The AI traffic you read in GA4 is a floor, and no setting will turn it into a total. GA4 only recognizes a visit from an assistant when a signal comes with it, and that signal depends on a link parameter and a header you do not control. What you can have is a clean floor: a written, tested, last-click rule that compares cleanly from one month to the next.
That floor is enough to run the tracking. It is not enough to prove a return on investment, and it cannot be fixed with a "hidden traffic" multiplier, because as of 3 November 2026 no public measurement supports one.
This stopped being theoretical in October. For about a week, teams watched ChatGPT fall in their reports while its links had, at least in part, lost their tag. A number with no documented rule does not survive that kind of episode.
Doesn't GA4 already count AI traffic?
Partly, since 13 May 2026. Google added an AI Assistant channel to the default channel group that day: the medium becomes ai-assistant and the campaign (ai-assistant) when the referrer matches a list of recognized assistants.
Referred AI traffic is the set of visits an analytics tool can tie to a conversational assistant, through its referrer or a link parameter. Anything else may exist, but it cannot be measured.
The native channel leaves three questions open:
- Which assistants? The announcement names ChatGPT, Gemini and Claude. The channel definition names ChatGPT, Gemini, Deepseek, Copilot and Grok. Perplexity appears in the 11 June note on the Source Group field. The help page points to a general spreadsheet of default channel group sources, without saying whether it covers assistants: open it before relying on it.
- Which visits? A tagged visit with no medium would miss the channel. Clickport (David Karpik, October 2026) applies Google's documented rules:
utm_source=chatgpt.comwith no medium giveschatgpt.com / (not set), hence Unassigned. He states he has not tested this in a GA4 property, and step 1 below checks it on yours. - Since when? Google does not say whether the channel applies to past data. The SEO Works wrote its own regex for its twelve-month benchmark because, in its account, the category only exists from June 2026.
AI Overviews and AI Mode stay outside this channel, in Organic Search, and under Web in Search Console. Search Console's generative AI report, open to all sites since late August 2026, gives impressions. For the French picture, see the first AI Overviews figures in France.
How do you build a reliable channel group?
Start from a copy of the default group and set a condition on source rather than referrer. You need the Editor role, and a standard property allows only two custom groups. Here is a starting regex:
(.*\.)?(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|copilot\.com|claude\.ai|chat\.mistral\.ai|meta\.ai|grok\.com|chat\.deepseek\.com)
GA4 uses full matching by default, so this pattern accepts subdomains and rejects lookalikes. It was tested outside GA4, in full-match mode, on 27 values: 14 assistant hosts accepted, 13 lookalikes rejected, including gemini.com, copilot.github.com, metaxai.com and a fake chatgpt.com subdomain. It is a base, not an exhaustive list.
Google's help page offers its own example regex, a very broad one. Run through the same kind of test, it misses grok.com and chat.deepseek.com, and accepts bombardier.com, lombard.fr, gptzero.me and gemini.com, because of patterns such as .*bard.* and .*gpt.*. The gap may be zero on your site. Check it rather than assume it.
- Audit. In Reports > Acquisition > Traffic acquisition, switch the dimension to Session source / medium and search for chatgpt, perplexity, gemini, claude, copilot. Note each variant:
/ referral,/ ai-assistant,/ (not set). Add any missing host to the regex. - Create the group. Admin > Data display > Channel groups > Create new channel group. You start from a copy of the default group.
- Add the channel. Add new channel, name it "AI assistants", then + Add condition group. Pick the Source dimension and "matches regex", paste the pattern, then Save channel.
- Order it. Traffic goes into the first channel whose definition it matches. With Reorder, place "AI assistants" below the paid channels, so a sponsored link stays paid, above Referral as Google instructs, and above the native AI Assistant channel. Then Apply and Save group.
- Keep the native channel as an alarm. Right below yours, it only receives assistants Google recognizes and your regex misses. When a row appears there, extend the pattern.
- Check. Select the group as primary dimension in Traffic acquisition. In a free-form exploration, filter Session source with the same pattern: over the same period, totals should match, paid visits aside. The
chatgpt.com / (not set)row, if you have one, should have left Unassigned.
Google says these groups apply to your reports retroactively. Still date each regex version in a shared document: history recalculated after a rule change looks like a trend.
Why do your numbers match no other tool?
Because each tool applies its own rule, and few publish it. A gap does not mean a tool is wrong. Most of the time it means they are not counting the same thing.
| Source of the gap | What changes |
|---|---|
| Source list | A short list drops minor assistants, a long one adds them up |
| Match mode | Strict equality misses host variants, "contains" catches lookalikes |
| Tagged visits with no referrer | Counted if the tool reads utm_source, lost if it only reads the referrer |
| Sponsored links | Filed as paid or as AI, depending on the rule |
| Attribution scope | Session in Traffic acquisition, the selected model (data-driven by default) for event-scoped dimensions |
| Collection | Consent, blocked scripts and bot filtering differ across tools |
Public panels show the same thing. Ahrefs measures 0.46% AI traffic in October 2026 across 118,671 sites. SE Ranking measures 0.32% across 101,574 sites from January to April 2026, and 0.2689% in the EU, down from 0.2858% to 0.2506% between February and April. The split diverges even more. At Ahrefs, Gemini is about 3% of October's AI visits (183,300 of 5.4 million, a ratio computed from the published figures), behind Meta AI. At SE Ranking, it is 11.56%, and Perplexity 7.23%. Neither panel explains the gap with the other.
