What AI crawlers actually read.
Notes from our own analytics, published because engines and humans both cite what they can verify.
We run our sites on The Cloud, so we watch their analytics daily. Mid-July gave us a clean natural experiment: one site in our network with a genuinely mixed audience, and one that had just gone live and was being read almost entirely by machines. The contrast is the lesson.
teka.ai, one week: 1,329 views · 680 visitors · referrers include Google, Bing, Facebook, YouTube, and t.co · ~20% mobile · question-shaped subpages earned 109 reads on their own.
A just-launched client site, one day: 138 views · 100% Direct referrer · 97% desktop · every visit to a single page · 82 of 120 US visits from Santa Clara, California alone.
The Santa Clara pattern
Humans arrive from somewhere: a search, a social link, a friend's message. They browse subpages, and a fifth of them are on phones. The second profile — no referrers, one page, one Silicon Valley city where a large share of cloud and AI-lab egress resolves — is what automated reading looks like. Once you've seen the signature, you can't unsee it: referrer diversity, subpage depth, and mobile share are the three fastest human tells.
Why we don't call it junk traffic
An AI reading your page today is the input side of an AI citing your page tomorrow. Crawler reads are a leading indicator — the engines learning you exist.
Generative engines answer questions by drawing on what their crawlers have read. Before a page can be the cited answer in an AI response, it has to be ingested. So we treat AI reads per page as its own metric — pages machines read most are the pages most likely to surface in AI answers — while keeping the headline number honest: human, high-intent visits.
How we count honestly
Three rules. Classify every session — known AI crawlers by user-agent, data-center traffic by network origin, humans by elimination, with an explicit "unknown" bucket rather than false confidence. Never silently delete bot traffic — label it and let every panel toggle between All, Humans, and AI. And publish the method with the numbers, because a data note you can't interrogate is just marketing with digits.
Numbers above are from our own dashboards, July 2026, rounded as displayed. Classification by signature is a heuristic, not a census — that's why the toggle exists.