Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Number You Are Reporting Is Too Low
AI-referred sessions are undercounted in GA4 for a mechanical reason: a meaningful share of them arrive with no referrer header, and a session with no referrer and no campaign parameters is classified as Direct. Nothing downstream can recover the origin, because the browser never sent it. The undercount is structural, not a settings mistake.
This matters more than the usual attribution grumble, because the traffic being lost is not a random sample. It is disproportionately the traffic from native assistant apps, and assistants are where buyers now start a category. The result is a report that shows AI contributing a rounding error while the actual contribution sits in a Direct bucket that everyone reads as "people who already knew us".
Be careful with the statistics circulating on this. Vendor write-ups put the share of AI sessions arriving without a referrer anywhere from about a third to over two-thirds. That spread is wide enough to establish that the problem is real and too wide to plan against. None of those figures are measurements of your site, and the point of the next thirty minutes is to stop borrowing them.
You can state your AI-referred session count and name the specific reason it is still an undercount.
What GA4 Shipped, and the Two Holes It Left
On 13 May 2026 Google added an AI Assistant channel to the GA4 Default Channel Group. Sessions from recognised assistants are assigned the medium ai-assistant and grouped into the new channel with no configuration. Broad availability across properties followed in early June 2026.
This is a genuine improvement and it does not finish the job. The first hole is that the change is forward-only. It classifies sessions from 13 May 2026 onward and does not go back and reclassify what came before, so your historical AI trend stays buried inside Referral. Any year-over-year comparison you run through this channel will show growth that is partly an artefact of when the channel started.
The second hole is the one no vendor can close. If the referrer header is missing, there is no signal to classify. The session lands in Direct and stays there. A channel group is a rule for interpreting data that exists; it cannot invent data that does not. This is why the fix below has two independent halves, and why the second half matters more than the first.
| Situation | Where the session lands | What fixes it |
|---|---|---|
| Recognised assistant, referrer present, after May 2026 | AI Assistant channel | Already handled |
| Recognised assistant, referrer present, before May 2026 | Referral | Custom channel group, applied retroactively in reports |
| Assistant GA4 does not recognise | Referral | Custom channel group with your own source list |
| Native app click, no referrer sent | Direct | UTM on links you control; otherwise unrecoverable |
| Answer with a citation nobody clicked | No session at all | Not an analytics problem; see the AI impressions report |
Build the Custom Channel Group
A custom channel group is a reporting-time rule, which means it applies to historical data as well as new data. That single property is why it is worth building even though GA4 now has a native channel: it is how you get one consistent AI line across the whole period rather than a line that starts in May 2026.
- Create the group in Admin, not in a report.
Admin, then Data display, then Channel groups, then create a new group based on a copy of the default. Editing a copy keeps every existing channel definition intact so your other numbers do not shift underneath you.
- Add one AI channel above Referral.
Order matters: channel rules evaluate top-down and the first match wins. An AI rule placed below Referral will never fire, because Referral will have already claimed the session. Put it above Referral and below your paid channels.
- Match on source with a regex.
Use a condition of
Sourcematches regex, covering the assistants and AI browsers that matter in your market. A reasonable starting list is chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, and meta.ai. Add to it rather than trusting a list you copied, including this one. - Include the native medium too.
Add a second condition matching medium equal to
ai-assistant, joined with OR. This makes the group a superset of Google's native classification rather than a competitor to it, so you keep working when Google adds a source you have not listed. - Verify against a known session.
Ask an assistant a question your site should answer, click through, and confirm the session lands in the new channel within the realtime report. An unverified channel group is a configuration you believe in rather than one you know works.
Where UTMs Help, and Where They Quietly Lie
UTMs are the only mechanism that survives a missing referrer, because the origin travels inside the URL rather than in a header the app may drop. They work on every link you place yourself, and they do nothing at all for the case you most want to measure: an assistant citing your page from its own index.
Be precise about the boundary, because this is where reporting goes wrong in a way that looks like success. Tag the links you control: documentation you submit, partner placements, profile links, anything you syndicate. Those sessions will attribute cleanly forever. But an assistant that discovered your page through ordinary retrieval will link to your canonical URL with no parameters, and no amount of UTM discipline changes that.
