric: traffic referred by Claude's native app carries no referrer header, they believe the same applies to other providers' native apps, and because the referral side of the calculation therefore only counts web-based tools, the published ratios may overstate the real gap by an amount they cannot quantify.
Read that again, because it has two consequences that point in opposite directions.
On the analytics side, a buyer who reads about you in ChatGPT's phone app and taps through arrives at your site anonymously. Google Analytics has nothing to classify the session by, so it lands in Direct next to people who typed your URL from memory. Your AI traffic is undercounted.
On the crawler side, the referral half of the crawl-to-refer ratio is missing those same visits. So the ratio looks worse than reality. AI extraction is overcounted.
One header, two distortions, in opposite directions. Any measurement plan built on referral clicks inherits both.
What GA4's AI Assistant channel does
On 13 May 2026, Google added AI Assistant as a default channel group in GA4. When Analytics detects a referrer it recognises as an AI assistant, it writes ai-assistant to the medium dimension, (ai-assistant) to the campaign dimension, and groups the session under AI Assistant in Default Channel Group reports. You will find it under Reports, then Acquisition, then Traffic acquisition. Nothing to configure.
This is a genuine improvement. It is also incomplete in four specific ways worth knowing before you build a report on it.
| Limitation | What it means for your numbers |
|---|---|
| No backfill | Default channel groups apply going forward only. Everything before the rollout stays in Referral, so any year-on-year comparison crossing that date is measuring a reclassification, not a trend. |
| AI Overviews are not included | Clicks from Google's own AI Overviews and AI Mode arrive via Google Search and count as Organic Search. The largest AI surface most companies face is not in the AI channel at all. |
| Referrer-less traffic still lands in Direct | The native app problem above. The channel is a floor on your AI traffic, never a ceiling. |
| The recognised list is not published | Google named ChatGPT, Gemini and Claude as examples and has not published the full set of referrers it matches. You cannot audit what you cannot see. |
The fix for the first and fourth is the same: keep a custom channel group running alongside the default one. Google's own documentation on custom channel groups walks through it — Admin, then Data display, then Channel groups — and the one step people miss is ordering. Your AI channel has to sit above Referral in the list, or sessions match Referral first and never reach it.
Nothing fixes the second and third. Those are properties of how the web works, and you should stop expecting a dashboard to solve them.
Crawl-to-refer: a good diagnostic, a terrible headline
Cloudflare's metric divides the HTML page requests a platform's crawlers make against the HTML requests arriving with that platform in the referrer. In their original published example, covering 19–26 June 2025, the range ran from Anthropic at roughly 70,900:1 down to Mistral at 0.1:1, where Mistral was sending ten referrals for every page it crawled.
Now watch what happened to that number as it travelled. Searching for Anthropic's current ratio today returns 2,237:1, 4,580:1, 10,300:1 and 23,951:1 across four different marketing blogs, each citing Cloudflare, each covering a different window, and not one of them carrying Cloudflare's own caveat that the figures may be overstated.
This is what a citation chain looks like when nobody opens the source. If you want the number, read it off Cloudflare Radar yourself, note the window, and quote the window alongside it.
Used properly, the metric is genuinely useful, but as a diagnostic on your own logs rather than as an industry statistic. Compute it per operator on your own server data over a fixed window. A high ratio for a training-focused crawler is expected and mostly uninteresting. A crawler fetching thousands of your pages while its assistant never names you is a different signal: you are being read and not retained, which is a positioning and authority problem, not a crawling one.
The three layers that actually survive
Strip out everything the referrer problem corrupts and three measurable things remain. This is the stack we run, and the order matters.
Layer 1 — Presence: are you named?
The question that decides revenue is whether an engine names you when a buyer describes their problem. Not whether it links you. Ray's research this year found engines routinely cite a page as a source while recommending a different brand in the answer, which we covered in the agency scorecard piece. A citation you can measure in your logs and a recommendation you cannot are not the same event, and the second one is the one that closes deals.
Build a fixed prompt set of 20 to 40 questions in the language your buyers actually use — problem statements, category comparisons, "best tool for" phrasings, and your brand against each named competitor. Run it on a schedule, on the engines your market uses. Record whether you were named, in what position, and who was named instead. Any monitoring tool does this; so does a spreadsheet and a disciplined half-day each month.
Treat every reading as a dated observation, not a ranking. Answers vary between runs on identical prompts, so a single check tells you almost nothing and a twelve-week trend tells you most of what you need.
Layer 2 — Retrieval: can they reach you?
A page that cannot be fetched cannot be cited, however well it is written. This layer lives in server logs, not analytics.
