# How can you see which prompts send visitors from AI assistants to your website?

> AI assistants never pass the prompt. How to segment AI traffic in GA4, match landing pages to cited prompts, and what our data says about lost referrers.

**Published:** 2026-11-04 · **Category:** GEO

[Version HTML](https://www.citeme.io/ressources/which-prompts-send-visitors-from-ai-assistants)

---

You cannot see which prompts send visitors from AI assistants to your website, because no assistant passes the prompt with the click. What you can do is reconstruct the most likely prompt: segment AI-referred visits in your analytics, look at the page each visitor landed on, and match that page against the prompts where an AI answer cites it. The result is a probable link, not a reading of the prompt.

**In short**: the click carries, at best, a label saying which assistant sent it. Often it carries nothing and lands in "direct". Prompt-level attribution comes from joining two data sets you control: landing pages and the prompts whose answers cite them.

## Why can't you see the prompt behind an AI visit?

Because the prompt never leaves the assistant. When someone clicks a source in an AI answer, their browser opens your page like any other link. Google stopped passing search keywords to sites years ago, and AI assistants never started. What reaches your site is, at most, one of three labels:

-   a Referer header from the assistant's domain, such as chatgpt.com or perplexity.ai;
-   a utm\_source tag added to the link, such as utm\_source=chatgpt.com;
-   on Android, the package name of the app that opened the link.

All three are set by the assistant or its app, not by you. Any of them can disappear with a release.

## What actually reaches your analytics

We measured this in two studies on our own data, published in the [Citeme research](https://www.citeme.io/research) section.

### Most assistants send at least some clicks with no label

In October 2026, in our lab, we asked nine assistants on the web and on Android to open and cite a test page, then clicked their citations. Seven of the nine sent at least one click that carried nothing at all: ChatGPT, Claude, Gemini, Meta AI, DeepSeek, Grok and Mistral. Only Perplexity and Kimi labelled every click we saw. Of the 42 clicks that came from an assistant (the lab logged 62 clicks in all, including test and control clicks), 30 carried no label. An analytics tool files those as direct traffic.

The label also depends on the surface, not only on the engine. Gemini and Meta AI kept their Referer on the web and lost it in their Android apps. In one ChatGPT answer on Android, the inline link carried a tag and the Sources sheet did not. Full method and limits are in [Who read this page?](https://www.citeme.io/research/who-read-this-page)

### ChatGPT's tag vanished for five days

From 10 to 14 October 2026, ChatGPT's citation links went out without their utm\_source tag. On the ten client sites with ChatGPT traffic before, during and after, tagged ChatGPT clicks fell from 40.1 a day to 16.2, a 60% drop. ChatGPT's live reads of the same sites barely moved, from 300 a day to 261. The likely explanation is that people kept clicking and the clicks lost their label. A dashboard that reads labels alone would have reported a collapse in AI traffic that probably never happened.

This is where the Beacon earns its place. Citeme's Beacon does not rely on the utm tag alone: it also reads the Referer and the app name, and it logs the assistant's own reads of your pages, which no missing tag can hide. Those reads are how we saw ChatGPT keep reading these sites while its tag was gone. For a landing with no label, the Beacon pairs it with the latest AI read of the same page, a match that named the right assistant in 7 of 7 held-out lab clicks (Step 3 covers why it works less often on real sites).

![Citeme settings for the source of AI visits, with the CiteMe beacon selected and code to send conversions from the site](https://cdn.sanity.io/images/xc5dg2px/production/132eb8b90d56b07d3d85ffd5b7674c7be0583962-4440x1726.png)

_The Beacon is the default source of AI visits, installed as a script, a Cloudflare worker or a Shopify app, with PostHog, Plausible and GA4 as alternatives._

### ChatGPT sends most of the labelled clicks

In an earlier study, we logged 385,057 AI-agent requests on twenty sites from January to September 2026. Of the 3,228 human visits we could attribute to an engine, ChatGPT sent 80.3%, Gemini 7.8%, Perplexity 6.2%, Claude 4.0% and Copilot 1.4%. Meta AI was the largest crawler, at 28.2% of all crawl volume in the study, and sent no human visit we could attribute. Details are in [What is actually crawling you](https://www.citeme.io/research/what-is-actually-crawling-you).