The failure mode is subtle. If you tag some AI-adjacent placements and not others, your AI channel becomes a measure of where you remembered to add parameters, which then gets reported as a measure of AI performance. If you tag, tag by policy and write the policy down, so that six months later you know what the number is made of. Keep the parameters simple and consistent, and never put them on internal links, where they will overwrite the original source of the session and destroy the attribution you were trying to protect.
Carrying the Source Through to Revenue
A channel group tells you sessions arrived. It cannot tell you they were worth anything. The bridge is a hidden field on every form that captures the landing source, written into your CRM on the record, so the origin survives the gap between first touch and closed revenue.
The mechanism is unglamorous and takes an afternoon. On first visit, store the source, medium, and landing page in a first-party cookie. Populate hidden fields from that cookie on every form. Map those fields to properties on the lead or contact record. Now a deal that closes in ninety days still knows it started with an assistant, and you can compute a conversion rate per channel that means something rather than a session count that does not.
This is also what turns the AI conversation from a visibility argument into a commercial one. Session volume from assistants is usually small enough to dismiss; revenue per session frequently is not. Independent studies report AI-referred visitors converting better than ordinary organic clicks by margins that range from modest to dramatic across industries, which is exactly the kind of range that should make you measure your own rather than cite theirs. The page-side of that opportunity is covered in what to change when an assistant sends you a visitor.
Reading the Result Without Fooling Yourself
Report AI sessions as a floor, never as a count. The honest sentence is "at least this many", because the referrer gap guarantees an unknown quantity sits in Direct. A floor that rises over time is still a perfectly good trend line, and it is the strongest claim the data supports.
One diagnostic is worth running monthly: watch Direct sessions to deep pages. Direct traffic to a home page is plausible as genuine type-in behaviour. Direct traffic to a long-tail article URL that nobody would ever type is almost certainly a stripped referrer. If that segment is growing while your AI channel stays flat, the gap is widening and your floor is drifting further from the truth.
Then put the number somewhere it will be seen next to the rest of search performance, rather than in a separate AI deck that gets its own optimistic narrative. The structure for that, including which metrics earn a place in a monthly review and which are decoration, is in our SEO reporting framework, and the visibility half of the same picture now comes from Search Console's generative AI report.
Questions People Ask About Tracking AI Traffic
- Does GA4 track ChatGPT and Perplexity traffic automatically?
Partly, since 13 May 2026. Google added a native AI Assistant channel to the GA4 Default Channel Group that recognises referrers from assistants such as ChatGPT, Gemini, and Claude, assigning the medium value ai-assistant with no configuration required. It is forward-only, so it does not reclassify sessions recorded before that date, and it cannot classify a session that arrived with no referrer at all.
- Why does my ChatGPT traffic show up as Direct in GA4?
Because a referrer header never reached your site. Clicks originating in native desktop and mobile apps frequently arrive without one, and any session with no referrer and no UTM parameters is classified as Direct by definition. This is not a GA4 bug and no channel group can fix it, because the information was never sent. The only reliable countermeasure is a link-level UTM where you control the link.
- How much AI traffic is being misattributed?
Nobody credible knows for your site, and you should be suspicious of anyone who quotes a precise figure for theirs. Vendor estimates of AI sessions arriving without a referrer range from roughly a third to over two-thirds, which is a spread wide enough to tell you the problem is real and not much else. The useful move is measuring your own gap rather than adopting someone else's number.
- Should I build a custom channel group if GA4 has a native one?
Yes, for two reasons. The native channel starts on 13 May 2026 and does not rewrite history, so a custom group is how you get a consistent trend line across the period before it. And the native grouping does not necessarily capture every assistant or AI-powered browser you care about, so a custom group is where you add the ones that matter to your market.
- Is AI traffic worth measuring if the volume is small?
Usually yes, because the value per session tends to run higher than the volume suggests. Visitors arriving from an assistant have typically had the category explained to them already and are further into a decision. Independent studies report conversion uplifts against organic ranging from modest to very large depending on industry, so the honest approach is to measure your own rate rather than assume a multiplier.
