How to track referral traffic from Perplexity, Claude, Gemini and other AI assistants
Every assistant sends visits from its own address, and several also live inside something else: Google search, Bing, X, WhatsApp. What to look for, where those visits hide, and why an AI fetch in your logs isn't a visitor.
In this article
ChatGPT isn't the only assistant that puts links in its answers. Perplexity, Claude, Gemini, Copilot, DeepSeek, Grok, Meta AI and Mistral do too, and each one leaves a slightly different mark in your analytics. Some are easy to spot. Others look like Google search, Bing, X, or nothing at all.
This is a field guide to those eight: the address each one sends visits from, where their visits get mixed up with other sources, and how to tell a person clicking a link from an AI reading your page. For ChatGPT itself there's a separate guide on seeing whether ChatGPT sends you visitors. Everything here works in any analytics tool.
What an AI visit carries
An analytics tool has two ways to tell that a visit came from an assistant.
- The referrer. When someone clicks a link on a web page, the browser usually tells the next site where the click came from. Unless the page asks for something else, browsers send only the site's address, not the page's: a click from
https://example.com/some/pagearrives ashttps://example.com/. Under that default, a referrer from an assistant holds its address and nothing about the conversation, which is one reason no analytics tool can show you the question someone asked. - A utm_source tag. A link can carry its source in the address itself, as in
?utm_source=newsletter. Whether an assistant adds one, and what it writes, is up to the assistant and can change without notice. Because the tag is part of the link, it still works when the browser sends no referrer.
Neither is guaranteed. A page can tell the browser to send no referrer at all, an app has no web page to report, and plenty of links carry no tag. So count a visit as AI traffic if either signal names an assistant, and expect to miss some.
Assistant by assistant
These are the addresses each assistant runs on, which is what a referrer from it would show. The third column is where the same assistant also lives, and that's where visits go missing.
| Assistant | Address to look for | Also worth knowing |
|---|---|---|
| Perplexity | perplexity.ai, www.perplexity.ai | Perplexity also fetches pages while it answers a question. Those fetches are a server-log matter, not visits (more below). |
| Claude | claude.ai | Also a desktop app for Mac and Windows, and a phone app. Links opened from an app usually arrive with no referrer. |
| Gemini | gemini.google.com, bard.google.com (older data) | Called Bard until February 2024. The old address now redirects to the new one. AI Overviews and AI Mode in Google Search are not Gemini referrals. |
| Copilot | copilot.com, copilot.microsoft.com, copilot.cloud.microsoft | As of November 2026, copilot.microsoft.com redirects to copilot.com, and Microsoft is moving the Microsoft 365 Copilot web app to copilot.cloud.microsoft. Microsoft's AI search inside Bing runs on bing.com. |
| DeepSeek | chat.deepseek.com | Also a phone app. |
| Grok | grok.com | Also built into X, on x.com and in the X apps. |
| Meta AI | meta.ai | Also inside WhatsApp, Messenger, Instagram, Facebook, Meta's AI glasses and the Meta AI app. |
| Mistral | chat.mistral.ai | Mistral renamed its assistant from Le Chat to Vibe in 2026. The address stayed the same. |
Sources: Google on the Bard rename, Microsoft on the Microsoft 365 Copilot address, Bing on its AI search, Mistral on the Vibe rename, Meta's app listing, X on Grok and Claude's downloads page.
Addresses change, as Bard and Copilot both show. Use the table as a starting list, then compare it with the referrers in your own report. In the utm_source report, look for the same names (perplexity, claude, gemini, copilot, deepseek, grok, meta.ai, mistral) or any of the addresses above.
Where AI visits hide
Several of these assistants also live inside products that already have their own place in your reports. Visits from there won't carry the assistant's address.
In Google search
AI Overviews and AI Mode sit inside Google Search, so a click from them comes from Google. Google Analytics files them under Organic Search, and Search Console counts them as part of web search. A Gemini referral is one from gemini.google.com.
