Which Engines Send People and Which Only Take

01Fifteen days, thirteen hundred reads, nobody sent back

Across fifteen days in August this blog took 1,358 machine requests and received zero referral clicks from ChatGPT, Perplexity, Claude, Copilot or Gemini, and exactly one from Google. I ran the count again for the first half of September. 1,365 requests to entry pages, zero visits from any assistant, six from Google. Two separate fortnights, the same answer.

A reader and a courier are different animals

Those numbers only look like failure if every machine is supposed to send somebody. They are not. Some agents read you to learn, to build an index or to answer a question in place, and nobody follows them back. Others fetch a page while a person waits and hand that person a link they may actually click. I argued a month ago that the machines never stopped reading. The better question is which of them ever brings anyone with them. Some of them are readers. Some of them are couriers. Only one type shows up in analytics.

Readers are not freeloaders. A training crawler reading a page today may be the reason a model describes you correctly next year, and an index builder decides whether a live fetch can find you at all. They pay in a currency analytics was never built to count.

02Fetches against visits, engine by engine

This blog is too small to answer that on its own, so I ran the same comparison across the other sites on the same server over the same September fortnight. On one side, live fetches: requests from the agent each company sends when a person asks a question. On the other, visits that arrived from that company's assistant. The ratio between fetches and visits is the most honest description of an engine I have found.

ChatGPT is the courier

ChatGPT stands apart. Its live fetcher made 10,867 requests and roughly 4,100 visits followed, about one visit for every 2.6 fetches. It also labels its own outbound links: 11,151 page requests in that fortnight carried a tag naming chatgpt.com as the source, which matters because referral headers get stripped all the time and a tag in the address survives. The engine that tags its links is the one that expects to send you somebody.

Claude reads the most for every visitor

At the other end, requests identifying as Claude's live fetcher came to 5,845 and visits from Claude came to 17, around 340 fetches for each person who arrived. Perplexity sat in between at about sixty. I cannot see inside either product and will not guess at why, but on my sites a Claude fetch almost never turned into a visit. Copilot and Gemini sent 111 and 137 visits, and neither runs a separate live fetcher I could count, so they have no ratio at all. A fetch that never becomes a visit may still be a mention you will never see.

03Measure each engine the way it behaves

For couriers, tag and count visits

An engine that sends people can be measured the old way. Give AI referrals their own channel in analytics, keep the tags that arrive on the address rather than stripping them, and watch which landing pages they reach, because those are the pages being quoted with a link. It is the one part of AI visibility your existing dashboard can see, provided you stop it from folding everything into one misleading line.

For readers, count fetches and mentions

An engine that mostly reads will never appear in that channel, so measure it where it does leave a trace. Server logs show which pages its live fetcher requests and how often, which is the closest thing to seeing what it uses. Then check what it says, with the same fifteen run routine from one question across five engines. Judge a courier by the visitors it brings and a reader by the pages it keeps coming back to.

None of this makes the readers less worth courting. It changes what you should expect from them, which is presence in answers rather than sessions. If a large share of your audience asks Claude or Perplexity, a dashboard that only counts visits will quietly advise you to stop the very work that is reaching them.

Two limits keep this honest. A user agent is a name anything can claim, and my sites sit behind a CDN, so every figure here is a declared count rather than a verified one. And one fetch can serve many answers, or none. The ratios are a description of behaviour on my sites over two weeks, not a law, and they will move.

Tomorrow: a site that got every technical detail right and still went uncited.