Check which AI user agents are hitting you, which pages they take, and — the part teams skip — which of your commercially important pages they never touch. Confirm nothing in your robots.txt, WAF or bot rules is blocking assistants you actually want traffic from, and confirm your decision-stage pages are reachable without executing JavaScript. This is unglamorous and it is frequently the entire problem.
Layer 3 — Demand: did it change behaviour?
Here is the uncomfortable part. Most AI influence produces no click at all. A buyer reads a comparison in ChatGPT, notes two names, and searches one of them in Google an hour later. The AI did the work. Google gets the credit. Your dashboard shows nothing.
So measure the shadow rather than the click. Watch branded search volume in Search Console, direct session volume, and the ratio of branded to non-branded queries. When AI visibility improves, these usually move before referral traffic does.
See which prompts name your competitors
Our private teardown runs three Google searches and three AI buyer questions in your category, and shows who gets named instead of you.
Request the teardown ↗The thing that beats every dashboard costs nothing
Add a "How did you hear about us?" field to your demo and contact forms, as free text rather than a dropdown.
That is it. That is the single highest-value AI measurement you can implement this quarter.
Self-reported attribution is the only method that captures zero-click influence, because it asks the one system that actually observed the whole journey: the buyer. It catches the ChatGPT conversation that led to a branded Google search three days later, which no analytics platform on earth can reconstruct. Free text matters — a dropdown only finds the answers you already thought of, and the useful signal here is the phrasing people volunteer unprompted.
It is unfashionable, it is not a chart, and nobody selling a subscription will lead with it. Run it for a quarter and read the answers verbatim.
What belongs in the monthly report
| Layer | What you report | Where it comes from |
|---|---|---|
| Presence | Named in X of Y tracked prompts, with the competitors named instead | Prompt set, run on a schedule |
| Retrieval | AI crawler coverage of commercial pages, and any blocked or unreachable ones | Server logs |
| Demand | Branded search volume, direct sessions, branded-to-non-branded ratio | Search Console and GA4 |
| Self-reported | Count and verbatim quotes of buyers naming an AI tool | Form field |
| Referral | AI Assistant channel sessions, footnoted as a known undercount | GA4 |
| Organic | Rankings on commercially important terms | Rank tracking |
That last row is not filler. Organic visibility and AI citations move together — Ray's February 2026 analysis found that sites losing organic rankings also lost citations across AI Mode, Gemini and ChatGPT, with ChatGPT more tightly coupled to Google's results than Google's own Gemini. If your AI numbers move while rankings sit flat, check your measurement before you celebrate. We set out why the two belong in one programme in GEO vs SEO.
Four numbers to stop reporting
- AI referral traffic as a headline KPI. Structurally undercounted, small for most B2B companies, and it measures leakage from AI answers rather than influence inside them. Keep it as a footnote.
- Composite "AI visibility scores". Every vendor defines theirs differently and none publishes the weighting. The number is not comparable across tools, across competitors, or across a tool's own version changes. Report the components.
- Share of voice with no prompt list. A share-of-voice figure is only as meaningful as the questions behind it. If the prompt set is not published in the report, the percentage is decoration — and a prompt set quietly tuned toward flattering questions will produce a lovely chart.
- Crawl-to-refer ratios quoted from a blog post. Either pull it from Radar with the window attached, or compute it on your own logs. Repeating a third-hand figure means inheriting a caveat you never read.
Frequently asked questions
Why is my AI traffic showing as Direct?
Assistants running in native apps generally strip the referrer header, so GA4 has nothing to classify the session by and files it under Direct. It is a property of how referrers work rather than a tagging mistake, and no configuration on your side removes it.
Should I still build a custom channel group?
Yes. The native channel does not backfill, so a custom group is the only way to reclassify sessions from before the rollout, and it catches sources Google's undisclosed list may miss. Order it above Referral or it will never match.
How often should I run a prompt set?
Monthly is enough for most categories, weekly if you are in an active competitive fight. Answers vary run to run on identical prompts, so read the trend across at least a quarter and resist reacting to single readings.
Do I need a paid tool for any of this?
Only for convenience. Layer one can be run manually, layer two lives in logs you already have, layer three is Search Console plus a form field. A tool saves time once the programme exists; it does not create one. We covered when a dashboard is the right purchase in the agency comparison.
Get the measurement baseline before you buy a dashboard
Six searches, one category, one commercially useful diagnosis of where you stand.
Request the teardown ↗Sources: Cloudflare, 1 July 2025, for crawl-to-refer methodology and the native-app referrer caveat. Google Analytics Help for custom channel groups. GA4 AI Assistant channel released 13 May 2026 per Google's Analytics Help Center release notes. Lily Ray, February 2026, on organic and AI citation correlation. Figures verified August 2026 and subject to change.