In Bing
Microsoft's AI search inside Bing, launched as Copilot Search in April 2025, runs on bing.com. A click from it looks like any other Bing search visit.
Inside X and Meta's apps
Grok used inside X starts on X, so a referrer, if there is one, names X rather than grok.com. Meta AI inside WhatsApp, Messenger, Instagram or Facebook starts inside an app, so don't expect meta.ai either.
In Direct
A link opened from a desktop or phone app reaches the browser with no web page behind it, so there's usually no referrer. The same goes for a link someone copies out of an answer and pastes into a new tab. Without a utm_source, these count as direct. There's more on what else ends up in Direct.
One filter for all of them
Most tools let you filter the source or referrer report. If yours accepts a regular expression, this pattern catches the names and addresses above, plus ChatGPT, so you get every assistant in one view:
perplexity|claude|gemini|bard\.google|copilot|deepseek|grok|meta\.ai|mistral|chatgpt|openaiThree things to check before you trust the result:
- Run it on both signals. If your tool shows the referrer and utm_source in separate reports, filter both. If it merges them into one source field, once is enough.
- Read the matches. Short names like grok, gemini, claude and mistral also turn up in other sites' addresses. If a match doesn't belong, tighten that part of the pattern, for example
gemini\.googleinstead ofgemini. - Check how your tool matches. Some tools compare the pattern with the whole value rather than any part of it, and some are case-sensitive. Google's help page on regular expressions in Analytics says both are the default there. If nothing matches, wrap the whole pattern in
.*(and).*, and look for capitalised values too.
Crawls are not clicks
AI companies don't only send people. Their software also fetches pages, and that shows up somewhere else: in your server logs or your host's request logs, under the fetcher's user agent. Three AI companies name their main ones like this, as of November 2026:
| Company | When a person asks | For search | For training |
|---|---|---|---|
| OpenAI | ChatGPT-User | OAI-SearchBot | GPTBot |
| Anthropic | Claude-User | Claude-SearchBot | ClaudeBot |
| Perplexity | Perplexity-User | PerplexityBot | None listed |
A fetch by one of the agents under "When a person asks" is worth knowing about: someone asked a question, and the assistant read your page to answer it. It isn't a visit. The person may never click.
These fetches usually don't reach analytics either. A study Vercel published in December 2024 found that the AI crawlers it measured didn't run JavaScript, among them OpenAI's three agents, ClaudeBot and PerplexityBot. The exceptions it named were Google's Gemini, which uses Googlebot's infrastructure, and Applebot. The study didn't cover every agent in the table, but a fetcher that doesn't run JavaScript never triggers an analytics script on the page.
So there are two numbers: fetches in the logs and clicks in analytics. Both are worth watching, but don't add them together. It's one of the reasons no two tools agree on your traffic.
Reading the numbers
- Look at landing pages, not only the total. Which pages assistants link to often tells you more than how many visits they sent.
- Read weeks, not days. If the numbers are small, a single day is noise.
- Treat the count as a minimum. App clicks, copied links and AI answers inside Google, Bing and X all go uncounted.
- Keep your own tags out of it. A link you post somewhere with
utm_source=perplexitycounts as a Perplexity visit, and then you can't tell your post from a real referral. Tag your own links with the place you posted them. A simple naming convention helps.
A small AI number is a floor, not a measurement.
A short checklist
- Check both the referrer and utm_source.
- Start from the address table, then compare it with your own referrer report.
- Remember the lookalikes: Google search, Bing, X, Meta's apps and Direct.
- Use one filter, and read the matches before you trust the total.
- Keep AI fetches in your server logs apart from visits in your analytics.
- Never put an assistant's name in your own UTM tags.
Lomen recognises visits from ChatGPT, Claude, Perplexity, Gemini, Copilot, DeepSeek, Grok, Meta AI and Mistral, by the referrer and by utm_source, and shows which of them sent people to your site. It reads the same two signals described above, so it has the same blind spots: a visit that names no assistant in either isn't counted as AI traffic. It counts from a script on the page, not from server logs, so for AI fetches you'll need your host's logs.