How to Find Which Pages AI Engines Cite Most

AI engines do not cite websites. They cite pages — usually a specific section of a specific page that answered a specific sub-question. So “are we visible in AI search?” is the wrong resolution. The useful question is which of your URLs are getting pulled into answers, and that turns out to be measurable, if imperfectly.

There is no report that just tells you. But three signals, read together, get you close.

Signal 1: referral traffic, segmented by landing page

The most direct evidence, and the most incomplete.

AI engines send referral traffic with identifiable sources — chatgpt.com, perplexity.ai, gemini.google.com and others. In GA4, build an exploration with landing page as the dimension and filter session source to those hosts. What you get is a ranked list of the pages people actually arrived on from an AI answer.

The setup details are in tracking AI traffic in GA4 and Search Console, including the referral-exclusion problem that silently drops some of this data.

What it misses: every citation nobody clicked — which is most of them. If you only measure clicked citations you will systematically undervalue the pages that answer questions completely, because answering completely is exactly what removes the reason to click.

Signal 2: AI crawler hits in your server logs

The signal nobody looks at, and the only one that is a census rather than a sample.

Every documented AI agent identifies itself. The important distinction is when each one fetches:

That last column is the one that matters. An index-building crawler visiting your page tells you it is eligible to be cited. A user-triggered agent like Perplexity-User visiting tells you a real person asked something and your page was fetched to help answer it — as close to a live citation signal as you can get without an API.

Perplexity documents both agents, and notes that Perplexity-User generally ignores robots.txt because the request is user-initiated rather than a crawl (Perplexity crawler docs). OpenAI documents its agents here.

Group your access logs by URL, filtered to those agents, and you get a ranked list of the pages AI systems actually reach for. On Netlify, Vercel or Cloudflare this means enabling log retention or drain — it is off by default on most static hosts, which is why almost nobody has this data.

What it misses: a fetch is not a citation. The page may have been retrieved and discarded. It also tells you nothing about how you were characterised.

Signal 3: sampled prompt runs that record URLs

The only signal that captures citations nobody clicked.

Run a fixed panel of real buyer questions against each engine, repeatedly, and record the exact URLs cited rather than just whether your brand appeared. Over enough runs you get a frequency table of your own pages — which ones the engines reach for, on which questions.

This is the page-level version of the discipline described in what is AI citation tracking. The difference is the unit: brand-level tracking answers “are we named”; page-level tracking answers “which asset is doing the work”.

What it misses: only the questions you thought to ask. Your panel is a guess at the question space, and the real space is far larger.

Reading the three together

Each signal is wrong in a different direction, which is what makes the combination useful.

• Cited in prompt runs + crawler hits, but no referrals — a page doing real work invisibly. This is the most common and most under-valued pattern. Do not cut this page because analytics says it is quiet.
• Crawler hits, but never cited — retrieved and rejected. Usually a structure problem rather than a topic problem; see content chunking for AI.
• Referrals but no crawler hits — the citation is coming from training data or a third-party page about you, not from your page being fetched. That is a digital PR result, not an on-page one.
• Nothing on any signal — the page is not in the retrievable set at all. Check rendering and crawler access first: JavaScript rendering and AI crawlers and robots.txt for AI crawlers.

What to do with the ranking

Once you know your most-cited pages, the highest-return move is almost always to improve them rather than write new ones. A page that already gets retrieved has cleared the hard part — it is in the index, it is being reached for, and the engine has decided it is relevant. Tightening its answers, refreshing its data and making its structure more extractable compounds on an asset that already works.

New pages start from zero and have to earn retrieval all over again.

If pages are being fetched but never cited, the cause is usually structural rather than technical. The crawl-side checks are in the technical SEO checklist for AI search; if a page is cited but the description of you is wrong, that is a different problem entirely, covered in when AI gets your brand wrong.

Where this leaves your reporting

If your AI visibility report is a single number, it is hiding the thing you can act on. The page-level view is where the decisions are: which assets to strengthen, which to restructure, and which topics you are absent from entirely. Start with server logs if you can turn them on — it is the only census available, and almost no one is reading it.

Want the page-level picture for your own site without building the pipeline? Our AI visibility audit runs the prompt panel and reports which of your URLs come back.