How-to

How to track Perplexity, Claude and Gemini referrals

ChatGPT is the one everybody names. The others are where the miscounting happens — two of these hosts were being added to the organic-search and referral numbers rather than left unattributed.

28 August 20266 min read

ChatGPT is the one everybody checks. These are where the miscounting happens

Not because the other assistants are harder — because their hostnames collide with rules you already have.

If you have set up any kind of AI-traffic tracking, you almost certainly matched chatgpt.com and stopped. That is a reasonable place to start and a bad place to finish, because the other assistants do not merely go uncounted — some of them get added to numbers you rely on for something else.

This post is the host-by-host reference, plus the two traps that made us change our own classifier. The pillar guide covers the concepts; this is the list.

The hosts, by assistant

What appears in your referrer when someone clicks through. Each of these is matched exactly, or as a true subdomain — never as a fragment of a longer name.

Assistant	Referrer hosts	Worth knowing
Perplexity	`perplexity.ai`	The straightforward one. A single host, no surprises.
Claude	`claude.ai`, `claude.com`	Both serve the assistant. Matching only the first loses every click from the second.
Gemini	`gemini.google.com`, `bard.google.com`, `aistudio.google.com`	The old Bard host still circulates. AI Studio is the trap — see below.
Copilot	`copilot.microsoft.com`, `copilot.cloud.microsoft`	Microsoft moved the product. The old host stays matched because links to it are still out there.
DeepSeek	`chat.deepseek.com`, `deepseek.com`	The apex serves the assistant as well as the marketing site.
Grok	`grok.com`, `x.ai`
Others	`you.com`, `phind.com`, `poe.com`, `meta.ai`, `mistral.ai`	Smaller by volume for most sites, and cheap to match.

That is 12 assistant products in total, last reviewed by a human on 2026-08-10. The count comes straight from the matching table, so it cannot drift away from what the product actually recognises.

The two hosts that were being counted as something else

Both fixed on our side. Both worth checking in whatever tool you use, because neither fails loudly.

Google AI Studio was being filed as organic search

Almost every channel-classification rule in existence identifies Google search with a broad pattern along the lines of google. — it has to, since there are dozens of country domains. aistudio.google.com matches that pattern.

So visitors arriving from an AI product were being added to the organic search number, which is a number people use to judge whether their SEO is working. Not merely uncounted as AI: actively inflating something else. The fix is order — the AI host list must be consulted before the search list is, so a more specific match wins.

Copilot's new host was falling through to generic referral

When Microsoft moved Copilot to copilot.cloud.microsoft, that host matched no AI rule and landed in the general referral bucket, next to forums and directories. Quieter than the AI Studio case, and the same shape of problem: a real assistant, silently filed as something ordinary.

Both were fixed in our classifier on 2026-08-10. If you maintain your own regular expression somewhere, these are the two to go and check right now.

What we deliberately do not match, and why

Two gaps that look like omissions and are decisions.

Assistant surfaces identified by a path

Some assistants live at a path on a bigger host — a chat surface under a search engine’s domain, or an assistant under a social network’s. It is tempting to match those, and we do not.

The reason is that a referrer is routinely delivered origin-only: many browsers and sites send just the hostname, with the path removed. So a path-based rule would fire for some visitors and not for others arriving from the very same product, depending on a privacy setting neither of you chose. You would get a number that is not wrong so much as arbitrary, and there is no way to tell from the report which visitors it captured.

A known gap you can reason about beats an inconsistency you cannot. So those arrivals stay in whatever channel their host implies, and we say so here rather than quietly banking a flattering number.

Two campaign tags that mean something else in the real world

When a visit has no referrer at all, a utm_source naming an assistant is read as a last resort. Two obvious-looking values are excluded from that:

utm_source=meta — on essentially every site that sets it, this is Meta Ads, not Meta AI.
utm_source=you — a word that appears in referral-programme links far more often than it names You.com.

Reading either would convert real paid-social and referral traffic into invented AI traffic, which is the exact failure this whole feature exists to avoid. The full hostnames meta.ai and you.com are still matched, because a hostname is unambiguous.

