You can't see the exact query. No AI assistant documents passing a user's prompt to the website they click through to — not ChatGPT, not Perplexity, not Gemini, Copilot or Claude. What you can know is which assistant sent the visitor, which page they landed on, and whether that visit became a lead and later revenue. Add one well-placed question to your form and you'll get closer to the real prompt than any analytics report can take you.
That same limit shapes the second question marketing teams keep asking: which ChatGPT prompts generate the most revenue? There's no report that answers it directly. But there is a method that gets you a defensible answer, and it's laid out in full below.
Key takeaways
- No AI assistant passes the user's prompt to your site. At most your analytics sees the assistant's domain, and sometimes not even that.
- ChatGPT adds
utm_source=chatgpt.comto links in its search results, according to OpenAI. That tells you the visit came from ChatGPT — not what was asked. - GA4 has had a built-in AI Assistant channel since 13 May 2026. It deliberately excludes Google's AI Overviews and AI Mode, which stay in Organic Search.
- The only source of actual prompt text is the lead themselves. A free-text "What did you ask?" field on your form beats every report here.
- "Which prompts drive revenue" is answered by joining three partial views — the landing page, self-reported prompts and prompt monitoring — to CRM revenue. No single tool does it for you.
Why the prompt never reaches your website
When someone clicks a link inside an AI answer, their browser loads your page and may send a Referer header saying where they came from. By default, modern browsers apply a strict-origin-when-cross-origin policy. MDN describes what that means for a click from one site to another: the browser sends "the origin (only)". So the most your server can receive is something like https://chatgpt.com/ — never the conversation URL, and never the prompt.
Often it receives even less. A link marked rel="noreferrer" is instructed, in MDN's words, "to omit the Referer header and otherwise leak no referrer information". A link that's copied out of a chat and pasted into a new tab carries no referrer at all. Those visits reach your analytics as direct traffic, with nothing to connect them to an AI assistant.
The assistants don't make up for this with their own reporting. OpenAI's help centre explains that ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to its search providers. That's traffic between OpenAI and a search partner, not something your site ever sees.
There's one narrow exception worth knowing. When ChatGPT writes a link itself, details from the conversation can end up inside the URL. OpenAI's own example is a car search: a link built for someone in London "transmits the information that you live in London to the car search service". That's the model composing a URL, not a referral feature, and it won't hand you the question that was asked.
Paying doesn't unlock the prompt either. ChatGPT now carries ads, and OpenAI is explicit that "advertisers never receive your chats, chat history, memories" and that they "only receive aggregate information about how their ads perform". Advertisers see impressions, clicks, spend and conversions — never the conversation that triggered the ad.
What each AI assistant actually tells you
Here's what each vendor documents about traffic it sends to websites, as of 30 September 2026. "Not documented" means the vendor publishes nothing either way. It doesn't mean nothing is sent, but you shouldn't build reporting on an assumption.
| Assistant | What arrives at your site | Can you see the prompt? | Reporting the vendor gives site owners |
|---|---|---|---|
| ChatGPT | utm_source=chatgpt.com on links in ChatGPT search results; a referrer, if any, is at most the origin |
No | Ads: aggregate metrics only. No report of which prompts cited your pages |
| Perplexity | Not documented | No | Analytics for Publisher Program partners; query text is not documented as included |
| Google AI Overviews and AI Mode | Counted as organic Google search | No | Search Console's generative AI report: impressions by page, country, device and date |
| Microsoft Copilot and Bing AI summaries | Not documented | No — "grounding queries" are grouped phrases, not user questions | Bing Webmaster Tools AI Performance: citations and grounding queries |
| Gemini app | Not documented | No | None documented |
| Claude | Not documented | No | None documented |
Two rows deserve a closer look.
The ChatGPT UTM is real but narrow. OpenAI's publishers FAQ says "ChatGPT automatically includes the UTM parameter utm_source=chatgpt.com in referral URLs". It says this in the context of ChatGPT search. OpenAI doesn't document the same for every link in every answer, so treat utm_source=chatgpt.com as a reliable floor for ChatGPT traffic, not a total.
Bing's AI Performance report is the nearest thing to query data any vendor provides, and Microsoft is careful to say what it isn't. Grounding queries are grouped and generalised, and the report "does not display individual user questions or highly specific long‑tail queries". Microsoft also notes that citations don't represent clicks. It shows where your content was used in Copilot answers, which is useful — but it can't tie a query to a visit, let alone a lead.
