How many times AI crawls a site before it sends one visitor

How we measured

Everything here comes from raw Nginx access logs on our own servers, for the 14 days from September 12 to September 26, 2026. Four sites, about 1.4 million log lines in total. No sampling, no analytics product in the middle.

A crawl means one successful page request from a named AI agent. We counted only HTML page responses, and removed static assets, API calls, error responses and the constant background noise of vulnerability scanners probing for WordPress files.

A visit means one page view whose referrer was an assistant domain such as chatgpt.com, perplexity.ai or claude.ai, made by a current browser user agent. That last filter matters. Referrer spam is heavy on every site we run, and we have caught bot networks faking assistant referrers before, so we excluded outdated browser versions and anything identifying as automated. The filter removed about five percent of raw referral hits.

The four sites are described by type rather than name: a CV builder, an SEO agency site, a legal information site, and this site. They differ enough in purpose to make the comparison the interesting part.

Results by engine

Crawler fetches and attributable visits, summed across all four sites for the 14 day window.

Two search engines sit outside this table because their crawlers serve ordinary search as well as AI answers, which makes a clean ratio impossible. For the record, Googlebot made 2,050 page fetches in the same window while the Gemini app sent 72 visits, and bingbot made 18,149 fetches while Copilot sent 43. Most of what those two crawlers do is not AI work at all.

The Anthropic line is the one people react to. More than a thousand page fetches, one attributable visit. Some of that is measurement rather than behaviour, because assistants do not always pass a referrer, but the direction is consistent with what other operators have reported publicly.

Results by site

The same numbers, split by site. This is where the averages fall apart.

One site collects 99 percent of the visits. The CV builder receives between 1,400 and 2,700 visits a day from ChatGPT alone, arriving from more than 3,700 distinct visitors on current versions of Chrome, Edge, Safari and Android browsers. It is ordinary human traffic that happens to be introduced by an assistant.

The legal information site is the mirror image. It is well written, it ranks, and AI agents read it constantly. In 14 days it earned nine attributable visits. Its content answers the question so completely that there is no reason to click.

Where the visits actually land

On the CV builder, the pages that receive assistant traffic are not the articles. The top landing pages were the builder itself, the home page, the live CV feature and the pricing page. Blog posts barely appear.

This is the single most useful finding in the dataset. Assistants answer the informational question themselves, then hand the person over at the moment they need to do something. Nobody clicks through to read your explanation of how to write a CV. They click through to build one.

It reframes what an AI visibility strategy is for. Your explanatory content earns the model's trust and supplies the answer. Your tool, template, calculator or product page collects the visit. A site with only the first half gets crawled forever and stays empty.

What this means for your site

Three practical conclusions come out of this data.

First, judge AI visibility by whether you have something to hand over, not by traffic alone. If your site is purely informational, expect a poor crawl to visit ratio no matter how good the writing is, and measure your value in mentions rather than sessions.

Second, give the assistant a reason to send someone. A free tool, a calculator, a template, a live checker or a product page is the destination that makes a referral worth making. Put it where the answer ends, not three clicks away.

Third, be careful with blanket blocking. Publishers who block everything have a defensible case, and the ratios above explain why. A business with something to sell is in a different position, because an assistant that reads a page and then recommends the product is delivering a customer rather than taking value. Blocking agents that act for a real person is where that goes wrong, which is exactly what several networks now do by default, as we covered in the Cloudflare crawler defaults change.

Caveats worth stating

Four sites is a small sample, and they are our own, so treat these as measurements rather than universal truths. Attribution is imperfect: assistants sometimes strip the referrer, apps often pass none at all, and a visit with no referrer is counted as direct here, which understates every engine to some degree. ChatGPT is the easiest engine to measure because it tags outgoing links, so part of its strong showing is better tracking rather than better behaviour. The window is 14 days, which is long enough for crawl patterns and short enough that a single viral answer could move the numbers. We publish the method so you can run the same counts on your own logs and compare.

The bottom line

AI crawling pays back unevenly. Averaged across four sites it looks like one visit per crawl, but that average is one tool site carrying everyone else. Content sites were crawled thousands of times for a handful of visits. If you want AI traffic rather than AI reading, own the thing a person needs once the answer is finished.

Frequently asked questions

Abd Shanti

Co-founder and strategy lead at Outline Technologies, the team behind FreeGPTSEO and AI Citation Monitor. Abd works on how brands get found and cited by search engines and AI assistants.