Most site owners we talk to have no idea whether ChatGPT, Perplexity or Copilot are sending them anyone. It is not an exotic measurement problem - the data is already in your analytics, under source names you have probably scrolled past.

This is how to find it, and how to read it once you have.

The source names

Assistant referrals arrive as ordinary HTTP referrers, so they land in your existing analytics as their own sources. The ones we see in practice:

chatgpt.com
Perplexity
copilot.com

In Plausible, Top Sources shows them by name with no configuration. In GA4 they appear under session source. There is nothing to install.

Two caveats. Not every assistant sends a referrer - some strip it, and those visits land in Direct. And people frequently read an answer, then type your domain in manually, which also lands in Direct. So whatever number you find is a floor, not a total.

Separate the humans from the bots first

Before you read any of it, you have to throw out the noise, because on a small site the noise is most of the traffic.

The tell is dwell time combined with bounce rate. Look at each source’s average visit duration:

A few seconds at a high bounce rate is automated traffic. It does not matter what the source claims to be.
Direct traffic with a very high bounce rate and near-zero duration is the usual bulk of it.
Referrals from large social platforms can be scrapers rather than people, especially if you have not posted there.

Assistant referrals look completely different, and that difference is the most useful thing in this whole exercise. In our own data, LLM-referred visitors dwell for minutes, not seconds - an order of magnitude longer than the automated traffic hitting the same pages. Someone who arrives from an assistant has already read a summary of you and decided to click through. They arrive further along than a search visitor.

So the first pass is: sort sources by visitors, then immediately re-sort by duration. The picture usually inverts.

Cross-slice source against landing page

Totals will not tell you anything actionable. The question worth answering is which pages assistants cite, and for that you filter by source and break down by entry page.

In Plausible you can do this straight from the dashboard by clicking a source to filter, then looking at Entry Pages. Via the Stats API it is a breakdown on visit:entry_page with a visit:source filter.

What you will probably find - we did - is that the distribution is brutally concentrated. A small number of pages take almost all of it, and the rest of the site takes nothing.

What actually gets cited

Looking at which of our own pages assistants send people to, the winners have an obvious shape in common, and it is not the shape most business blogs are written in.

Narrow beats broad. A page that answers one specific question gets cited. A page that surveys a topic does not. “What to check in a website audit” is a survey; it gets nothing. A comparison of named tools with an opinion about each gets cited repeatedly.

Current beats evergreen. Pages with a year in the title and genuinely current content do well. This runs directly against the standard SEO advice to write evergreen content and avoid dating yourself. Assistants are asked what is good now, and a page that commits to a date is easier to cite for that.

Opinionated beats neutral. “Where it works well” and “where it doesn’t” is citable. A neutral feature table is not. If your page has no position, there is nothing to quote.

Utilities get cited disproportionately. A single-purpose tool that answers a question someone would otherwise delegate - check this, look up that - punches far above its weight in referrals. Whether those referrals are worth anything commercially is a separate question, and worth asking honestly. Ours were not.

Add an llms.txt, with realistic expectations

An llms.txt at your root is a plain-text summary of your site for assistants: what you do, your key pages, how to contact you. It is cheap to add and it cannot hurt.

Be honest about what it does, though. We shipped one that catalogued every page we wanted cited, and it did not measurably redistribute anything - the same one or two pages kept taking the traffic. Assistants appear to cite what they have independently found useful rather than what you nominate. Treat llms.txt as making your site easier to summarise correctly, not as a placement mechanism.

While you are at it, check your robots.txt actually allows the assistant crawlers you want citing you. Blocking them and then wondering why you are not cited is a common own goal - our post on controlling AI crawlers covers which user agents to name and what llms.txt is really for. For the content side of it, see generative engine optimization.

Set up conversion goals before you optimise anything

This is the step everyone skips, ourselves included, and it invalidates everything downstream.

If you have not configured goals in your analytics, your form submissions and button clicks are being recorded as events but not surfaced as conversions. You can then spend months improving a funnel with no idea whether anyone has ever entered it. We had exactly this: the events were firing correctly and arriving, no goals were registered, and it turned out our main form had never been submitted. Not a low conversion rate - zero, for six months.

That single number would have changed what we worked on, and it was sitting there the whole time.

So: configure the goals, wait for real data, and only then decide whether the channel deserves more content. A traffic source you cannot connect to an outcome is a vanity metric no matter how good the dwell time looks.

The honest framing

LLM referral is a real channel and, from what we can see, a growing one. It is also small in absolute terms for most sites, and it converts on the same arithmetic as everything else. Doubling a small number gives you a slightly less small number.

Measure it, feed it cheaply with the kind of content that actually gets cited, and size your expectations to the traffic rather than to the growth rate.

How to Tell Whether LLMs Are Sending You Traffic