Google AI Overviews and AI Mode are a separate case. In our lab, a copy of AI Mode's link markup sent a www.google.com Referer, the same as an organic result. We did not test AI Overviews clicks directly, but if they behave the same way, analytics cannot tell them apart from organic search.

> **In short**: labelled AI clicks are a floor. ChatGPT dominates them, a share of every assistant's clicks arrives as direct, and Google's AI clicks probably look like organic search.

## Step 1: segment AI traffic in GA4

Start by isolating the visits you can label. In GA4, create a custom channel group (Admin, Data display, Channel groups) with a channel named "AI assistants" placed above Referral, using this condition: Session source, partially matches regex:

```
chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|chat\.mistral\.ai|deepseek\.com|grok\.com|meta\.ai|kimi\.com
```

The regex catches both the Referer domain and utm\_source=chatgpt.com, since GA4 reads the tag into Session source. Check your own source list for assistants missing from it before you rely on it. Google has also added an AI Assistant group to GA4's default channels (release note of 13 May 2026); compare its count with yours.

Then fix the two traps we found on client sites:

-   Paid clicks that look like citations. One client site has run a paid campaign in ChatGPT since 8 October 2026, tagged utm\_source=chatgpt&utm\_medium=paid. Counting utm\_source alone filed 100 landings on one campaign page in nine days as organic citations. Add a condition that excludes Session medium = paid (or cpc) from the AI channel.
-   Tags copied onto internal links. Some sites carry the landing's tags onto every internal link, so one visitor becomes several AI visits. Across at least 15 client sites we found 551 such repeats, and on the worst site they were 16% of its ChatGPT clicks. Count sessions, not page views, and check that your internal links do not carry utm parameters.

Finally, open the Landing page report filtered on your new channel. That list of pages is the raw material for the next step.

> **In short**: one regex gives you labelled AI sessions by landing page. Remove paid ChatGPT clicks and repeated tags before you trust the number.

## Step 2: match landing pages to the prompts that cite them

An AI-referred visitor almost always lands on a page the assistant cited. So the question becomes: for which prompts does an AI answer cite this page? If you track a set of prompts across engines, you already have that list.

![Citeme view of a ChatGPT answer to a French prompt that cites the brand first, with the sources listed beside it](https://cdn.sanity.io/images/xc5dg2px/production/48985f45d81a4ebddf4bf9cac12ebb7da770e4e8-2632x2576.png)

_One tracked prompt in Citeme: the ChatGPT answer, the brand's position and the five sources it cited, the URLs you join against your landing pages._

The method fits in a spreadsheet:

1.  Export AI-referred sessions by landing page and by assistant from GA4.
2.  Export the prompts you track, with the engine and the URLs cited in each answer.
3.  Join the two on the URL and the engine. Each landing page now has a list of candidate prompts.
4.  Weight the candidates. A page cited for one prompt gives a strong guess. A homepage cited for twenty prompts gives a weak one.

The quality of the result depends on your pages more than on your tools. A page that answers one precise question ("how much does X cost for a team of 10") ties back to a prompt cleanly. A generic page does not.

It also depends on your prompt list. If buyers reach a page through a question you do not track, the join cannot find it. When a page gets AI visits but matches none of your prompts, treat it as a hint that a prompt is missing from your list.

> **In short**: landing page plus the prompts that cite it gives a probable prompt per visit. Precise pages give precise answers, generic pages give guesses.

## Step 3: what about the visits with no label?

You can recover only a few of them, one by one.

When an assistant reads your page live to write an answer, its fetcher hits your server just before the person clicks. In our lab, every assistant that read the page live could be identified from its requests, and matching a click with no label to the most recent read of the page (the assistant's server fetching it) named the right assistant in 7 of 7 clicks we had set aside to test the match. But the lab forced a live read, because every prompt contained the page address.

On real sites the picture changes. We hid the label on 2,179 ChatGPT clicks and looked for a ChatGPT read of the same page in the two minutes before. Only 1.1% had one beyond what chance produces. ChatGPT mostly answers from copies it already holds, so there is usually no fresh read to match. Webmail links also arrive cross-site with no Referer, which makes them look like assistant clicks.

