How to measure if AI assistants cite your site
Last updated: May 27, 2026
TL;DR
There is no “Google Analytics for AI” in 2026. The four signals that exist today are: referrer headers(Perplexity and ChatGPT Search send them, most others don't), server-log fingerprints (AI bots have identifiable user-agents), direct prompting (ask the assistants and read what they say), and branded-query lift on Google. Combined, they give you a defensible picture. In isolation, each one lies.
Why this is genuinely hard
Classic analytics works because browsers send referrers and execute tracking scripts. AI assistants do neither, consistently. A ChatGPT user reading an answer that quotes your page never visits your site at all — the citation lives inside the chat, and your server sees nothing. Even when they do click through, the assistant frequently strips the referrer. So “am I being cited” cannot be answered by traffic alone. You need a stack of imperfect signals that, taken together, triangulate the truth.
The four signals that actually work
1. Referrer headers (when you're lucky)
Some assistants do send a Refererheader on click-throughs. Perplexity is the most reliable (you'll see perplexity.ai in your logs and in any analytics tool). ChatGPT Search sends chatgpt.com on links inside answer cards. Google AI Overviews send the standard google.comreferrer — indistinguishable from organic search unless you correlate against the SERP. Filter your analytics by these hostnames and you get a floor — the real number of citations is higher (most don't produce a click) but anything you see here is real.
2. Server-log fingerprints (which bots are actually visiting)
AI crawlers identify themselves. Grep your access logs for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Google-Extended. The pattern matters: a crawler that hits once a month is doing training; one that hits seconds before a user sees an answer (ChatGPT-User, OAI-SearchBot) is doing retrieval — that's a live citation in flight. The volume and rhythm of these fingerprints tells you whether your content is in the assistants' rotation, even when no click follows. Bonus: it's also how you catch a CDN that's silently 403-ing bots, which we covered in robots.txt for AI crawlers.
3. Direct prompting (the only way to know what they say)
The most direct signal — and the one most teams skip because it doesn't fit a dashboard. Pick 10–20 queries a real customer would ask, run them weekly across ChatGPT, Claude, Gemini, Perplexity and Copilot, and record whether your brand appears, in which rank, and with what attribution. This is the only way to catch hallucinated facts about you, missing entity recognition (“a tool that does X” without naming you), or competitors being cited where you should be. Keep a versioned log; trends matter more than any single run.
4. Branded-query lift on Google
The indirect-but-reliable proxy: when an assistant cites your brand by name, a meaningful fraction of users will Google the brand to verify or find the official site. A rising floor of branded searches in Google Search Console — especially uncorrelated with your ad spend or PR cycle — is a strong lagging indicator that your AI citations are compounding. It's noisy at low volume but reliable once your brand has any baseline recognition.
What no tool gives you (yet)
Honest list: there is no API from OpenAI, Anthropic, Google or Perplexity that reports “here are the queries where we cited your URL last week.” Anyone selling that as a product is, at best, extrapolating from sampled prompts and search-engine signals — useful, but not ground truth. Be skeptical of dashboards that claim to count “ChatGPT impressions” without telling you their sampling methodology.
A pragmatic monitoring stack
For most teams in 2026: log analysis weekly (10 minutes), a recurring prompt set across five assistants weekly (30 minutes), GSC branded-query trendline monthly (5 minutes), and a per-agent crawlability audit after every infrastructure change. The last one is the unglamorous one that surfaces 80% of the “why did citations drop” mysteries — usually a CDN rule, a robots.txt change, or a framework upgrade that broke SSR. You can run a free audit per URL, or take the 2-minute readiness quiz for a directional read. For continuous tracking, the paid plans add scheduled re-runs and regression alerts.
FAQ
Does Google Analytics show ChatGPT traffic?
Partially. You'll see chatgpt.com referrals for clicks from ChatGPT Search, but only a fraction of citations produce clicks, and pure-chat citations are invisible. GA is a floor, not a measure.
How often should I run the prompt set?
Weekly is the sweet spot for trend visibility without burning hours. Daily is overkill (assistants don't re-rank that fast); monthly is too sparse to catch regressions before they cost you.
Should I block PerplexityBot if it doesn't bring traffic?
Usually no — Perplexity is one of the few that sends real referrers and tends to attribute citations explicitly. If you genuinely want to monetise rather than be cited, look at Bot Paywall instead of blocking, which charges specific bots per URL.
Want the per-agent breakdown for your own site? Run a free audit, or read the companion piece on why AI assistants ignore your site for the most common reasons your citations are flat.