How do I find out if AI agents are citing my content? Our agency method
How we find out if AI agents are citing a client's content: manual prompts, referral traffic, bot logs and Promptwatch citation analytics, and what each one misses.
Clients ask whether AI agents are citing their content more often than almost anything else. The short answer is that it takes more than one check, because there are three different things that can be true. An agent can fetch your page, cite your page, or send a visitor to your page. Each needs a different kind of evidence.
Start by hand
We always begin manually. We take the questions a client's content is supposed to answer and run them in ChatGPT with search, Perplexity and Google, and we read the sources. It takes minutes and gives us a feel for the situation.
It is also a sample of one. The same prompt can show different sources tomorrow, so we never report it as a finding. For a closer look at one answer we use the free GEO Inspector Chrome extension from Promptwatch, which shows ChatGPT's fan-out queries and fetched versus cited sources and runs locally.
Look at referral traffic
Next we check analytics. We filter sessions by referrer for ChatGPT, Perplexity and Gemini and see which landing pages get visits. A page with AI-referred visits is being cited and clicked, which is a strong signal. We then note the limits to the client: some apps strip the referrer, so some visits appear as direct, and a citation nobody clicks leaves no mark.
Read the bot logs, carefully
Server logs show which AI bots fetched which pages. The names matter: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, PerplexityBot and Perplexity-User are different bots with different jobs. We keep two warnings in mind. A fetch is not a citation, because a bot can read a page and never use it. And user-agent strings can be faked, so we do not trust a log line without IP verification.
Move to continuous tracking
Manual checks and logs do not scale, so on client programs we run Promptwatch. Four features do the job.
Citation analytics. Page- and domain-level citations with URL positions. We use Page Tracker to watch specific URLs, the client's or a competitor's, and we read the offsite mentions view to see where a brand is named on third-party pages without a link. The Reddit, YouTube and LinkedIn reports show the community side.
Agent Analytics. Real-time, IP-verified AI crawler logs, so spoofed bots are filtered out. It shows a crawl-to-citation path with a citation rate per page and error tracking with failure explanations. This is how we separate "never fetched" from "fetched but never cited."
Visitor analytics. AI-referred sessions and conversions, via a lightweight script, a GTM template or the WordPress plugin.
Prompt monitoring. ChatGPT (Search), Perplexity, Google AI Overviews, Google AI Mode, Copilot, Gemini, Alexa and Meta AI through the real interfaces, and Claude, DeepSeek, Mistral and Grok through APIs, run daily.
Plan note: crawler logs start on Professional at $245 a month. The free Explore tier covers 10 prompts on ChatGPT only.
How we read the three situations
Cited, and it is the right page. We keep the page stable and note it. Citation Trends show first cited, ramp-up days, peak rate and decay half-life, so we can see whether it holds.
Fetched but not cited. The bot reads the page, and the engine prefers something else. This is usually a content-fit problem. We look at the cited sources for the prompt and adjust the page.
Never fetched. This is an access problem. We check robots rules, firewall behavior and errors before touching content.
We have seen programs where access was the whole story. Promptwatch's own customer case study from TreeHouse, behind Pararius and Huurwoningen.nl, tells a similar tale: AI search grew to about 25% of organic conversions, 99% from ChatGPT, after crawler access was fixed. That is that customer's result, not a promise.
What we report
A monthly note with three lines per priority page: fetched or not, cited or not, visits and conversions. It is plain enough for a client to follow and traceable back to data.
The limit
Nothing can force an engine to cite a page, and answers vary between runs. What the data can do is tell you which situation you are in, so you fix the right thing.