Step 1: Separate AI visits in your analytics
Before you can connect anything to revenue, AI traffic has to stop hiding inside "Referral".
GA4: the AI Assistant channel is already built in
If you last looked at this problem a year ago, GA4 has moved. Google added an AI Assistant channel to the default channel group; its What's new log dates it to 13 May 2026, describing a way to "identify how users are discovering your site through chatbots like ChatGPT, Gemini, and Claude". The channel definition covers sources "like ChatGPT, Gemini, Deepseek, Copilot, or Grok". When a visit's referrer matches Google's list, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant).
Read the exclusion carefully, because it trips up a lot of dashboards. The AI Assistant channel "excludes Google's AI Overviews and AI Mode". GA4's Organic Search channel explicitly includes them. So Google's own AI answers sit in Organic Search, blended with ordinary search clicks — more on that below.
When to build a custom channel group anyway
The default channel is a good start, but default channel groups can't be edited in GA4. You'll want a custom group if you need to split AI traffic by assistant, or to control exactly which sources count.
The setup is in Admin → Data display → Channel groups. You create a group by copying an existing one — the default, if you have none — and adding a channel whose condition is Source matches regex. Order matters: GA4 puts traffic in "the first channel whose definition it matches", and Google tells you to move your AI channel above Referrals.
Google's help page offers an example regex beginning ^.*ai|. Don't paste it in unchanged. That first alternative matches any source containing the letters "ai", so email sources like mail or gmail would land in your AI channel. A tighter pattern names the domains outright:
^(chatgpt\.com|chat\.openai\.com|(www\.)?perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai)$
Four documented details are worth knowing before you rely on it:
- It's retroactive in reports. Google states that custom channel groups "can be applied to your reports retroactively", so you see history without waiting.
- There are limits. A standard property gets 2 custom groups and a 360 property gets 5, each with up to 50 channels.
- It isn't in BigQuery. Custom channel groups "aren't available in the BigQuery Export Schema". If you join GA4 data to your CRM in BigQuery — which you'll likely need to, see Step 2 — you'll classify AI sources in SQL instead.
- The key events paths report can't use it. Plan other reports around that gap.
Adobe Analytics: set it up now, because it isn't retroactive
In Adobe Analytics, AI referrals belong in a Marketing Channel processing rule, under Analytics → Admin → Report Suites → Edit Settings → Marketing Channels → Marketing Channel Processing Rules. Adobe now offers a built-in condition, Referrer Is Conversational AI, backed by a "Conversational AI tools" referrer type with a published domain list. That list is also a sensible reference for your GA4 regex.
Two rules of thumb from Adobe's own FAQ. First, "Rules process in order", and referring-domain rules should be among the last. Second — and this is the one that costs teams months of data — "Changing processing rules do not affect data retroactively." Every week you wait is a week of AI traffic you'll never be able to separate.
Customer Journey Analytics: retroactive, but you build it
Customer Journey Analytics takes the opposite approach. Adobe says plainly that it "does not label these visits as AI traffic on its own". You build an AI-referral dimension yourself as a derived field, and Adobe publishes the recipe: use the URL Parse and Classify functions to classify referrers. Because data view settings "are retroactive and non-destructive", that field covers data you've already collected — the reverse of Adobe Analytics' processing rules.
Step 2: Put the source on the lead, not just the session
Analytics platforms record sessions. Your leads and your revenue live in a CRM. Unless you deliberately carry the source across, "AI Assistant" stays a traffic metric and never becomes a revenue one.
Capture these on every form, as hidden fields:
- The full landing page URL, including its query string. That's where
utm_source=chatgpt.comwill appear. - The referrer, read from
document.referrerat page load. It'll often be empty or origin-only, and that's still informative. - Your analytics identifier — the GA4 client ID, or your own user ID — so the lead can be matched to its session later.
Then decide how revenue gets back to analytics. GA4 ships lead-lifecycle events such as close_convert_lead, which feeds a Converted leads metric. But the timing limits matter for anything with a sales cycle. Measurement Protocol events "can be backdated up to 72 hours", and event data import only accepts timestamps from the previous two calendar days. A deal that closes six weeks after the first visit can't be pushed back onto that session through either route.
For real B2B cycles, do the join in BigQuery. Match the CRM record to GA4's user_pseudo_id, and read the source from the session_traffic_source_last_click record. Because custom channel groups don't reach BigQuery, classify the AI domains in your SQL using the same list as your regex.
