How many times does ChatGPT read your site for a single visit? An AI-traffic X-ray of a 990-page site
In short: on e-slovenie.com, ChatGPT reads about 560 pages a day on behalf of its users and sends back 7 visitors a day. A ratio of 81 reads for 1 click. Only 58 % of page requests come from humans. AI assistants account for 2.1 % of visitors, up 173 % in a month. All of it can be measured, provided you stop looking for the answer in Search Console.
The question that started it
After my post about the 12 August eclipse and the traffic it generated, a reader asked the right question: “Do you have a way to tell this AI traffic from real human traffic in your logs, or is it still at the hypothesis stage?”
The answer has three parts, and most of the “AI traffic” numbers going around never say which one they are using.
| Level | Can it be measured? | How |
|---|---|---|
| Bots | Yes | User-triggered agents (ChatGPT-User, Claude-User, Perplexity-User) identify themselves in the user-agent. CDN logs count them per page per day. |
| Humans | Yes, but under-counted | Someone who clicks a link inside an answer arrives with the assistant’s domain as referrer (chatgpt.com, copilot.microsoft.com, perplexity.ai, gemini.google.com, claude.ai). Apps often strip that referrer: it is a floor. |
| No | AI Overviews and AI Mode are folded into the ordinary “Web” totals of Search Console, with no filter. Only proxies remain. |
The gap between the first two levels is what I call ghost traffic: reads that never become a visit.
What the logs say (measured)
Three days of Bunny CDN logs (22 to 24 August 2026), every request classified by user-agent. Out of 8,664 page requests a day:
- Humans: 5,067 (58 %)
- Classic search engines (Googlebot, bingbot…): 1,120
- Unknown bots: 739
- AI search indexers (OAI-SearchBot, PerplexityBot, YouBot, Amazonbot…): 660
- On-demand reads for AI users: 584
- SEO tools: 302, model training (GPTBot, ClaudeBot, Bytespider): 135
The dominant agent is ChatGPT-User, at 562 requests a day. That is the user-agent OpenAI sends when a user asks a question and ChatGPT goes to read the page right then. One request is roughly one answer that relied on the page. Over three days it read 101 distinct pages of the site.
Per platform, in requests per day:
| Platform | Crawl (training) | AI index | On-demand read | Human clicks | Ratio |
|---|---|---|---|---|---|
| ChatGPT | 16 | 89 | 562 | 6.9 | 81 |
| Claude | 57 | 0 | 20 | 0.7 | 29 |
| Perplexity | 0 | 44 | 0 | 0.2 | none |
| Gemini | 3 | via Googlebot | invisible | 0.7 | unknown |
| Copilot | 0 | via bingbot | invisible | 0.7 | unknown |
| Alexa / Rufus | 0 | 193 | 0 | 0 | none |
Two lessons. Perplexity indexes the site but never reads it for a user: being indexed is not being used. Gemini and Copilot have no reading agent of their own, they ride on Googlebot and bingbot: their clicks exist, their reads are invisible.
The pages assistants read
Almost only English pages, almost only “answer” pages: the new traffic rules on the Vršič pass, the Soča valley without a car, access to the Peričnik waterfall, the Bohinj car train, the Mangart road, the Vintgar gorge. Dated data, checked on site, laid out in tables. What an assistant cannot make up, it comes to fetch.
On the human side, the landing pages from ChatGPT are the same topics, with one clear bias: transport questions make people click. A car-free itinerary or a car-train timetable is something the assistant cannot summarise on its own.
Humans coming out of an AI answer (measured)
Over 30 days in Umami: 276 visitors from an AI assistant, 2.1 % of the site’s 13,377 visitors, against 101 in the previous 30 days (+173 %). ChatGPT accounts for 207 (+204 %), Gemini 21, Claude 20, Copilot 20, Perplexity 6. Mixed engines that blend search and AI (DuckDuckGo, Brave) add 708 visitors (+340 %), counted apart because the two cannot be separated there.
And Google? (inferred)
Search Console exposes no AI metric, not even in the searchAppearance dimension. But the proxies are clear and consistent, 28 days against the previous 28: the share of long queries (6 words and more) goes from 3.9 to 8 % of impressions, question-shaped queries from 6.8 to 10.7 %, impressions from outside the French-speaking area from 71 to 80 %, across 196 countries.
One confounder not to hide: the 12 August eclipse explains most of the window’s +222 % impressions, and it is what pushes the top-3 CTR up (7 → 16 %), the opposite of an AI Overviews signature. The rest of the evidence holds: the pages receiving more than 90 % foreign impressions across 40 to 80 countries are exactly the ones ChatGPT-User reads most. Two systems that never talk to each other pick the same pages. The hypothesis is solid: AI Mode cites these pages in dozens of languages, as already observed on the Vintgar page in July (528 visible queries in 27 languages, 13 of 21 clicks from conversational citations).
Reproduce the analysis on your site
I turned the method into an open-source skill for Claude Code, geo-xray, in the github.com/lipaonline/seo-geo repository. Three dependency-free scripts (Bunny logs, Umami referrers, Search Console API) and a per-platform funnel. Adaptable to another CDN or analytics tool by editing a single file. One caveat: the Bunny logging API keeps only 3 days, so enable permanent log storage on the pull zone or archive daily.
The full 9-page report: AI-traffic X-ray, e-slovenie.com, August 2026 (PDF, in French).
Takeaways
- AI traffic can be measured, provided you look in the right place: logs and referrers give facts, Search Console gives hints. Do not blend the two into one number.
- The reads-per-click ratio is the real AI-visibility metric. 81 : 1 here. In three months it will say whether assistants read more, send more, or read without ever sending.
- Being indexed is not being used.
- The pages that get read are “answer” pages with dated data.
- What earns the click is what the answer cannot absorb.