How to Actually Measure AI Search Traffic in 2026

For two years the honest answer to "how much traffic am I getting from AI" was that nobody could tell you properly, including the people charging to tell you. The data was not exposed. Everything else was inference dressed up as measurement.

That changed twice this year, about three weeks apart. Both changes are real improvements and both are partial in ways that are easy to miss and expensive to misread. This is what each one actually gives you, what it quietly leaves out, and what I would set up if you have an hour.

Change one: the GA4 AI Assistant channel

On 13 May 2026 Google added an AI Assistant channel to the default channel group in Google Analytics 4. It needs no configuration. When a session arrives with a referrer Google recognises as an AI assistant, the medium is set to ai-assistant, the campaign is labelled (ai-assistant), and the session lands in a channel called AI Assistant.

To find it: Reports, then Acquisition, then Traffic acquisition, and set the primary dimension to Session default channel group.

Now the part that matters. Here is Google's own definition, from the Analytics help documentation:

AI Assistants is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok. It excludes Google's AI Overviews and AI Mode.

Read the last sentence twice, because it undoes a lot of what people assume this feature does.

A click from an AI Overview or from AI Mode is a click from Google Search, and GA4 files it under Organic Search alongside every ordinary blue-link click. There is no dimension that separates them. So if your worry is Google's AI answers eating your traffic, which for most UK businesses is the bigger exposure by far, this channel does not measure it. It measures the assistants, which is a different and much smaller thing.

Three more limitations worth writing down.

It is not retroactive. GA4 processes data daily and does not reprocess history, so anything before mid-May stays where it was originally filed. Your AI Assistant channel starts the day it switches on, and any comparison to last year is not available.

The list is Google's, not yours. Google names ChatGPT, Gemini, Deepseek, Copilot and Grok as examples. Perplexity is not among them, and Perplexity is a meaningful share of AI referrals for some sites. If an assistant is not on Google's list, its traffic sits in Referral as before. Worth checking against your own Session source data rather than assuming.

The medium can be overwritten. If you have UTM tags on a link, or an existing channel group rule that catches these referrers first, the tagging wins and the session goes elsewhere. Custom channel groups you built as a workaround in 2025 are the most common culprit, so check whether yours is now double-counting or intercepting.

Change two: Search Console generative AI reports

Three weeks later, on 3 June 2026, Google announced Search Generative AI performance reports in Search Console. This is the one that covers what GA4 excludes: AI Overviews and AI Mode on Search, plus generative AI features in Discover.

What it gives you: impressions, the pages that appeared, countries, devices for Search, and dates with granularity down to hourly.

What it does not give you: clicks, click-through rate, or query data.

So it answers whether you were shown. It does not answer whether anyone came, or what they asked. That is a genuine limitation and it is worth being clear-eyed about, because an impressions-only report is very easy to over-read. Appearing 4,000 times in AI answers last month is not 4,000 anything. It is 4,000 opportunities you have no click data for.

It also rolled out to a subset of sites, starting in the UK. I have written up why the UK went first and what the CMA required of Google separately. If it has not reached your property yet, that is normal.

The dark traffic problem, honestly

Some proportion of visits that begin in an AI assistant arrive at your site with no referrer attached, and land in Direct. That happens for ordinary technical reasons: links opened from a desktop app rather than a browser, referrer policies, privacy settings, copy and paste.

You will see figures quoted for how big this is. A range of 35% to 70% circulates widely. I have not found a primary study behind it that I would rely on, and the sources quoting it tend to cite each other. So I am going to do the unhelpful thing and tell you the honest position: the effect is real, the size of it is not well established publicly, and anybody giving you a precise percentage is probably passing on somebody else's estimate.

What follows from that is a rule rather than a number. Whatever your AI referral figure says, it is a floor and not a total. Use it to watch a trend, never to calculate a share of your traffic.

