AI Traffic Analytics: Set It Up in GA4 and Know What It Misses

AI traffic analytics is measuring the visits that ChatGPT, Perplexity, Gemini, Claude, Copilot and other AI assistants send to your site, then what those visitors do. In GA4 you get it from the built-in AI Assistant channel plus a custom channel for the sources it misses. Whatever number you get is a floor: visits that arrive without a referrer land in Direct, and clicks from Google’s AI Overviews stay inside Organic Search.

This guide covers the setup, the blind spots, how to separate bots from people, a worked example that turns AI visits into value per landing page, and a table for debugging numbers that look wrong. If what you want is whether assistants mention your brand, that is a different job, covered in our guide to LLM tracking.

What AI traffic analytics measures (and what it doesn’t)

Three different things get called “AI analytics” in marketing. They come from different tools and answer different questions, so keep them apart from the start.

This article is about the first two. Most of the ranking guides on this topic agree on that split, and it matters most when you report: add a visibility number to a traffic number and nobody can tell what the total means.

GA4’s built-in AI Assistant channel

GA4’s default channel group now includes a channel called AI Assistant. Google describes it as the channel for users arriving from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, and says it excludes Google’s AI Overviews and AI Mode. The detail most guides skip is how it works: when the referrer matches Google’s list of AI assistants, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and the channel rule is simply “medium exactly matches ai-assistant”.

Three consequences follow:

• Any filter on medium = referral now misses these sessions. Old explorations and Looker Studio reports that isolate AI sources with “medium is referral” will undercount. One top-ranking guide even recommends adding a medium condition such as ^(referral|organic|cpc|organic_search)$ to the AI channel; with the new medium, that condition excludes the sessions you want.
• Google does not publish the full list. Perplexity and Claude are not among the named examples. Check your own source report: if perplexity.ai or claude.ai still shows up under Referral, you need a custom channel for them.
• It only works when a signal exists. Like any channel rule, it classifies visits that carry a referrer. Visits without one go to Direct, as below.

How to set up AI traffic analytics in GA4

The built-in channel is a good start. The custom setup below catches the sources it skips, keeps one row for “AI” in every standard report, and takes about as long as it takes to read. You need Editor or Administrator access to the property.

Step 1: Check whether AI assistants already send you visits

Open Reports → Acquisition → Traffic acquisition, switch the primary dimension to Session source / medium, set the date range to the last 90 days and search for chatgpt, perplexity, gemini, copilot and claude one at a time. Note every AI source you find and its medium (referral or ai-assistant). That list is the input for step 2, and it tells you which sources are already handled by the built-in channel.

Step 2: Write and test one source pattern

Channel conditions use regular expressions. Here is a pattern anchored to whole domains, so it matches chatgpt.com and www.perplexity.ai but not a site that merely contains those letters. It also catches the bare chatgpt or perplexity source some links carry in their UTM tags. Add or remove domains to match what step 1 showed you.

Test before you save. Paste the pattern into any regex tester and check it against a list that includes real AI sources and lookalikes:

Step 3: Build the channel and put it above Referral

• Go to Admin → Data display → Channel groups and copy the default group to create a new one (for example “Channels with AI”).
• Add a channel named AI traffic with two conditions joined by OR: Source matches regex (the pattern from step 2), and Medium exactly matches ai-assistant. The second condition keeps everything the built-in channel already recognizes.
• Drag the new channel above Referral, and above the copied AI Assistant channel if your copy contains one. Google’s documentation is explicit that traffic goes to the first channel whose definition it matches, in the current order. Leave it below Referral and it stays empty.

Standard properties can have two custom channel groups beyond the predefined one, with up to 50 channels each, so do not burn both slots on experiments. You can also make your custom group the primary channel group, which changes how the property classifies traffic in reports going forward.

Step 4: Confirm the result in standard reports

Back in Traffic acquisition, pick your new group from the channel dropdown. You should see an AI traffic row, and chatgpt.com should no longer appear when you drill into Referral. Google says custom channel groups can be applied to your reports retroactively, so you should see past months too.