How to test each one yourself, in about fifteen minutes

The only evidence that is actually about your site.

Reading a table of hostnames tells you what to look for. Watching a real visit arrive tells you what your own setup does with it, which is the thing you actually need to know — and it is surprisingly quick.

Open a real-time view in whatever analytics you use, and exclude your own traffic first if you can, so you are not hunting for yourself among your own noise.
Ask each assistant a question your site genuinely answers. Something specific enough that you would expect to be cited — a product you sell, a guide you wrote, a question you rank for.
Click the citation when it appears, from a device you can recognise in the report.
Write down which channel each visit landed in. Do it for Perplexity, Claude and Gemini separately. They will not all behave the same, and the differences are the finding.

Try it once in a desktop browser and once from the assistant’s phone app. The app case is the interesting one: it usually opens a fresh browser context with no referring page at all, so the visit typically lands in direct — which is a large part of why AI referrals feel under-reported everywhere.

Why it is worth naming them separately at all

An 'AI traffic' total is a number you cannot do anything with.

The obvious objection to all of this is that assistants are a small slice for most sites, so why not keep one bucket. Three reasons, none of them speculative:

They behave differently, and you cannot see that in a total. One assistant may send fewer people who read far more; another may send browsers who bounce. That is a real difference in the value of being cited, and a combined figure averages it into nothing. We are deliberately not telling you which assistant is which — that is a fact about your site, and anyone quoting you industry averages for it is guessing.
They are cited for different pages. Which of your pages an assistant reaches for is a genuine signal about what you are considered authoritative on, and it varies by assistant.
They change independently. A hostname move, a product rename, or a shift in how one assistant links out will show up as an unexplained wobble in a combined number and as an obvious, attributable step in a per-assistant one.

What happens to your chart when a list like this changes

The part almost nobody discloses.

Here is a problem every analytics product with a matching list has, and very few mention. Classification is stamped onto each visit when it is recorded, and it is not revisited afterwards. So on the day aistudio.google.com was added to our AI list, every earlier visit from that host stayed filed as search — for ever.

The practical effect is a step in your chart on the day the list changed, which looks exactly like a real change in your traffic and is not one. If a tool never tells you this happened, you will eventually spend an afternoon explaining a jump that was an internal edit.

What Coruve does about it

Every change to the signature lists writes a note onto the affected charts, on the date it happened, saying that the list changed and what was added. The note is not optional and not a setting. Rewriting the historical rows would be the fuller fix and is not something the product can currently do — so the annotation is the mechanism, and we would rather show you the seam than let you find it yourself.

What if the assistant you care about is not on the list

A list like this is never finished, and the useful skill is finding the gap yourself.

New assistants launch, existing ones move hosts, and regional products serve audiences that a list assembled elsewhere will not have thought about. So it is worth knowing how to spot one rather than waiting for somebody to add it.

Sort your referrer report by volume and read the unfamiliar names. That is the whole technique. Any host you do not recognise is worth thirty seconds of looking up, and assistant hosts tend to stand out — they are usually a product name rather than a publication or a directory.

Two things to be careful about when you find one:

Match the host, not a fragment of it. A rule that looks for a product’s name anywhere in the referrer will also match domains that merely contain it, and those are registerable by anyone.
Check it is the assistant and not the company. Plenty of AI companies serve their marketing site and their product from related hostnames. A visit from a pricing page is a normal referral; only the assistant surface is an AI referral.

If you find one we are missing, tell us — a corpus like this improves by being corrected, and the alternative is that it quietly ages.

Checking your own site

Three ways, in ascending order of effort.

Run the free checker. The AI traffic checker reads seven days and shows the split per assistant, with its limits stated on the page. No account.
Search your referrer report for each host in the table above. Any analytics tool with a referrer list can do this, and it will tell you immediately whether the traffic is material for you.
Grep your server logs. More work, and the only method that sees fetches which never ran your analytics script — which is most crawler activity. Why that is true, and what it means for every number in this area, is the subject of the crawlers-versus-referrals post.

See which assistants send you people

The free checker reads seven days of your traffic and shows the split, with the caveats attached. No account, and nothing to install to see the answer.

Run the free checker All posts