How do I measure the clicks AI answers send back?
Segment the assistant referrers in your analytics (chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, copilot.microsoft.com) and count the visits. Then put that number next to the AI reads in your server logs, because the ratio between the two describes the channel.
The search-referral decline is real and probably permanent. The answer economy runs a small return channel most sites have never measured.
Why your analytics sees half the funnel
The AI funnel has two hops, and standard analytics misses the first entirely.
- The read. An AI assistant fetches your page to answer a person's question. Bots do not execute JavaScript, so script-based analytics records nothing. Server logs are the only place this hop exists. Ours record more than 200M such fetches a day across 10,000+ sites.
- The click. Some answers cite sources, and some people click them. Those visits arrive as ordinary human traffic with an identifying referrer.
Sites that only look at hop two conclude that AI sends almost nothing. Sites that see both hops can compute reads to referrals: how often being the answer's source becomes a human visit.
What the return channel looks like
From our network logs, 14 days in August 2026, one topic cluster:
- AI assistants referred 1,836 human visits into the cluster's pages. Modest against the old search firehose, but concentrated and observable.
- Referrals track reads. The pages assistants fetched most were largely the pages people arrived on from answers.
- A few pages convert at striking rates. One explainer was fetched 307 times by assistants and received 48 human click-throughs.
- Referred visitors arrive deep, on the page that answered them, not on a homepage.
The ratio to expect
The 307-to-48 page is the exception. Here is the rule:
| Site or region | AI reads to AI referrals |
|---|---|
| tollbit.com, Sep 15 to Oct 14 2026 | about 1,000 to 1 (44K reads, 47 referrals) |
| European sites, Q1 2026 (State of the Bots) | 150 to 1 |
| European sites, Q2 2026 (State of the Bots) | 227 to 1 |
Our own 47 referrals split ChatGPT 37, Claude 7, Perplexity 3, plus 126 from DuckDuckGo with search and assistant mixed. The shape matters more than any one figure: the ratio runs in the hundreds or thousands, and in Europe it widened from Q1 to Q2.
Why referrals are a trailing, shrinking metric
A referral requires three things in sequence: the assistant read your page, cited it, and the person clicked. Citations are on the order of 5–15% of what an assistant reads, and advanced agents often cite nothing. Each step loses most of the volume, so the click count is a trailing echo of the reading behind it.
It also shrinks as agents do more. A computer-use agent that compares products, fills a form or books a table completes the task in place, and nobody clicks because the person never saw a link. In September 2026, Resy deactivated accounts whose personal assistant was hitting the site about 200 times an hour. Those were visits with intent, and none would show up as a referral. Cloudflare's CEO put automated traffic at about 60% of requests in May 2026. The reads grow. The clicks do not.
Treat referrals as the confirmation metric and reads as the leading one.
The measurement playbook
- Instrument the referrers today. Segment the assistant domains in your analytics. It costs nothing.
- See your reads. Analytics tools cannot. Server logs can. Until you can see what AI fetches from you, the referral number has no denominator.
- Be the fetched source. Only retrieved pages can be cited, and only cited pages can be clicked. Answer-first pages with clean structure win retrieval.
- Judge the channel on its own economics. AI-referred visits are fewer than search referrals were, and they land with intent on exact-match pages. Measure their conversion separately.
- On Google, AI Overviews attribution is limited and evolving. Measure the assistants that identify themselves cleanly.
TollBit Analytics shows both hops from the same forwarded logs: reads by agent and page, blocked fetches, and the referrals that came back.
Frequently asked questions
Which referrer domains identify AI assistants? The set worth segmenting now: chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, copilot.microsoft.com. The list grows as assistants launch browsing surfaces, so review it quarterly.
Is the click decline AI's fault? Independent studies consistently find reduced click-through when an AI answer appears, and the trend is not in serious dispute. The strategic question is presence in the answers, plus capture of the clicks they still send.
Can I get AI traffic back by blocking bots? Blocking removes you from answers entirely. Reads stop, then mentions stop, then the few referrals stop. It maximizes control and minimizes presence, and for most commercial sites the goal is the reverse.