What you can do with unlabelled traffic is watch it in aggregate. A rise in direct sessions on pages that AI answers cite, at the same time as a drop in labelled AI sessions, is the pattern we saw during ChatGPT's tag outage.

> **In short**: matching server reads to clicks works when an assistant reads live, which ChatGPT rarely does. Treat unlabelled AI traffic as a trend, not as a list of visits.

## Doing it with a tool: Citeme AI Impact

Citeme is our product, and AI Impact automates the method above.

-   A script or a one-click Cloudflare connection records every visit and sorts it into four intents: training crawl, AI search indexing, real-time retrieval, and a human arriving from an AI answer. Only the last one can buy.
-   Each human visit is matched to the AI answers that cite its landing page, across the prompts Citeme tracks, and shown with its most likely prompt. The product labels it as probable, and it is only as fresh as the last audit.
-   When a visitor gives an email through one of your forms, their earlier anonymous visits are merged into one record. Purchases and signups are recorded with their value over a 30-day window.
-   With HubSpot, GoHighLevel or Attio connected, deals your team marks as won come back as CRM-verified revenue, per currency. On Stripe, Citeme computes a verified AI MRR from subscriptions actually charging.

The limits are the ones described in this article. The prompt link is an estimate, strong on a page cited for one prompt and weak on a homepage. A click that arrives with no label, and no live read just before it, cannot be tied to an assistant. AI Impact is included from the Growth plan ($149 a month); see [AI Impact](https://www.citeme.io/features/ai-impact), the crawler side in [AI search analytics](https://www.citeme.io/features/analytics), and plans on the [pricing page](https://www.citeme.io/pricing).

![Citeme Live AI fetches screen with weekly citations of the site, cited pages with their clicks, and pages cited but never visited](https://cdn.sanity.io/images/xc5dg2px/production/244a0e47877242bbd552540a8d4f70a5ce2ebdf4-5120x2576.png)

_Pages that AI answers cite, with their citation and click counts, next to a list of cited pages that nobody clicked._

> **In short**: AI Impact joins visits, landing pages and tracked prompts automatically, then follows the visitor to a conversion or a won deal. The prompt shown is the most likely one: a strong guess on a page cited for one prompt, never a certainty.

## FAQ

### Can Google Analytics show which ChatGPT prompt sent a visitor?

No. GA4 sees the source of a session, such as chatgpt.com, when ChatGPT passes a Referer or a utm_source tag. It never sees the prompt, because ChatGPT does not send it. You can only infer a likely prompt by matching the landing page with the prompts whose answers cite it.

### Why does AI traffic show up as direct in GA4?

Because many AI clicks arrive with no Referer and no tag. In our October 2026 lab, seven of nine assistants sent at least one click with no label, and 30 of the 42 clicks that came from an assistant carried nothing. GA4 files those sessions as direct.

### What does utm_source=chatgpt.com mean?

It is a tag ChatGPT adds to many of its citation links so the site can see the visit came from ChatGPT. It is not guaranteed: from 10 to 14 October 2026 the tag was missing, and tagged ChatGPT clicks fell 60% on the ten client sites we measured. A pasted link keeps the tag too, so it can also appear when someone shares a ChatGPT link elsewhere.

### Can I tell AI Overviews clicks apart from Google organic clicks?

Not reliably from the click. In our lab, a copy of Google AI Mode's link markup arrived with a www.google.com Referer, like an organic result, and we have no sign that AI Overviews clicks differ. To estimate them, track which of your pages AI Overviews cites for your target queries, for example with the free AI Overviews tracker.

### How accurate is prompt-level attribution?

It depends on the page. A page cited for one prompt gives a strong guess. A homepage cited for many prompts gives a weak one. In both cases the prompt is probable, never certain. It also depends on tracking the right prompts: a visit from a question you do not track cannot be matched to it.

## Sources

- [Who read this page? Identifying AI assistants from their network calls (October 2026)](https://www.citeme.io/research/who-read-this-page) · Citeme Research
- [What is actually crawling you (September 2026)](https://www.citeme.io/research/what-is-actually-crawling-you) · Citeme Research
- [AI Impact](https://www.citeme.io/features/ai-impact) · Citeme