Measuring AI visibility: how to track citations in ChatGPT, Perplexity and AI Overviews

Why AI visibility has to be triangulated

Classic search gives you a report: queries, impressions, clicks, position. Answer engines give you nothing of the kind. OpenAI, Perplexity, Anthropic and Microsoft publish no per-site citation data, and Google folds AI Overview traffic into its existing search totals. Measurement therefore means combining four partial signals, each of which sees a different slice of what happens between a question and your page.

When referrals, user-triggered fetches and sampled citations move in the same direction for a topic, the trend is real. When one moves alone, check the instrument first.

Referral traffic from answer engines

The cheapest signal is already in your analytics. When someone clicks a citation, the visit arrives with a referrer, a UTM parameter, or both.

OpenAI's Publishers and Developers FAQ states that ChatGPT automatically includes utm_source=chatgpt.com in referral URLs so that publishers can track the traffic in tools such as Google Analytics. The other engines are identified by referrer host:

• chatgpt.com (and the older chat.openai.com) for ChatGPT
• perplexity.ai for Perplexity
• copilot.microsoft.com for Microsoft Copilot
• claude.ai for Claude
• gemini.google.com for Gemini

Grouping them in Google Analytics 4

GA4 files these under Referral by default. Google's custom channel group documentation includes an "AI assistants" example built on a matches regex condition against the Source dimension, and notes that custom channel groups apply retroactively. Google's example pattern is deliberately broad; a tighter expression that matches only the hosts above is:

^(www\.)?(chatgpt\.com|chat\.openai\.com|perplexity\.ai|copilot\.microsoft\.com|claude\.ai|gemini\.google\.com)$

Place the new channel above Referral so it is evaluated first. A standard GA4 property allows two custom channel groups.

What referrals undercount

Referral traffic is a lower bound on citations. Answer engines resolve many questions without a click, so a citation often produces no session. Native apps and in-app browsers frequently drop the referrer, so the visit lands in Direct. And some engines put citations in a collapsed source list that few users open. A flat referral line means clicks are rare, which is the normal case, not that you are uncited.

AI crawler hits in server logs

Your access logs record every fetch, including the ones no analytics script sees. The vendors publish their user agents, and the names mean different things.

OpenAI's crawler documentation lists three relevant agents: GPTBot crawls content that may be used for training; OAI-SearchBot surfaces websites in ChatGPT's search features; and ChatGPT-User handles certain user actions in ChatGPT, where, because the fetch is initiated by a user, robots.txt rules may not apply. Perplexity documents PerplexityBot for its search index and Perplexity-User, which may visit a page when a user asks a question and which, in Perplexity's words, generally ignores robots.txt. Anthropic documents ClaudeBot (training), Claude-SearchBot (search quality) and Claude-User (fetches on behalf of a user), all of which it says honour robots.txt. Google uses ordinary Googlebot for AI features; Google-Extended is a robots.txt control token with no user agent string of its own, so it never appears in a log.

That gives a three-way split worth counting separately:

1. Training crawlers (GPTBot, ClaudeBot, CCBot). Volume tells you about training exposure, not about answers.
2. Search-index crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot). A fetch means the page is a retrieval candidate.
3. User-triggered fetchers (ChatGPT-User, Perplexity-User, Claude-User). A fetch means the page was retrieved while a live question was being answered. This is the closest thing you have to a real-time citation signal.

A short pipeline over an Nginx or Apache combined log gives a daily count per agent and URL:

grep -E 'GPTBot|OAI-SearchBot|ChatGPT-User|PerplexityBot|Perplexity-User|ClaudeBot|Claude-SearchBot|Claude-User' access.log \
  | awk '{
      match($0, /GPTBot|OAI-SearchBot|ChatGPT-User|PerplexityBot|Perplexity-User|ClaudeBot|Claude-SearchBot|Claude-User/);
      agent = substr($0, RSTART, RLENGTH);
      split($4, d, ":");
      print substr(d[1], 2), agent, $7
    }' \
  | sort | uniq -c | sort -rn | head -50

User agent strings are trivially forged, so verify before trusting a spike. OpenAI publishes IP ranges per agent (openai.com/gptbot.json, openai.com/searchbot.json, openai.com/chatgpt-user.json), Perplexity publishes perplexity.com/perplexitybot.json and perplexity.com/perplexity-user.json, and Anthropic publishes claude.com/crawling/bots.json.

Prompt-based citation sampling

Referrals and logs are passive. The only way to learn which questions you are cited for, and who is cited instead, is to ask the engines yourself and record the answers.

Build the panel

Write twenty to fifty questions that your pages are meant to answer, phrased the way users phrase them rather than as keywords. Give each question a stable id, version the panel, and never silently edit a prompt: an edited prompt is a new prompt.

Run each prompt in a fresh, logged-out session on each engine, because answers vary with history, account and location. Generation is non-deterministic, so run each prompt three times and record every run.

Record the answers

One row per answer:

Define the metrics

• Citation rate: answers in which your domain is cited, divided by all answers, per engine.
• Share of answer: your citations divided by all citations across the panel's answers, per engine. This is the answer-engine equivalent of share of voice, and the number that falls when a competitor starts winning.
• Mean citation position: average of position over cited answers. A rising mean with a stable citation rate means you are being pushed down the list.
• Competitor share: share of answer computed for every domain in all_domains. Keep the whole distribution.
• Coverage: answers that showed any sources at all. A coverage drop is not a visibility drop.

The GEO paper by Aggarwal et al. formalises the same idea with two finer measures: Position-Adjusted Word Count, which weights the words attributed to a citation by an exponentially decaying function of its position, and Subjective Impression, which uses a model to rate how prominent a source looks in the answer. A weekly sheet with cited and position columns is a coarse approximation of both.

What Google Search Console can and cannot tell you

Search Console is the only first-party source for Google, and its limit is precise. Google's documentation on AI features and your website says that sites appearing in AI features such as AI Overviews and AI Mode "are included in the overall search traffic in Search Console" and are "reported on in the Performance report, within the 'Web' search type". As of this writing there is no filter that isolates them, so an AI Overview click and an ordinary click on the same query are indistinguishable in the export.

What you can still do:

• Export query-level impressions and clicks weekly for the topics in your panel, and watch click-through rate. Impressions holding steady while CTR falls is consistent with an AI Overview absorbing the click, but also with a new competitor or a changed layout.
• Record whether an AI Overview appeared for each query, from a rank tracker that detects it or from your own sampled searches, and whether your URL was among its sources.
• Google's documented controls for AI features (nosnippet, data-nosnippet, max-snippet, noindex) are levers, not measurements, but they allow a before-and-after comparison on one section of the site.

Categories of commercial tools

The market is young and the products overlap. Four categories cover most of it, and none replaces the definitions above.

1. Prompt-monitoring platforms automate the panel across engines on a schedule. Ask whether they query the consumer interface or an API (the two can cite differently), how many runs per prompt they take, how they handle regions, and whether raw answers can be exported.
2. Bot analytics in CDNs and WAFs classify crawler traffic by user agent and, in the better cases, by verified IP range. Ask whether verification is done or only claimed.
3. Analytics add-ons and dashboard templates package an AI channel group and a report. The regex above is most of the product; the value is presentation and alerting.
4. Rank trackers with AI Overview detection record whether an AI Overview appeared for tracked keywords and whether your URL was cited in it. Google only, but it fills the Search Console gap.

Whatever you buy, keep the spreadsheet. The metric definitions should be yours, so that changing tools does not reset the series.