Step 3: Ask the lead — the only way to get the actual prompt
Everything above tells you that a lead came from an AI assistant. Only the lead can tell you what they asked.
Add a "How did you hear about us?" question with an explicit AI option, followed by an optional free-text field: "What did you ask? Paste or paraphrase the question if you remember it." People who found you through an AI answer often remember the question well, because they typed it themselves minutes earlier.
A few practical points make the data usable:
- Keep it optional. A required field costs you leads, and forced answers are the least reliable ones.
- Ask where it doesn't add friction — the thank-you step, the demo-booking confirmation, or the first sales call.
- Normalise answers into themes. "Best GA4 consultants in the UK" and "who can migrate us from Universal Analytics" are two different buying intents, and your reporting should treat them that way.
- Treat it as a sample, not a census. Self-reported attribution is incomplete and leans on memory. It's still the only real prompt text you'll ever collect.
Which ChatGPT prompts generate the most revenue?
No tool reads this off a report, because nothing in the data carries the prompt to your revenue. You build the answer by joining three partial views to CRM revenue. Each is imperfect alone; together they're defensible.
View 1: the landing page stands in for the question. AI assistants cite pages that answer specific questions. A visit that lands on "How to migrate from Adobe Analytics to Customer Journey Analytics" tells you what problem the visitor was trying to solve, even without their words. Group your AI Assistant sessions by landing page, then follow those sessions through to leads and closed revenue.
View 2: self-reported prompts give you the words. The free-text answers from Step 3, grouped into themes, attach real questions to real deals. They also catch visits your analytics filed as direct.
View 3: prompt monitoring shows where you're cited. List the questions your buyers actually ask. Then check which of your pages the assistants cite for them — by hand, or with one of the monitoring tools built for it. Vendor data helps at the edges. Bing Webmaster Tools gives Copilot citations and grouped grounding queries, and Search Console's generative AI report shows which pages earn impressions in AI Overviews and AI Mode. Neither connects a query to a visit, but both tell you which pages are in play.
Join the three and the result looks like this:
| Prompt theme | Evidence | Landing page | AI sessions | Leads | Pipeline | Closed revenue |
|---|---|---|---|---|---|---|
| Choosing between GA4 and Adobe Analytics | Self-reported and landing page | Platform comparison guide | from GA4 | from CRM | from CRM | from CRM |
| Finding a migration partner | Self-reported | Partner directory page | from GA4 | from CRM | from CRM | from CRM |
| Fixing a specific tracking error | Landing page and cited in monitoring | Troubleshooting article | from GA4 | from CRM | from CRM | from CRM |
The themes in that table are illustrative — yours come from your own data. The structure is the point. The row that closes the most revenue is your answer to "which prompts generate the most revenue", with its evidence stated rather than assumed.
Expect the answer to surprise you. Pages that attract the most AI visits are rarely the ones that close the most revenue. Comparison and "who can help me with this" questions tend to be further down the funnel than the educational queries that drive volume.
The AI traffic you'll never see: the direct-traffic gap
Every number above is a floor. GA4 defines (direct) / (none) as traffic "that doesn't have a clear referral source". A meaningful share of AI-originated visits ends up there — the copied link, the stripped referrer, the link opened from somewhere that sends no referrer. The HTTP standard (RFC 9110) requires that a request from a source with no URL either excludes the Referer header or sends about:blank.
No vendor or standards body publishes what share of AI traffic arrives this way, and anyone quoting a precise figure is estimating. What you can do is watch for the signature. If direct traffic rises on the specific pages that AI assistants cite for your topics, while branded search stays flat, AI answers are the likeliest explanation. Your self-reported answers are how you size it.
Google AI Overviews and AI Mode: the hardest case
Google's own AI answers are the least visible of all.
In GA4 they're counted as Organic Search, not AI Assistant. In Search Console, Google states that AI features "are reported on in the Performance report, within the 'Web' search type". An AI Overview "occupies a single position in search results, and all links in the AI Overview are assigned that same position". A follow-up question in AI Mode counts as a new query.
Google added a dedicated generative AI performance report, announced in June 2026 and rolled out to all sites on 31 August 2026. It gives "dedicated views of your impressions within generative AI features on Search", broken down by page, country, device and date — and the same data stays included in your overall Web totals. It tells you which pages Google's AI features show. It doesn't give you the queries behind those impressions, and it can't separate an AI click from an ordinary organic one.
So the honest position for Google AI traffic is this. You can see which pages AI features surface. You can't isolate AI clicks from organic clicks, and you can't see what was asked.
