How Can I Find the Prompts or Intent Behind ChatGPT Referrals?

What data exists, and what does not?

Some ChatGPT visits arrive without a referrer, for example from the app, and appear as direct traffic. So analytics understates ChatGPT's share. How to set up tracking is covered in how to track ChatGPT and other AI referrals in GA4.

Can analytics show which AI search queries or prompts drove website visits?

No. No mainstream analytics tool receives prompts from ChatGPT, Gemini, Perplexity or Claude, and the assistants do not pass them. What analytics can show is the landing page and behaviour, which narrow the likely question considerably. A visit that lands on your pricing page and goes straight to a trial signup came from a different kind of question than one that lands on a blog guide and reads three articles. Treat any tool that claims to show exact prompts behind your traffic with caution. Ask where the data comes from. Often it is a panel of users or modelled estimates, not your visitors' actual conversations.

How do you infer intent from what you have?

1. Group ChatGPT sessions by landing page. Each landing page implies a small set of questions it answers.
2. Look at what visitors do next. Pricing visits followed by signups suggest a buying question. Documentation visits followed by support contacts suggest a "how do I" question.
3. Check server logs around the visit. A ChatGPT-User request for a page shortly before a ChatGPT visit to it suggests a user asked about that page or topic in a conversation.
4. Recreate the questions. Ask ChatGPT the likely questions for that landing page and see which ones cite it. Browser developer tools can also reveal the search queries ChatGPT runs when it searches, as covered in how to use browser tools to reveal ChatGPT's search queries and fan-outs.
5. Write down the inferred intents for each landing page and review them with sales and support.

How should you ask visitors?

• Add "AI assistant (ChatGPT, Gemini, Perplexity, Copilot)" to "how did you hear about us?" on forms.
• Add an optional follow-up: "What did you ask the assistant?" Short free-text answers are often surprisingly specific.
• Ask in sales calls and onboarding: "What were you looking for when you found us?"
• Use a short on-site prompt for visitors arriving from AI sources, asking what brought them. Keep it optional and unobtrusive.

Record the answers in your CRM so they can be counted. How to find the questions people ask about your business more widely is covered in how to find the questions people ask ChatGPT about your business.

How do you match log fetches to visits?

ChatGPT-User requests are the closest thing to a live signal that someone is asking about you. OpenAI describes ChatGPT-User as the agent that fetches a page when a user's request needs it, as opposed to OAI-SearchBot, which crawls to build the search index. So a ChatGPT-User request for your pricing page at 10:14 means that, around 10:14, a conversation needed your pricing page. It does not tell you what the question was, but the page narrows it.

A practical way to use this:

• Export ChatGPT-User requests from your server or CDN logs with URL and timestamp, and keep them separate from OAI-SearchBot.
• Count fetches by URL per week. Pages fetched often are pages conversations need, whether or not users click through.
• Compare fetches with visits. A page fetched 300 times with 10 ChatGPT visits is being read and summarised in answers; users get what they need without clicking. A page with few fetches but many visits is probably cited as a link people want to open.
• Look for clusters. Several product pages fetched within seconds of each other usually means one comparison question.
• Confirm the user agent against OpenAI's published IP ranges, since other bots copy the name.

The fetch-to-visit ratio is often more revealing than referrals alone. It shows which pages answer questions inside the conversation, which is where most AI influence happens.

What about tools that claim to show prompts?

Several visibility tools advertise "prompt volumes" or "the prompts people use". None of them can see your visitors' conversations. Their numbers come from one of three places: panels of users who agreed to share browsing or chat data, extrapolation from search keyword volumes, or prompts the tool itself runs. Each can be useful for spotting topics, but none tells you why a particular visitor arrived. Before paying for such data, ask how many real users the panel covers, which countries and devices it includes, and how a figure for your niche was estimated. Treat estimates for small or specialist categories with particular care, since panels are thinnest there.

What are the privacy limits?

Asking visitors what they typed is fine when the question is optional and clearly explained. Keep a few rules:

• Do not try to capture prompts by other means, such as scripts that read clipboard contents or text from other tabs. That is both unreliable and a serious privacy breach.
• Treat free-text answers as personal data where they might contain names, health details or company information. Store them with the rest of your form data, under the same retention rules.
• Report inferred intents in aggregate. "Price comparison" as a share of AI-referred sessions is useful. A log of individual guesses tied to named leads is rarely needed.

How do you turn this into something useful?

Combine the inferred intents and the visitors' own words into a question set. Track that set across assistants over time. Use it to:

• Improve the landing pages that receive AI visitors, so they confirm what the answer said.
• Find gaps: questions visitors mention that no page answers well.
• Prioritise content, based on which intents bring converting visitors.
• Brief sales on what AI-referred prospects already know when they arrive.

What might an illustrative case look like?

Consider an illustrative HR software company that saw steady ChatGPT referrals but no idea why. Grouping sessions by landing page showed most arrived on its "UK holiday entitlement calculator" page and its pricing page. Visitors to the calculator rarely converted, while pricing visitors often started trials. Server logs showed ChatGPT-User fetching the pricing page in bursts during working hours. The company added an optional "What did you ask the assistant?" field. Within a month, answers such as "HR software for a 30-person company under £5 per person" and "Is [company] cheaper than [competitor]?" showed that buyers were comparing prices. It added a comparison page and clearer per-person pricing, and trial conversions from ChatGPT referrals rose.

What do people ask about ChatGPT referral data?

Will OpenAI ever share prompts with websites?

There is no sign of it. Conversations are private, so plan around inference and direct feedback.

Why do ChatGPT visits sometimes show as direct traffic?

Some clicks, especially from apps, arrive without a referrer or tag. That traffic cannot be separated from other direct visits.

Are Perplexity and Gemini referrals any different?

The same limits apply. You see the source and landing page, not the question.

How many visitors will answer the "what did you ask" question?

Usually a minority. Even a small number of answers is valuable, because they are in the visitor's own words and often name the exact comparison or problem.

Does ChatGPT's agent or built-in browser change this?

OpenAI retired its Atlas browser in August 2026 and moved browsing into the ChatGPT desktop app, but the problem is the same. Agent-driven visits may load pages like a browser and may not carry a referral tag, so they can be harder to separate from other traffic. Server logs and on-site behaviour are the main clues.