The Referrers That Never Clicked

01Thirty six visits from Google, four clicks

In the seven days from 6 to 12 October, the access log of an agency site I run recorded 36 page loads that said they came from Google. Search Console, which counts clicks on Google's side when they happen, recorded 4 for the same site and the same days, out of 310 impressions. Both numbers cannot describe the same people. One of them is a count of what visitors said. The other is a count of what happened. A referrer is a claim the visitor makes about itself, and anyone can make it.

Who was claiming

Ten of the 36 were easy. They came from Stripebot, the site checker of a payment company, which sent a Google referrer along with every visit. The rest wore ordinary desktop browser names, mostly Linux and Windows builds of Chrome, and landed on the about page, the homepage and a couple of guides. Some were probably people. Most were not, and I cannot tell which from the log alone, because the log records the address of the network edge in front of the site, not the visitor. That is the honest limit of a raw log: it can show you a lie, but it cannot always show you the truth.

02Agent names are costumes too

ChatGPT referrals that were not visits

The log also held 244 requests carrying a ChatGPT referrer. None of them was a page load. They were small files a browser fetches in the background after someone lands, and they came from very few sessions. The landings themselves, tagged by ChatGPT with its own marker in the address, numbered 3 that week. One was a bot announcing itself. Two were people, and both went to the careers page. Two hundred and forty four requests and two people is what a referral count looks like before you read it.

Assistants that never fetched a page

The agent names were worse. Requests calling themselves Perplexity-User numbered 68, and not one fetched a real page. Claude-SearchBot: 68 requests, zero pages. GPTBot: 84 requests, 3 pages. Across all the agent names, 238 requests that week asked for things like private key files, admin paths and configuration files that no assistant answering a question would ever want. Scanners wear the names of trusted crawlers because some sites wave those names through. Only ChatGPT-User behaved like itself, with 135 successful page fetches out of 214. I counted the costume problem in your traffic chart is lying; it has not gone away.

03Three rules for counting machines and clicks

Clicks come from the engine

For search, take clicks from Search Console and Bing Webmaster Tools, never from referrers. They lag by a few days and they hide rare queries, but they count clicks where clicks happen. For assistants, count only landings that carry the assistant's own tag in the address, and check the user agent of each one, because bots copy the tag too. A referral number you cannot reconcile with the engine's own report is a rumour with a decimal point.

Four clicks from 310 impressions is a click rate of 1.3%, which is ordinary for a site whose pages mostly sit beyond the first page of results. It is a small number, and it is true. I would rather plan around four clicks I can trust than thirty six I cannot, because every decision made on the bigger number would be a decision about bots.

Agents count only when they fetch something real

For crawlers and live agents, count successful fetches of real pages, never all requests with the name. A 404 on a configuration file is a scanner, whatever it calls itself. Better still, check the published address lists the AI companies maintain, which needs the visitor's real address in your log; behind a CDN that means restoring it at the origin first. A name in a log is a claim, the warning in nobody reads your blog, and an address check turns it into a fact. An agent that never fetched a page was never an agent.

None of this makes logs useless. They are still the only place you see machines at all, the method behind how to know if AI is reading your site. They just answer a different question. Logs tell you who knocked. The engine tells you who it sent.

Tomorrow: four posts that cited us with numbers we never wrote.