Step 5: Build an exploration by source and landing page

In Explore, start a blank exploration. Rows: Session source / medium and Landing page + query string. Values: sessions, engaged sessions, engagement rate, key events and, if you sell online, total revenue. Filter on your custom channel group exactly matching AI traffic. Add a line chart by week to see the trend. This is the view you will use most: which assistant sends people to which page, and whether they do anything there.

Step 6: Add conversions and value

Sessions do not tell you whether AI traffic matters. Mark the actions that do as key events (purchase, demo request, signup, form submit, phone click) and give each one a value, then compare value per session across channels. Our guide to GA4 conversion tracking covers the key event setup, and how to calculate conversion value covers putting a dollar figure on a lead.

Why some AI traffic is invisible to analytics

Analytics tools classify a visit from two signals present when the session starts: the referrer and any UTM tags on the URL. When both are missing, the visit is Direct. Every guide on this topic agrees on the main ways that happens:

• Apps that open links without a referrer. Some assistant apps and in-app browsers do not pass one. The top guides disagree on exactly which ChatGPT surfaces strip it (see below).
• Copy and paste. Someone copies your URL out of an answer, or types your brand name after reading it. Nothing links the visit to the assistant.
• Google’s AI Overviews and AI Mode. A click from an AI Overview arrives from Google like any other organic result, with no parameter to tell it apart. GA4’s AI Assistant channel explicitly excludes these, and the same limit applies in Adobe Analytics and every other tool. Our explainer on what an AI Overview is covers how those answers work.
• Sources not on any list yet. A new or regional assistant lands in Referral until someone adds it to a pattern.

One signal survives a stripped referrer: a UTM tag the assistant itself adds. Links cited in ChatGPT often carry utm_source=chatgpt.com, which is why it can appear as a source even when the referrer is empty. Check for it in your source report. You cannot add tags to links an assistant writes, so advice to “watermark” your content with UTMs does not apply to organic citations; tags only help on links you control, such as ads or listings.

Search Console’s AI report shows impressions, not traffic

Search Console now has a generative AI performance report covering AI Overviews and AI Mode. It counts an impression when a link to your site was shown in one of those features, and breaks it down by page, country, device and date. Google says the rollout reached all sites as of August 31, 2026.

It has no clicks, so it cannot tell you how much traffic AI Overviews send. Use it as the visibility layer for Google’s own AI features: if a page’s AI impressions rise while its organic clicks fall, that is a pattern worth investigating, not proof of cause. The help page offers an export button and does not mention API access, so plan for manual exports if you want this in a dashboard.

Bots are not visitors: crawlers vs AI-referred people

AI companies also hit your server with automated agents, and these show up in server logs and CDN dashboards, not as people. OpenAI, for example, documents several agents: GPTBot (used for its foundation models), OAI-SearchBot (used to surface websites in ChatGPT’s search features) and ChatGPT-User (used when a user’s question in ChatGPT makes it visit a page). Anthropic and Perplexity publish their own crawler names.

A ChatGPT-User fetch means an assistant read your page while answering someone. It is not that person visiting. Keep three counts separate: crawler hits (logs, by user agent), assistant fetches (logs, by user agent) and human sessions (analytics, by referrer). A page with many fetches and few sessions is being read and summarized but rarely clicked, which is a content question, not a tracking bug.

Estimating dark AI traffic, with a worked example

You cannot measure visits that carry no source, but you can bound them. The usual signal, mentioned by several of the ranking pages, is Direct sessions landing on deep pages: real people rarely type the URL of a long how-to article from memory. Turn that into a number like this.

All figures below are illustrative, for a made-up B2B software site with about 40,000 sessions a month.

• Recognized AI traffic (your custom channel) last month: 600 sessions. This is the floor.
• Direct sessions landing on content pages, excluding the homepage, login and app URLs: 2,500 last month, against an average of 1,800 for the same months a year earlier, before AI referrals showed up in your reports.
• Excess Direct on deep pages: 2,500 − 1,800 = 700 sessions.

