To track AI traffic in GA4 properly you need two figures on the same slide: the share of sessions, which sits around or below 1% for most B2B sites, and the conversion rate, which looks higher for B2B and falls apart under testing almost everywhere else. GA4 added a native AI Assistant channel on 13 May 2026, and it fixes less than the coverage suggested: Google doesn’t name Perplexity in it, says nothing about past traffic, and can’t see sessions that arrive with no referrer or source tag, which land in Direct. A custom channel group closes most of that gap, and you should always report the share as a floor.
Most guides on how to track AI traffic in GA4 stop at the setup, and the setup is the easy half. GA4 now has a native AI Assistant channel, a custom channel group still catches more and applies to your history, and both miss the AI visits that turn up with no source attached, which is why every AI traffic figure you report is a floor. The harder half is reporting it. The conversion multiples everyone quotes (23x, 4.4x) come from vendors measuring their own sites and categories, and the one multi-site study that tested significance found no overall difference between AI and organic traffic, while its B2B slice showed AI converting at close to double. So the slide that survives a leadership meeting shows the share, the conversion rate and the raw conversion count side by side, and attaches a decision to them.
Where does AI traffic land in GA4?
AI traffic in GA4 lands in one of four channels, and only one of them is labelled as AI. On 13 May 2026 Google added a “New AI Assistant traffic measurement” to Analytics, naming “chatbots like ChatGPT, Gemini, and Claude” and assigning the medium ai-assistant “when the referrer matches a recognized AI Assistant” (Google Analytics). Before that, AI sessions went to Referral or, with no referrer, to Direct.
| Where it lands | What sends it there | What it misses |
|---|---|---|
| AI Assistant (native, since 13 May 2026) | A referrer on Google’s recognised list. | Sources Google hasn’t listed, and Google hasn’t published the full list. |
| Referral | An AI referrer that isn’t on Google’s list. | Nothing labels it as AI unless you build a channel for it. |
| Direct | No referrer and no source tag at all. | The biggest leak, and the hardest to size. |
| Organic Search | Clicks from Google AI Overviews and AI Mode. | Indistinguishable from an ordinary search click. |
Google’s channel definitions give ChatGPT, Gemini, DeepSeek, Copilot and Grok as examples of recognised sources, and file AI Overviews and AI Mode under Organic Search (Google Analytics). Perplexity isn’t named in either the release note or the definition, so check where your Perplexity sessions sit before you put the channel in a board deck. Google’s note also says nothing about past traffic, and since the channel keys on a medium GA4 assigns as sessions arrive, don’t expect your pre-May history to move into it.
Sessions only show up at all if AI crawlers and agents can reach your site, which is worth checking if you’re on Cloudflare, since its AI crawler defaults changed in September.
Why does AI traffic end up in Direct?
AI traffic ends up in Direct when a visit arrives with neither a referrer nor a source tag, because GA4 has nothing to attribute it to. How big that leak is depends on the tool, the app and the week. Analytics vendor Clickport found that 35.7% of the AI sessions it could identify across its customers’ sites carried a source tag but no referrer, so the tag alone kept them out of Direct; its own sample is self-selected, which it says plainly (Clickport).
October showed how fast that can change. ChatGPT links stopped carrying their utm_source=chatgpt.com tag and opened in a way that suppressed the referrer, and OpenAI confirmed it was a bug, with a fix rolling out on 15 October. Adobe’s Cedric Huesler looked at about 500 pages where ChatGPT had been at least 5% of visits, comparing 9 to 13 October with the week before: ChatGPT traffic fell by about 31,000 visits a day while Direct rose by about 26,000 on the same pages (Search Engine Land). If you report on AI traffic, annotate 8 to 15 October 2026 in GA4 and don’t read that week as a trend.
How do you track AI traffic in GA4 with a custom channel group?
You track AI traffic in GA4 most completely with a custom channel group on top of the native channel, because Google says custom groups “can be applied to your reports retroactively” and you control which sources they cover (Google Analytics). It takes about ten minutes in Admin, and standard properties get two custom groups besides the default, so don’t waste one.