Ordinary Google and Bing search is a different story — there, real query data does exist, in Search Console, the Google Ads search terms report and Bing Webmaster Tools. Our guide on how to know what users searched before clicking your website covers where to find it and its limits.
What not to do
- Don't buy a promise of "the ChatGPT prompt behind every lead." No vendor passes it. A tool that shows prompts is either reading self-reported data or running prompts itself to monitor citations — useful, but ask which before you pay.
- Don't paste Google's example regex unchanged. Its leading
.*aialternative sweeps mail sources into your AI channel. - Don't assume
utm_source=chatgpt.comis all your ChatGPT traffic. OpenAI documents it for search results, not every link. - Don't count citations as clicks. Microsoft says outright that its citation data doesn't represent traffic.
- Don't wait on Adobe Analytics. Its processing rules aren't retroactive, so the delay is permanent.
- Don't report AI traffic without its gap. Label every AI figure as a floor and show direct traffic beside it.
Where an implementation partner earns its fee
A single analyst can do most of Step 1 in an afternoon. The harder work is Steps 2 and 3: hidden fields that survive your form tool and CRM sync, a BigQuery join that holds up across the sales cycle, and a theme model for self-reported answers that sales and marketing both trust. That's where most teams stall, and where an experienced partner saves months.
If you're on GA4, you can compare GA4 implementation partners or start from our GA4 RFP template to brief one on exactly this scope. Adobe teams can compare Adobe Analytics partners and Customer Journey Analytics partners. If you're not sure which platform the problem belongs to, find a partner by need instead.
Sources
- OpenAI — Publishers and developers FAQ
- OpenAI — Searching the web with ChatGPT
- OpenAI — ChatGPT-generated links
- OpenAI — Ads in ChatGPT
- OpenAI — Testing ads in ChatGPT
- Google Analytics Help — Default channel group
- Google Analytics Help — Custom channel groups
- Google Analytics Help — What's new in Google Analytics
- Google Analytics — Sending Measurement Protocol events
- Google Search Central — AI features and your website
- Google Search Central Blog — Generative AI performance reports
- Search Console Help — Performance report metrics
- Bing Webmaster Tools — AI Performance
- Adobe Experience League — Marketing channel processing rules
- Adobe Experience League — Marketing channels FAQ
- Adobe Experience League — Referrer type dimension
- Adobe Experience League — Customer Journey Analytics derived fields
- Adobe Experience League — Classify AI traffic with derived fields
- Adobe Experience League — Data views are retroactive
- MDN — Referrer-Policy
- MDN — rel="noreferrer"
- RFC 9110 — HTTP Semantics, Referer
Frequently Asked Questions
Can I see the exact ChatGPT prompt that sent a visitor to my website?
No. No AI assistant documents passing a user's prompt to the site they click through to. At most your analytics receives the assistant's domain, and sometimes nothing at all. The only source of actual prompt text is the visitor, so add an optional "What did you ask?" field to your form.
Does ChatGPT add UTM parameters to links?
Yes, for ChatGPT search results. OpenAI's publishers FAQ says ChatGPT automatically includes utm_source=chatgpt.com in referral URLs. OpenAI doesn't document the same for every link in every answer, so treat it as a floor for ChatGPT traffic rather than a complete count.
Does GA4 track ChatGPT traffic automatically?
Yes. GA4 added an AI Assistant channel to its default channel group, which its What's new log dates to 13 May 2026. It covers sources such as ChatGPT, Gemini, Copilot, Deepseek and Grok, and sets the medium to ai-assistant when the referrer matches Google's list.
Are Google AI Overviews and AI Mode included in GA4's AI Assistant channel?
No. Google's definition says the AI Assistant channel excludes AI Overviews and AI Mode, and the Organic Search channel explicitly includes them. In Search Console they are counted within the Web search type, so AI clicks cannot be separated from ordinary organic clicks.
How do I find out which ChatGPT prompts generate the most revenue?
Join three partial views to your CRM revenue: the landing page as a stand-in for the question, self-reported prompts from your forms, and prompt monitoring that shows where assistants cite you. No single tool reports it, because no assistant passes the prompt through to your analytics.
Why does some AI traffic show up as direct in GA4?
Because the referrer is often missing. Browsers send at most the assistant's origin by default, links marked noreferrer send nothing, and a link copied out of a chat into a new tab carries no referrer at all. Treat your AI channel numbers as a floor, not a total.