The excess is an upper bound, not an AI count: bookmarks, internal links in emails, and seasonal demand also move Direct. Report it as a range with the word “estimate” attached, and never add it to the measured 600.

Now the more useful question: is the measured 600 worth anything? Same illustrative site, by landing page:

Traffic alone ranks the setup guide first. Value ranks the comparison page first by a wide margin. That is the decision this analysis exists to support: which pages to update, protect or build more of. The per-conversion values here come from a lead value calculation; our pillar on AI conversion tracking walks through that full chain from AI visit to revenue.

What to analyze once the channel exists

• Share of total traffic. AI sessions divided by all sessions, by month. Published benchmarks vary widely between studies, so your own trend line is more useful than anyone’s average.
• Source mix. Which assistants send visits, and whether the mix shifts. Sources are volatile: a page can be cited one week and dropped the next, so look at multi-week trends, not single days.
• Landing pages. Which pages assistants send people to. This is your best clue to what people asked, since no assistant passes the prompt.
• Engagement and conversion versus other channels. Engagement rate, key events, value per session. Small numbers swing hard, so wait for a few dozen conversions before declaring AI traffic better or worse than organic.
• Audience. Country, device and browser, which tell you where assistant users are coming from and whether the mobile experience of your top AI landing pages needs work.
• Branded search. A rise in branded queries in Search Console can be an indirect effect of AI mentions. Treat it as a supporting signal, not attribution.

When AI traffic numbers look wrong: debugging

Where the top guides disagree

Do custom channel groups work on past data?

One top-ranking guide says channel groups are not retroactive in standard reports; another says they are. Google’s custom channel groups documentation settles it: they “can be applied to your reports retroactively.” If past months look empty, check the channel order before assuming a data limit.

Which ChatGPT users send a referrer?

One guide says free ChatGPT users never send a referrer and paid users do; another blames the mobile app. Neither cites a primary source, and we found no OpenAI documentation of referrer behavior by plan. Treat both as unverified. What you can check is your own data: look at the share of ChatGPT sessions that arrive with a referrer versus only a utm_source tag.

How much AI traffic is hidden?

The ranking pages quote vendor studies whose estimates of untracked AI traffic differ by roughly a factor of two, measured on different samples. There is no neutral benchmark. Your own Direct-on-deep-pages estimate, labeled as an estimate, is more defensible in a meeting than a borrowed percentage.

Tools: which one answers which question

Every tool that reads referrers hits the same wall, so pick on what happens after the visit is recognized. If the question you need answered is conversions and value by landing page rather than sessions, SEOConversion identifies visits from ChatGPT, Perplexity, Claude, Gemini and Copilot from referrers and supported UTM signals and reports their conversions and value per landing page. When an assistant sends no referrer, the visit stays Direct; it does not guess.

How to report AI traffic without overclaiming

• Report recognized AI visits as “at least”, with conversions and value next to them.
• Report dark AI as a labeled range, never added to the measured number.
• Report visibility (prompt tracking, Search Console AI impressions) on its own line, because it is not traffic.
• Show the trend over several months; one month of a small channel is noise.
• Update your source pattern every quarter and note the change date, so a jump caused by a new domain is not read as growth.

If you meant a different kind of AI traffic analytics

The phrase has two other common meanings. In transportation, it refers to cameras, sensors and AI models that count vehicles and adjust traffic signals, which is where questions like “is AI in traffic lights?” come from. In IT security, network traffic analysis uses machine learning on network data to spot bottlenecks and threats. Neither is about website visitors.

And if you are asking what website analytics is in general: it is the collection and analysis of how people find and use a site, from sources and pages to conversions. AI traffic analytics is one slice of that, the visits that start in an AI assistant. For the organic search slice, see our pillar on SEO conversion tracking.

Frequently asked questions