Start in Admin, then Data display, then Channel groups, then Create new channel group, and put the year in the name, because you’ll be editing the regex. Add a channel called AI Search with the condition Source matches regex, and paste a hostname-anchored pattern, since an unanchored perplexity matches any domain containing the string:
^(chatgpt\.com|chat\.openai\.com|claude\.ai|perplexity\.ai|www\.perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|deepseek\.com|grok\.com|meta\.ai|you\.com|poe\.com)$
Matching on Source catches both the referrer and a utm_source tag, so chatgpt.com from either lands in the channel.
Then drag the AI channel above Referral, because GA4 puts traffic in “the first channel whose definition it matches”, so a freshly built group sits empty while the traffic keeps showing up under Referral if you skip this step; Google’s own example places AI assistants above Referrals for the same reason.
Last, put a quarterly review in the calendar: new assistants appear and existing ones change how they pass referrers, so a regex written in August is stale by spring, and Google’s help page says as much.
Write down what this setup doesn’t catch, and keep it next to the report so nobody discovers it in a meeting: visits with no referrer and no tag, Google AI Overview clicks, and any assistant released since you last edited the pattern. The native channel misses those too.
Does AI traffic convert better than organic?
For B2B, probably yes, though nowhere near the multiples in circulation. Ahrefs’ Patrick Stox found AI search visitors converted at 23x the rate of traditional organic on Ahrefs’ own site over 30 days, with AI search bringing 0.5% of traffic and 12.1% of signups (Ahrefs). Semrush’s 2025 AI search study put the average AI visitor at 4.4x the value of an organic one “based on conversion rate”, across 500+ digital marketing and SEO topics, without publishing the rates behind it (Semrush).
Neither works as the general claim it gets used for: Ahrefs is one site selling SEO tools to people who ask AI assistants which SEO tool to buy, close to the best possible case, and Semrush studied topics inside its own category.
Amsive ran the test: Will Guevara analysed six months of GA4 data across 54 websites with validated conversions and found organic converting at 4.60% and LLM referrals at 4.87%, with a paired t-test showing no significant difference (p = 0.794). His conclusion: “LLM traffic did not convert significantly differently from Organic.” LLM traffic beat the site average on 56% of sites and fell below it on 41% (Amsive).
| Source | Sample | Finding | The catch |
|---|---|---|---|
| Ahrefs, June 2025 | Own site, 30 days. | 0.5% of traffic, 12.1% of signups (23x). | One vendor site with an ideal category fit. |
| Semrush, 2025 | 500+ digital marketing and SEO topics. | AI visitors worth 4.4x organic. | Own category, vendor-published, rates not shown. |
| Amsive, September 2025 | 54 sites, 6 months, paired t-test. | No significant difference overall (p = 0.794). | The only one of the three that tested significance. |
| Amsive, B2B subset | Within the same 54. | 2.17% LLM against 1.16% organic. | A smaller subset, but the segment that counts here. |
The B2B row is the one to hold onto: the overall premium disappears under testing, while the B2B slice still shows LLM traffic converting at close to double organic.
Why can’t your own site answer the conversion question yet?
Your own site can’t answer it because the sample is too small. At 1% of sessions on a site doing 10,000 a month, that’s 100 AI sessions, and at a 3% conversion rate you’d expect three conversions, and next month it could be one or six, neither of which tells you anything about the channel.
My rough rule: a segment needs about 30 conversions before it’s worth looking at, and around 100 before its conversion rate stops swinging on noise. Most B2B marketing sites will reach that somewhere between next year and never.
The fix is free: report the raw conversion count next to the rate every time and let the denominator do the arguing. It’s the same discipline I used when reporting pipeline contribution, which I wrote up in the 48% marketing-sourced pipeline figure and how it was produced.
How should you present AI traffic to leadership?
Once you track AI traffic in GA4, present it as one slide with four lines, in this order, and never lead with the multiplier.
- The share, stated as a floor. “AI referrals were 1.1% of sessions in July, and that’s a starting point, because AI visits with no referrer or tag land in Direct.”
- The counts. Sessions and conversions as whole numbers: “112 sessions, 4 conversions.” Leadership reads counts faster than rates, and counts can’t flatter a small sample.
- The comparison, with the right caveat: “Converting above site average on a sample too small to test; the only multi-site study with a significance test found no overall difference, though its B2B subset showed nearly double.” That sentence separates a content lead who read a headline from one who read the study.
