Tracking AI Assistant Referrals in Google Analytics 4 and Matomo: A British Setup Guide, and the Two Lists That Disagree

Google Analytics 4 and Matomo disagree about what a visit from ChatGPT is, and the disagreement sits in two public files a British marketing team can read in ten minutes. Matomo's published search engine list, fetched on 16 October 2026, holds 296 named engines and ChatGPT and Perplexity are two of them. Google's published list of default-channel-group sources, fetched the same day, holds 819 rows, 129 of them categorised as search, and names neither. So one British visitor, one assistant, one afternoon, and two products put the arrival in two different channels. Settle which of the two definitions your reporting runs on, put that choice in writing, and only then build anything on top of it.

What Matomo's list actually says

Matomo keeps its engine definitions in an open repository, which is the useful part: you read the entry instead of asking support about it. Fetched on 16 October 2026, the file gives ChatGPT two hostnames and an empty parameter list, and Perplexity one hostname with a q parameter. Claude, Copilot and Gemini appear in neither that file nor the companion social networks file, which holds 88 entries.

Entry in Matomo's list, 16 October 2026	Hostnames recorded	Keyword parameter
ChatGPT	chat.openai.com, chatgpt.com	None, and the keyword is marked hidden
Perplexity	www.perplexity.ai	q
Claude, Copilot, Gemini	Not in the file	Not applicable

The empty parameter on the ChatGPT entry is the honest detail. Matomo will tell you a visit came from a search engine called ChatGPT and will not tell you what the person asked, because the question never travels in the URL. You gain a channel label and not a keyword, which is a fair trade as long as nobody in the meeting assumes otherwise. The Perplexity entry names one host with the www prefix, so an arrival from the bare domain is worth checking in your own data rather than assuming.

What Google Analytics 4 does with the same visit

Google takes a different route to the same decision. Its channel rules are published, and the one that catches an assistant is the Referral rule, which fires on a medium of "referral", "app" or "link" and is described as covering non-ad links from other sites and apps. Whether anything reaches Organic Search instead depends on a separate artefact: a spreadsheet of sources and categories, offered for download beside those rules.

That spreadsheet runs to 819 rows. Counting by category on 16 October 2026 gives 592 social, 129 search, 52 shopping and 46 video. We looked for engines whose absence would have shown the file had not loaded, and found Bing, DuckDuckGo, Yahoo across more than twenty country hosts, Baidu, Ecosia, Qwant, Naver and Yandex. Assistants are not among the 129. There is an address for asking Google to add a source, with a review cycle of at least annually and often sooner, which makes 129 a figure with a date on it rather than a constant.

For a British property the cost of leaving this alone is quiet rather than dramatic. Assistant arrivals settle into Referral beside trade press and partner links, the organic line keeps its shape, and a change in where enquiries come from never surfaces in a chart anybody reads.

The setup, in the order that avoids rework

• List the hostnames from your own data first. Open the referrer report for the last ninety days and read the hosts rather than guessing them. A British business to business audience rarely uses the same mix as a consumer brand's.
• In GA4, build a custom channel group with one condition per host. Keep it separate from Organic Search so the search line stays comparable with last year, and name it something a finance director understands without a glossary.
• In Matomo, confirm what the built in list is already doing. Because ChatGPT and Perplexity are in the shipped file, those visits may already sit under Search Engines. Check before adding a rule, or you will double count.
• Write the definition into the report. One line at the top: which hosts are counted, whose channel rules you followed, and the date. Two analysts comparing the tools will otherwise spend an afternoon rediscovering it.
• Segment the unattributed visits by landing page. Run as a desktop application, an assistant frequently sends no referrer at all, which is the same situation Google's help page describes when it lists links from offline documents such as PDFs and Word files among the causes of (direct) and (none). Watch for unsourced entries on narrow pages rather than on the home page.

The British layer: consent, and what you are allowed to record

Analytics cookies in the UK fall under PECR, which requires consent for storage that is not strictly necessary, alongside UK GDPR for the processing itself, and the Information Commissioner's Office publishes guidance on both. That bears directly on the numbers above, because the visitors who decline analytics storage are missing from every channel you report, assistant traffic included. The honest form of an assistant referral figure therefore carries the consent rate beside it. A share of sessions published with no note on how many visitors were measurable at all is describing a sample it has not sized.

Matomo is often chosen in the UK for exactly this reason, since it can be self hosted and run without cookies, which changes the consent question. That is a decision for your data protection lead, and it belongs before the channel group rather than after. The ICO's own domain turns up in our measurement work too: in the British half of our 15 October 2026 run, gov.uk was cited in 57 readings and ico.org.uk in 28, and neither appeared in the American set.

What the two products together still leave out

Arrivals are the only thing either product is built to count. A shortlist delivered inside a chat, three names long, with nobody sent anywhere, generates no arrival and therefore no row. Our own readings of that gap are published twice over, and the earlier one, taken on 15 October 2026, is what AI assistants actually cite across ChatGPT and Gemini in three markets. A second pass the next day asked 630 questions and held 1,260 readings. Narrow that pass to Britain and the count for our own domain reads 3 on ChatGPT, out of 210, against nothing at all on Gemini. Add the two passes up and Gemini has reached for us 0 times in 921 readings, which is the finding this programme has repeated most. Neither pass went beyond two engines, so there is nothing here about Perplexity, Claude or Google AI Overviews.

Whether any of this reaches an analytics report at all turns on the difference between a mention, a citation and a clickable link, and the scoring rules are in our 30 answer citation and link audit. If the query list for that exercise needs a starting point, take the full keyword set your site already ranks for, which a keyword site checker gives you in one pass.

Questions British teams ask about AI referrals in GA4 and Matomo

Why do GA4 and Matomo report different channels for the same visit?

Different source lists, nothing deeper. One of the two ships with ChatGPT and Perplexity written into its engine file, so the visit reads as search. The other works from 819 published source rows with no assistant in them, so the visit reads as a referral. Nominate one as your figure of record and say so on the report.

Does Matomo tell me what the buyer asked ChatGPT?

No, and the file says why: the ChatGPT entry carries no keyword parameter and marks the keyword as hidden, because the question is not in the URL. You learn the channel. To learn the question, put the question to the assistant yourself and record what comes back.

How does UK consent law affect these numbers?

PECR requires consent for non essential cookie storage and UK GDPR covers the processing, so your measured population is the consenting one. Record the consent rate next to any channel percentage you publish, and ask your data protection lead whether a cookieless Matomo configuration suits your site before you build reports on top of either tool.

Which hostnames should the channel group match?

Start from the hosts in your own referrer report over ninety days rather than a list copied from an article, then add the assistants you know your buyers use. Re-read Google's source spreadsheet and Matomo's engine file each quarter, since both are maintained files.

The visit you can count and the answer that came first

Set both tools up properly and you will know how many people reached your British site from an assistant and what they did next. That is worth having, and it is a small share of the story, because the assistant answered most of those buyers without sending them anywhere. The question worth asking this month is not which channel the clicks landed in. It is whether your company was one of the three names in the answer at all, on the questions your buyers type. We sell the measurement that reads that: AI Knows Us puts blind buying questions to ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, records every source behind every answer, and reports each engine separately because the two we have measured rarely agree. Our British number from that method was 3 citations in 420 readings on 16 October 2026, and it is on this page because a supplier who will not publish its own worst figure has not given you a way to judge its best one.