So do not try to reconcile tools. Pick one reference, write down its rule, and track how it moves.
How much AI traffic is hiding in Direct?
Nobody knows, as of 3 November 2026. The figures in circulation do not answer that question, even when presented as if they do.
- 70.6% at Loamly (February 2026). 14,413 visits with no referrer out of 20,428 AI visits, in a database of 446,405 visits. Those visits are detected by the vendor's model, and the vendor sells that detection. No period is given.
- 35.7% rising to 49.7% at Clickport (April and October 2026). These visits carry a tag and no referrer: 935 of 2,619, then 958 of 1,928. They are not lost, since the tag identifies them. The vendor warns its sample is uneven and self-selected, and counts identified visits only.
- 0.003% at Kevin Indig (Growth Memo, 19 October 2026). The share of Direct with an identifiable AI trace, averaged over 12 B2B-leaning sites. From January to September 2026, engaged Direct stayed flat at -1.0% while total Direct gained 140.9%, largely from bots. His conclusion: "An effect could still exist. I just couldn't isolate it."
David Rosam traced the often repeated "60%" back to a Loamly estimate in September 2026, and restates a distinction the retellings erase: the share of AI visits with no referrer is not the share of Direct that comes from AI. Nobody, he writes, has measured the second.
October's incident is the clearest clue, though it gives no lasting order of magnitude. As reported by Search Engine Journal, Cedric Huesler (Adobe) compared 9 to 13 October 2026 with the week before, on about 500 pages where ChatGPT drives at least 5% of visits: about 31,000 fewer ChatGPT visits a day, about 26,000 more Direct visits, on four days of sampled data. He ties the fall to links sent out with no tag and no referrer, and says OpenAI confirmed a bug to him on 14 October, fixed from the 15th. He adds that most ChatGPT clicks still arrive with no referrer, and advises counting the tag. Search Engine Journal notes that OpenAI's release notes do not mention link tracking, and quotes another vendor for whom aggregated data cannot separate a lost visit from a mislabeled one.
Three practical consequences. Annotate 8 to 15 October 2026 in your reports, approximate bounds, and log it alongside model changes: GPT-6 reached ChatGPT the day before, another break that moves the numbers without your brand doing anything (see how to date a series break). Track the share of AI visits that rests on the tag alone: it measures your dependence on a parameter you do not control. And do not multiply your number by a borrowed coefficient.
What can UTMs and server logs recover?
Better signal on what already exists, not the invisible part. You cannot tag links an assistant writes for you. You can stop losing the ones that arrive.
- Protect parameters. Check that redirects (http to https, www, language, trailing slash) keep
utm_source. A tag dropped at a redirect turns an identifiable visit into an anonymous one. - Tag what you publish. On links you place in a custom GPT or third-party tool, use an explicit source and medium. Never on internal links.
- Read logs for what they are. They typically keep the full URL, referrer and user agent, even with no analytics script or consent. So they let you check the gap with GA4 on tagged URLs.
Watch for the opposite trap. OpenAI says ChatGPT-User may visit a page when a user asks a question. That is a robot reading, not a click: adding it to sessions would inflate your number, and its handling belongs in the AI bots guide. A visit with no tag and no referrer stays, in logs too, a visit with no origin.
Finally, a rise in Direct proves nothing on its own. Indig's data points to bots first, which is why engaged sessions are the line to read.
Does AI traffic really convert better?
That is not established. The largest study puts organic ahead, the best documented paired test finds no significant gap, and the others disagree.
| Study | Sample | Result |
|---|---|---|
| Kaiser and Schulze, 2025-2026 | 973 e-commerce sites, Aug 2024 to Jul 2025, last click | Organic about 13% ahead of ChatGPT, which beats only paid social |
| Amsive, September 2025 | 54 sites, paired test | 4.87% vs 4.60% for organic, p = 0.794 |
| The SEO Works, 2026 | 63 sites, Sep 2025 to Aug 2026 | 7.18% vs 6.08%, AI worse on 23 sites, 0.62x on B2B products |
| Marie Haynes, October 2026 | 100 sites over 12 months, 55 of them tracking conversions | 13.8% for ChatGPT vs 3.02% for Google organic, no statistical test shown |
Sources: see the list at the end. Conversion definitions differ across studies.
None of these studies resolves selection: someone clicking out of an answer may already be further along. To run the comparison properly on your data:
- Compare by page type. An AI visit landing on a product page and an organic visit landing on an article are not the same contest.