- The decision, which can be that nothing changes. Usually that means keep publishing citable material, don’t move budget on a 1% channel, and revisit at 30 conversions.
When I tracked 17+ KPIs a quarter across content, social and organic at a B2B SaaS company serving 200+ higher education institutions, the metrics that survived contact with leadership were the ones I’d stress-tested myself first. A figure that collapses the moment someone asks about sample size costs you the rest of the deck. There’s more on which metrics earn a place in the metrics that tell you if your content strategy is working, and on what to say about the AI visibility you can’t prove in measuring AI visibility without analytics.
What should you track while the AI traffic is too small to read?
Track the signals that move before sessions do, so you have a baseline by the time GA4 has enough to test. Citations show up long before referrals, so a monthly prompt log across ChatGPT, Claude, Perplexity and Gemini gives you a trend while GA4 is still on single digits (how I track whether AI is citing my content), and the prompts themselves come out of prompt research.
Search Console works the same way, since impressions move before clicks do (impressions up, clicks down). Pair citation presence with the GA4 channel and you have a story with a direction long before you have one with statistical power, which is also what to report when organic sessions fall and nothing is wrong. The content work doesn’t change either way: the same substance-first approach as the AEO/LLMO content optimisation framework, and the same reason AEO won’t save a weak content strategy.
Where does this advice run out?
The 1% figure is a B2B starting point, and your property can sit well above or below it. Rand Fishkin noted in June 2026 that AI tools “send less than 1% of all traffic out” (SparkToro), and the Datos Q1 2026 State of Search report put AI tools at 1.65% of US desktop events and 1.72% in the EU and UK (PPC Land on Datos). A research-heavy vertical will sit higher than a B2B SaaS site, so check your own property before quoting anyone’s percentage, including mine.
I also can’t tell you what share of your Direct traffic is AI, and neither can anyone else. Clickport’s 35.7% describes AI sessions it already recognised, and the October bug showed the leak can jump in a week, so any single estimate would be false precision.
The significance argument cuts both ways. Amsive’s study is the most rigorous available, and it’s still 54 sites over six months in 2025, before the native channel existed, in a channel that grew fast during and after the measurement window. A finding of no significant difference then doesn’t prove there’s no difference now, but it does show the 23x headline was never supported by the evidence people cite it with, which is why the slide needs the share and the conversion rate side by side instead of whichever one flatters the argument you walked in with.
Frequently asked questions
Use two layers. GA4 has had a native AI Assistant channel since 13 May 2026 that needs no setup and labels sessions whose referrer matches Google’s recognised list, with ChatGPT, Gemini and Claude among the examples Google names. On top of that, build a custom channel group under Admin, Data display, Channel groups, with a hostname-anchored regex on Source covering ChatGPT, Claude, Perplexity, Gemini, Copilot and the smaller assistants, then drag it above Referral so GA4 matches it first. The custom group covers sources Google hasn’t listed and applies to your past data in reports.
AI traffic shows up as Direct when a visit arrives with no referrer and no source tag, because GA4 has no source to assign. That happens with some apps and link settings, and in October 2026 it happened at scale when a ChatGPT bug stripped the tag and suppressed the referrer for about a week, moving tens of thousands of daily visits into Direct on the pages Adobe analysed. Clicks from Google’s AI Overviews and AI Mode are a separate leak, counted as Organic Search. Either way, any AI traffic figure you report is a floor.
For B2B, probably, though not by the multiples in circulation. The widely quoted 23x came from Ahrefs’ own site and the 4.4x from a Semrush study of its own category, both vendor-published and untested for significance. Amsive’s study of 54 sites over six months found no statistically significant overall difference (p = 0.794), with LLM traffic beating the site average on 56% of sites and falling below it on 41%. Its B2B subset showed LLM traffic converting at 2.17% against 1.16% for organic, which is the claim worth making.
Report the share and the conversion rate together, put the raw conversion count next to the rate so the sample size is visible, and state the share as a floor. If the segment has fewer than about 30 conversions, give the count and skip the rate, because a percentage off single digits swings from month to month and gets challenged the first time it moves. At this volume the point of the report is the baseline, so attach a decision anyway, even if the decision is to change nothing and revisit at a set threshold.
