How to Track Whether AI Engines Are Citing Your Website

You can track AI citations in three ways: read your server logs for AI crawler visits, filter your analytics for referrals from AI hosts, and prompt the AI engines directly to see whether your pages come back. No single method is complete, and none of them is as tidy as Search Console. Together they give you a usable picture of whether AI systems read your site, quote it and send people to it.

This has become a real gap in most reporting setups. A visitor who asks ChatGPT a question, reads the answer, and clicks a cited source arrives at your site through a path your standard dashboards were never designed to describe. Worse, a visitor who reads the answer and does not click has consumed your content without producing a single line in any report you own. If your organic traffic is flat but your content is being used, you would currently have no way of knowing.

What follows is a measurement setup you can build this week, mostly with tools you already have, plus an honest account of what it cannot tell you.

What are the three layers of AI visibility?

Being useful to an AI engine happens in stages, and each stage has its own measurement. Conflating them is the main reason people conclude their AI data is meaningless.

Notice the funnel is leaky in an unusual way. A page can be crawled heavily and never cited. It can be cited constantly and clicked rarely, because the answer satisfied the user. And it can produce clicks you cannot attribute at all, because some AI interfaces strip the referrer. Treat the three numbers as three separate health checks, not as a conversion funnel.

How do you check which AI crawlers visit your site?

Server logs are the only honest record of who fetched what. Every AI company that crawls the open web identifies itself with a user agent string, and those strings are published. If you have access to raw access logs through cPanel, your host’s control panel or a CDN dashboard, you can count visits per bot in a few minutes.

# Count AI crawler hits in an Apache or Nginx access log
grep -Eoi "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-Web|
PerplexityBot|Perplexity-User|Google-Extended|Applebot-Extended|
CCBot|Bytespider|Amazonbot|meta-externalagent" access.log 
  | sort | uniq -c | sort -rn

# Which URLs is GPTBot actually fetching?
grep "GPTBot" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

Two notes on that table. Google-Extended is not a crawler at all — it is a control token you use in robots.txt to allow or block Gemini training, so it will not appear in logs as a fetch. And ChatGPT-User is the most interesting line to watch: it means a real person, mid-conversation, caused your page to be fetched. A rising count there is closer to a traffic signal than anything else in your logs.

Before you measure any of this, confirm you are not blocking the bots you want. A robots.txt that disallows GPTBot guarantees zero citations from ChatGPT’s index, and plenty of sites carry such a rule without realising, inherited from a security plugin or a template. Check your own file first.

How do you see AI referral traffic in analytics?

AI referrals already appear in your analytics — they are just scattered across the referral report under hostnames you are not looking for. The fix is to group them into one named channel so you can watch the trend instead of hunting for rows.

1. Open your referral or source report and sort by sessions. Scan for any host containing chatgpt, openai, perplexity, copilot, gemini, claude, bing with a chat path, or you.com.
2. Create a custom channel group. In GA4 this is Admin → Data display → Channel groups. Add a channel named “AI assistants” with a rule matching source contains any of those strings.
3. Apply it going forward. GA4 custom channel groups are not retroactive beyond a limited window, so create this before you need the history, not after.
4. Add a landing-page breakdown. Which pages AI engines send people to tells you which of your pages they consider authoritative — a far more actionable signal than the total.
5. Watch engagement separately. AI-referred visitors arrive with a question already half-answered. Their behaviour usually differs sharply from search visitors, and judging them by the same benchmarks will mislead you.

One caveat that matters more than it sounds. Some AI interfaces, particularly inside mobile apps, pass no referrer at all, so those visits land in your Direct bucket and are indistinguishable from someone typing your URL. If your Direct traffic has grown without explanation, a portion of it may be AI referrals you cannot see. There is no clean workaround; the practical response is to treat Direct growth as a weak corroborating signal rather than a mystery.

How do you test whether AI engines actually cite you?

This is the layer that matters most and the one no dashboard gives you. The method is unglamorous: build a list of the questions your customers really ask, put each one to each engine, and record what comes back. Done consistently, it becomes a rank-tracking equivalent for AI answers.

Record the competitors who get cited in your place. That list is the most useful output of the whole exercise, because it tells you which sources the engines currently treat as authoritative on your topic — and those sources are where your own citations and mentions need to appear.

Several commercial tools now automate this prompt-and-score loop across engines. They save time and nothing more; the underlying method is the one above, and running it by hand for a month teaches you more about your market than a dashboard will. If you do buy a tool, check that it runs fresh sessions, covers the engines your audience actually uses, and reports who was cited instead of you.

Does Search Console show AI Overview data?

Partially, and not in a way you can isolate. Google has said that impressions and clicks from AI Overviews are included in Search Console’s Performance report, but they are blended into the overall Search totals with no separate dimension to filter by. There is no “AI Overviews” row to click.

What you can do is read the pattern. The signature of a page being summarised rather than clicked is impressions holding steady or rising while clicks fall, which shows up as a declining click-through rate on queries that have not changed position. That is suggestive, not conclusive — seasonality and layout changes produce the same shape — so look for it across a group of related queries before drawing a conclusion.

If you see that pattern, the response is not to fight it. The page is doing its job as a source; what it needs is a reason to click that the summary cannot replace — a calculator, a template, current prices, a download, your own data. Pages whose entire value is a definition will keep losing clicks, and no amount of measurement changes that.

What should you actually report each month?

Six lines is enough. The temptation with a new measurement area is to build something elaborate, and elaborate reports on noisy data produce confident wrong conclusions. Keep it small, keep it monthly, and look at the direction rather than the absolute numbers.

What do you change once you have the data?

The measurement is only worth the hour if it changes what you publish. Three findings come up repeatedly, with three fairly clear responses.

• Crawled heavily, never cited. The engines can read you but do not find you quotable. Usually a structure problem: the answer is buried three paragraphs in, the heading is a slogan rather than a question, the key fact is inside an image. Rewrite answer-first.
• Cited but never clicked. Normal and not always a problem. Add something to the page the summary cannot carry, and make sure your brand name sits inside the sentence likely to be quoted, so you get the mention even without the link.
• Competitors cited consistently in your place. Look at what those pages have that yours does not. Frequently it is first-hand data, a named expert author, a clear publication date, or simply more specific numbers.

All three fixes are content work, not technical work — which is the quiet lesson of AI visibility measurement. The levers are the ones covered in our guide to appearing in Google AI Overviews and the broader piece on ranking factors in the GEO and AEO era. On the technical side, make sure the pages you want cited are indexable in the first place — see crawled but not indexed — that their facts are machine-readable via schema markup, and that crawlers can reach them through sound internal linking. If you are weighing whether to publish an llms.txt file, that piece sets out what it does and does not do.

Start this month, even roughly. The value of AI visibility tracking is almost entirely in the comparison between readings, which means the single most useful thing you can do today is take the first one. More practical guides sit in our technology and digital marketing section.