- Set a threshold. The SEO Works requires 100 AI sessions and 1,000 organic sessions. That is a minimum: with 10 conversions on 200 AI sessions (5%) against 1,200 on 30,000 organic (4%), the 95% confidence interval for AI runs from 2.7% to 9.0%, which settles nothing. Arithmetic illustration, fictional numbers.
- Freeze attribution. Paid and organic last click ignores Direct, according to Google. Keep the same model and the same report month after month.
- Look at revenue per session, not just rate. On revenue per session, Kaiser and Schulze find the same order: ChatGPT ahead of paid social and behind the other channels.
What should you report to your CMO?
A floor, its rule and a trend, on one page. That format answers a leadership team without promising what the measurement cannot say.
| Line | How to read it |
|---|---|
| AI sessions and share of traffic | Last-click floor, dated rule |
| Attributed key events and revenue | What is observed, not the total effect |
| Six-month trend | With the 8 to 15 October note and each regex change |
| AI vs organic, by page type | Only above the threshold you set |
| Share resting on the tag alone | The risk of a measurement break |
| Generative impressions (Search Console) | Exposure, no clicks, read alongside |
Three sentences to avoid: "x% of our AI traffic is hidden", "AI converts x times better" in general, and any return on investment drawn from this table alone. The incrementality protocol also treats these visits as a last-click floor, and the GEO, SEO and paid budget decision uses them as its lower bound. Influence without a click gets tested. It cannot be deduced from GA4.
What remains is the repetitive part: replaying this reading each month, by assistant, landing page and market, and reading it with your presence in answers. The AI Traffic page of Nanga's GEO module reads GA4 once the connection is set up with the Nanga team. It covers acquisition and engagement, landing pages, mobile and desktop, country and language. It applies the same AI traffic rule as Nanga's SEO module, and sits alongside AI Visibility Index tracking, weekly by default and daily as an option.
To see AI traffic and visibility in answers in one place: request a demo.
Frequently asked questions
What is the AI Assistant channel in GA4?
It is a channel in the default group, announced by Google on 13 May 2026. It collects visits whose referrer matches a list of recognized assistants, such as ChatGPT or Gemini. AI Overviews and AI Mode are excluded and stay in Organic Search.
Why does my AI traffic regex give a different number from the native channel?
Because it reads the source, while the native channel depends on the referrer or medium. According to Clickport, a visit tagged utm_source=chatgpt.com with no medium lands in Unassigned in the default group. The author has not tested this in GA4, so check your own property.
Should I adjust GA4 AI traffic with a hidden traffic multiplier?
No. Public figures run from 35.7% to 49.7% of tagged visits with no referrer at Clickport to 70.6% of visits with no referrer at Loamly, on different measures, and Kevin Indig finds almost no AI traces in Direct. None of them supports a multiplier.
Does AI traffic convert better than SEO?
That is not established. Kaiser and Schulze, on 973 e-commerce sites, put organic about 13% ahead of ChatGPT. Amsive, on 54 sites, finds no significant difference. Marie Haynes, on 100 sites, sees a large premium, with no statistical test. Compare on your own pages, by page type.
Can I see AI Overviews clicks in GA4?
Not separately. Google says this traffic is included in search data, under Web in Search Console, and its generative AI report announced on 3 June 2026 gives impressions. In GA4 these visits stay in Organic Search.
Sources:
- Google, New AI Assistant traffic measurement and New Source Group field, What's new in Google Analytics, May and June 2026.
- Google, Default channel group, accessed November 2026.
- Google, Custom channel groups, accessed November 2026.
- Google, About regular expressions (regex), accessed November 2026.
- Google, Scopes of traffic-source dimensions, accessed November 2026.
- Google, Get started with attribution, accessed November 2026.
- Google Search Central, AI features and your website, accessed November 2026.
- Google Search Central, Introducing Search Generative AI performance reports in Search Console, June 2026, updated August 2026.
- OpenAI, Overview of OpenAI Crawlers, accessed November 2026.
- Ahrefs, AI vs Search Traffic Analysis, October 2026 data.
- SE Ranking, Analysis of Top AI Search Engines: Who Is Catching Up to ChatGPT?, June 2026.
- The SEO Works, AI Referral Traffic 2026, 2026.
- David Karpik, Clickport, Why ChatGPT Traffic Shows as Direct in GA4 (2026), updated October 2026.
- Marco Di Cesare, Loamly, The AI Traffic Attribution Crisis, November 2025, updated February 2026.
- David Rosam, Everyone quotes 60, nobody measured it, September 2026.
- Kevin Indig, Growth Memo, Does AI drive more direct traffic?, October 2026.
- Matt G. Southern, Search Engine Journal, ChatGPT Traffic Drop Appears To Be A Tracking Problem, October 2026.
- Kaiser and Schulze, presented by Frankfurt School, Where ChatGPT already works in online shopping, June 2026, and by Digiday, E-commerce sites see low sales from ChatGPT traffic, October 2025.
- Amsive, Does LLM Traffic Convert Better Than Organic? A New Data-Backed Study, September 2025.
- Marie Haynes, GEO is a distraction, October 2026.