Can you find out which prompts sent ChatGPT to your site?

What you can actually know from your own site, what you can only estimate, and how to do both without wasting money.

Every week someone asks a version of this question in an SEO or AEO forum:

Is there a tool that shows the actual prompts where my website is cited in ChatGPT, Gemini, Perplexity or Claude? I want the prompt, whether I was mentioned, the URL that was cited, and the prompts where competitors are cited and I’m not.

The replies are usually a mix of tool names, “no tool has this”, and a few people quietly getting most of it for free. All three camps are partly right. Here is the whole picture.

The short answer

The rest of this article explains each row: why the prompt is out of reach, what your server sees instead, and how to run probing without fooling yourself.

Why the prompt never reaches you

It helps to know what happens between a question and a visit.

1. Someone types a question into ChatGPT (or Perplexity, Claude, Gemini, Copilot).
2. The assistant decides whether it needs to look something up. If it does, it rewrites the question into one or more search queries and runs them against its own index, or a search engine’s.
3. From the results it picks a few pages and fetches them, right then, to read the text.
4. It writes an answer and shows some of those pages as links.
5. If the person clicks a link, their browser lands on your site.

Your site is involved at steps 3 and 5 only. At step 3 the assistant’s fetcher arrives with a name (ChatGPT’s is ChatGPT-User, Perplexity’s is Perplexity-User, Claude’s is Claude-User) and asks for one page. It does not send the question, and it does not send the search query it made from the question. At step 5 the visitor’s browser sends a referrer — chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com — and ChatGPT adds utm_source=chatgpt.com to the link. That tells you which assistant. It never contains the prompt.

So the prompt is held in exactly one place: the assistant’s own logs. None of the companies publish those, for the same reason Google never published the full search history behind your organic clicks.

There is a second, less obvious problem. The “prompt” that produced a citation is often not a clean question at all. Thirty messages into a conversation about dinner, the message that made ChatGPT fetch your recipe page might have been “with sauce?”. Even if you had that string, it would tell you nothing. What actually drove the fetch was the whole conversation, plus whatever the assistant remembers about that person.

What about the tools that show prompt volumes? Some products will tell you that “best raw dog food” is asked in ChatGPT some number of times a month. Look closely at the method. Most of them take the keywords out of your prompt and give you Google search volume for those keywords, relabelled. That is a fair way to size a topic. It is not a count of people typing that sentence into a chatbot, and no tool has that count.

What about “real user prompt data”? One or two enterprise tools buy panel data: a sample of people who agreed to have their browsing recorded, extrapolated to everyone. That is real data about real prompts, from a small and probably unrepresentative group, scaled up. Treat it the way you would treat any survey: useful for spotting big themes, not for counting.

The one assistant that shows its queries

Microsoft is the exception. Bing Webmaster Tools has a report called AI Performance (still marked beta) that covers “Microsoft Copilots and Partners”. It lists the grounding queries — the searches Copilot ran while building an answer — that led it to your pages, and for each one how many times your pages were used and your citation share: what fraction of everything Copilot cited for that query came from you. Each query opens to show which of your pages it pulled in; each page opens to show the queries behind it. Microsoft adds AI-generated intent and topic labels, and lets you download the lot.

Three things to keep in mind when you read it. A grounding query is not the person’s prompt; it is the assistant’s rewrite of it (step 2 above), so you will see “public holidays 2027”, not whatever the person actually typed. A “citation” here means Copilot used the page as a source, not that anyone clicked it — the two numbers are usually far apart, and this report is where you can see just how far. And Microsoft says the data is a sample, so treat the counts as proportions rather than totals.

Even so, it is the only window of this kind anyone provides, and it is free. If you have not verified your site in Bing Webmaster Tools, do it; it takes ten minutes.

Google Search Console does not do the equivalent for AI Overviews or AI Mode. Clicks from those land in your reports as ordinary Google search clicks.

What your own server can tell you

Now the useful half. Everything below comes from requests your server already receives; nothing has to be estimated.

Three kinds of bot visit

AI companies send three different kinds of visitor to your site, and they matter differently.

• Training crawls — GPTBot, ClaudeBot, Bytespider, CCBot and others. Bulk downloads of your pages to build future models. No person is on the other end. Lots of hits, no direct connection to any answer.
• Search crawls — OAI-SearchBot, PerplexityBot, Claude-SearchBot. Index building, like Googlebot. Also no person on the other end, but a page that is never indexed is rarely cited.
• Live reads — ChatGPT-User, Perplexity-User, Claude-User. A person asked something just now, and the assistant fetched this page to answer them. This is step 3 from the list above. It is the closest thing to evidence that your page took part in an answer.

