How do you measure AI search traffic, and why does most of it show up as Direct?

Most businesses that ask us about AI search have the same first question: how much of my traffic comes from it? The honest answer is that no report can tell you precisely, and the reasons are structural, not a settings problem. This article walks through what each tool shows, where each one undercounts, and how to build a number you can defend. If you have not yet checked whether AI names your business at all, start with the 30-minute audit and come back.

Why there is no single AI traffic number

AI search touches your site in three different ways, and each one is recorded by a different tool.^[2] A shopper who asks ChatGPT for a recommendation and clicks the link arrives as a referral from chatgpt.com. A shopper who reads an AI Overview on Google and clicks through arrives with a google.com referrer, indistinguishable from a classic result. A shopper who copies the link out of a chat app, or whose app strips the referrer, arrives with nothing, and lands in Direct.

So the traffic is real, but it is split across three buckets, two of which are labelled as something else. That is why a business can be gaining AI visibility while its reports look flat.

What the GA4 AI Assistant channel captures

Google added AI Assistant to GA4's default channel groups in its May 13, 2026 release.^[1] When a session arrives from a referrer on Google's list of recognised assistants, GA4 sets the channel to AI Assistant and the session medium to ai-assistant automatically. Google's current definition names ChatGPT, Gemini, Deepseek, Copilot and Grok as examples, and states that it excludes AI Overviews and AI Mode.^[2]

Four things keep this number low:

• Perplexity is not on the list. Its referrals fall into the generic Referral channel.^[4]
• Google's AI is excluded by design. AI Overview and AI Mode clicks stay inside Organic Search, where nobody is looking for them.^[2]
• It only classifies sessions that carry a clean referrer. Everything else is Direct.^[4]
• It is not retroactive. Classification started on May 13, 2026, so a year-over-year comparison from the default grouping is not possible until May 2027.^[4]

One more wrinkle: the list Google maintains lags reality. Target Internet broke down three months of its own assistant referrals and found the same assistants appearing under AI Assistant, under Referral and under a medium of "not set" at the same time.^[2] Check your Referral report for AI domains before you trust the channel total.

Fix it with a custom channel group

Build a custom channel group in GA4 that matches the AI referrers you care about, including Perplexity and Claude, and place its rule ahead of Organic Search and Referral.^[2] Unlike the default channel, a custom group applies to historical data, so you can look back. ChatGPT often appends utm_source=chatgpt.com to the links it shows; matching on that as well as the referrer catches more of it.^[4]

What Search Console's generative AI report shows

Because Google's AI features never reach the GA4 channel, you measure them on Google's side. Search Console launched a Generative AI performance report on June 3, 2026, and as of August 31, 2026 it is available to all websites worldwide.^[3] It covers impressions in AI Overviews and AI Mode on Search, with a separate report for Discover, broken out by page, country, device and date.^[5]

Two rules shape how you read it. First, it is impressions only: there is no click, click-through rate or position column.^[5] Second, impressions are aggregated by property, so if two pages from your site appear in the same AI answer, that counts as one impression in the chart.^[5] If the report does not appear for your property, Google's help page says the usual cause is too few impressions in generative AI features so far.^[5]

The Direct-traffic blind spot

The largest gap is the traffic that arrives with no referrer. Desktop and mobile chat apps in particular often send visitors without a referrer header, and GA4 has nothing to attribute the session to, so it lands in Direct.^[4] Published estimates of the referrer-less share range from about a third to two-thirds of AI traffic, and the width of that range is the point: nobody can measure it precisely, including Google.^[4]

You cannot recover a referrer that was never sent. You can watch for its signature. A rise in Direct sessions that land on deep pages, such as a specific product page or a specific guide, rather than the homepage, is unusual behaviour for a returning visitor and consistent with someone following a link an assistant gave them.^[2] Build a segment for Direct sessions whose landing page is not the homepage and track it as an indicator, not a measurement.

Why the number is worth getting right

Small as the floor looks, the traffic behind it behaves differently. Shopify's Q1 2026 commerce data found that AI-referred sessions on its platform converted at nearly 50% higher rates than organic search, with average order values 14% higher.^[6] In Q2 2026, AI-referred sessions to Shopify storefronts grew 197% year over year while organic grew 12% on a much larger base.^[7] Shopify's Q1 post put it plainly: organic still refers more sessions to its merchants than all tracked AI platforms combined.^[6] Fast growth from a small base: exactly the pattern a flat-looking report hides. We cover what those numbers do and don't mean for a smaller brand in a separate article.

Put it together: the measurement we run

This is the setup we use for clients. It takes a couple of hours once and then runs itself.

• GA4: the AI Assistant channel plus a custom channel group that also catches Perplexity, Claude and utm_source=chatgpt.com. Compare engagement and conversion rate for AI-referred sessions against organic. That comparison is usually a stronger argument than the volume.
• Direct indicator: a saved segment of Direct sessions landing on non-homepage pages, watched month over month.
• Search Console: the Generative AI report as its own trend line, checked monthly. Weekly numbers are noisy; the trend reads over quarters.
• Visibility tests: a locked set of ten to fifteen customer questions, run across ChatGPT, Perplexity, Gemini and AI Mode on a schedule, logging whether you are named, linked or absent. This is the only one of the four that measures the answers people see, not just the clicks that follow. The method is in the audit article.

Report the GA4 figure as a floor, the Direct segment as an indicator, the Search Console impressions as presence, and the visibility log as the headline. Four honest numbers beat one inflated one.