If you only count “AI bot hits”, the training crawls swamp everything and the number means little. Separate the live reads and you have a per-page count of times an assistant read this page to answer someone. That is a real number, not a score.

The read-then-referral pair

Put the live reads next to the referrals and a pattern appears:

14:02:11  ChatGPT-User  fetched  /guides/choosing-hiking-boots/
14:03:40  visitor       arrived  /guides/choosing-hiking-boots/   referrer: chatgpt.com

The assistant read the page, then a person clicked through to it. You still do not know what they asked, but you know your hiking-boots guide answered it, and you know it happened at two in the afternoon on a Tuesday. Do this for every page for a month and you have:

• which pages assistants read most often,
• which of those reads turn into visits (many never do — the person got what they needed from the answer),
• pages that are read a lot but never clicked (the answer was good enough on its own; you are being used as a source, not a destination),
• pages that get referrals without a preceding live read (the assistant answered from its index or memory and still linked you).

That last group is worth its own label. A referral with no live read in the previous hour usually means the page was in the assistant’s search index and it linked you without re-reading — which tells you the search crawls are doing their job.

Why Google Analytics only sees half of this

GA4 sees step 5 only: the human click. It cannot see step 3, because the assistant’s fetcher does not run your analytics script. So GA4 can give you the cited URL when the referrer survives, but never the read that preceded it, and never the reads that produced no click at all.

It also misses more of step 5 than you would expect. Some assistant apps strip the referrer, so the click arrives as “direct”. Some visitors have blockers that stop GA4 loading. And if your site serves cached pages, the request may never reach WordPress for a server-side plugin to see either — which is its own topic, covered in our caching guide.

The practical answer is to count on the server, not in the browser. That is what the Pro version of Forge AI Bot Log does: it records live reads and referrals as they hit WordPress, pairs them per page, and shows the two side by side. We mention it once here and leave it at that; the method works with raw access logs too, if you are comfortable reading them.

Probing: asking the assistants yourself

Your server tells you what happened. It cannot tell you what didn’t happen — the questions where a competitor was cited and you were not. For that you have to ask the assistants, and the only honest way to do it is on a fixed schedule with a fixed list.

Build the list. Write twenty to thirty questions a real buyer would ask before choosing you. Not the questions you would like to rank for; the ones people actually ask, including the awkward ones (“is X worth the money”, “X vs Y for a small business”, “why does X keep breaking”).

Include messy phrasings. The useful signal is usually in the variants you did not think of. Add a few long, rambling versions and a few that misspell your product category.

Run them weekly. Same list, same assistants, same day. Record three things per answer: were you mentioned, which URL (if any) was linked, and which competitors appeared. A spreadsheet is fine to start with. When the list gets too long to run by hand, this is the one part of the job worth paying a tool for; several will run your list across ChatGPT, Perplexity, Gemini and Claude and keep the history.

Know what you are measuring. A probe measures the assistant’s answer to your list, on that day, from your account. Assistants give different answers to the same question on different runs, and they personalise. Two weeks of data is noise; two months starts to be a trend. And the exact wording matters less than you would think — the assistant rewrites everything into search queries anyway, so two differently worded prompts on the same topic tend to surface the same pages.

Don’t pay for prompt volume. See above. Size the topic with ordinary keyword volume if you must, then stop.

Putting the two together

Neither source is the picture on its own. Probing tells you what the assistants say about a topic; your server tells you what they actually read and where they sent people. A workable weekly routine:

1. Run your probe list. Note mentions, cited URLs, competitors.
2. Open the per-page live-read report for the week. Note which pages were read most, and which reads turned into visits.
3. Look for mismatches. A page that probes well but is never read live may only be in the index. A page that is read live constantly but never appears in your probes is answering questions you have not written down yet — go and write them down.
4. Open the AI Performance report in Bing Webmaster Tools. Add any new grounding queries to your probe list, and note which pages Copilot cites that your other sources never show.

Over a quarter, that gives you the four things people ask for, in the only forms they exist: the cited pages and reads from your own data, competitor gaps from probing, and a steadily better guess at the questions behind it all.

Forge AI Bot Log is a free WordPress plugin that shows which AI bots crawl your site and separates training crawls, search crawls and live reads. The Pro add-on adds the visitors those assistants send you, paired per page with the reads that preceded them. Both are made by Forge Creative in New Zealand.