What you can and cannot know

An hour of setup that is actually worth doing

1. Find the AI Assistant channel and screenshot it. Reports, Acquisition, Traffic acquisition, primary dimension Session default channel group. Whatever it says today is your baseline, because there is no history behind it.
2. Check your custom channel groups. If you built a regex rule for AI traffic in 2025, decide now whether to keep it or retire it. Running both without checking the order is how you end up with two numbers that disagree and no idea which is right.
3. Look at Session source directly. Set the primary dimension to Session source and search for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. This catches anything the default channel misses, and tells you whether Perplexity matters for you.
4. Turn on the Search Console generative AI report if you have it. Note the impressions figure and the pages list. If it is not available yet, diarise a check for next month.
5. Write down your three highest-impression pages. Those are the pages that carry your AI visibility. Everything you do next should start with them, not with the home page.
6. Add one question to your enquiry form. Where did you hear about us, free text, not a dropdown. It is unfashionable, the data is messy, and it is the only method that catches the customer who read an AI answer, thought about it for two days, then typed your name into Google. No analytics package will ever see that journey.

That last one is worth more than people expect. The whole measurement problem here is that AI discovery frequently ends in a branded search or a direct visit, which analytics attributes to Google or to Direct. A human being telling you they asked ChatGPT is better evidence than any channel report.

The small numbers problem

Here is the thing nobody selling AI visibility reporting wants to lead with. For most UK small businesses, these numbers are going to be tiny, and tiny numbers behave badly.

Some scale, from Semrush's analysis of 17 months of US clickstream data. ChatGPT referral traffic grew 206% year on year between January 2025 and January 2026. Total monthly referral visits plateaued at around a billion from November 2025. Those referrals were spread across roughly 170,000 unique domains by February 2026, and the top ten domains alone took just over 30% of the total, with Google itself taking 21.6%.

Divide what is left across the remaining domains and the average is a few thousand visits a month, and averages lie badly here because the distribution is so skewed. The median site is getting a small fraction of that. A local business in the UK, not in that panel at all, should expect numbers in the single or double digits per month.

Which matters for one specific reason: you cannot calculate a conversion rate from eleven sessions. One enquiry out of eleven is a 9% conversion rate, and it is also just one enquiry. Two would be 18%. Neither number means anything, and both will end up in a slide if somebody is not careful. This is the main reason the widely quoted claims about AI traffic converting several times better should be treated with care, which I have gone into properly in the piece on whether AI traffic really converts better.

What small numbers are good for is direction over months, not rates within a month. Watch the impressions trend across a quarter. Watch whether the pages appearing in AI answers are the pages you would want appearing. Ignore anything calculated on a denominator under about a hundred.

Questions people ask

Short sourced answers. The working is above.

How do I see ChatGPT traffic in Google Analytics?

Go to Reports, then Acquisition, then Traffic acquisition, and set the primary dimension to Session default channel group. Since 13 May 2026 there is an AI Assistant channel that captures traffic from assistants Google recognises, with no setup required. For anything that channel misses, switch the primary dimension to Session source and search for chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai directly.

Does GA4 track AI Overviews traffic?

Not separately. Google's own definition of the AI Assistant channel states that it excludes AI Overviews and AI Mode. A click from an AI Overview is a click from Google Search, so GA4 files it under Organic Search along with ordinary results, and there is no dimension that splits them apart. If you want visibility into AI Overviews and AI Mode, the Search Console generative AI report is the place to look, and it reports impressions rather than clicks.

What do the Search Console generative AI reports include?

Impressions inside generative AI features on Search, which covers AI Overviews and AI Mode, plus generative AI features in Discover. You get impressions, the pages that appeared, countries, devices for Search, and dates down to hourly granularity. You do not currently get clicks, click-through rate, or query data. The reports launched on 3 June 2026 and rolled out to a subset of sites, beginning in the UK.

Why is my AI traffic showing as Direct?

Because a share of visits that start inside an AI assistant arrive with no referrer information attached, usually because the link was opened from a desktop application, or because of referrer policy and privacy settings. Analytics has nothing to attribute the visit to, so it files it as Direct. The effect is real, the size of it is not well established in public research, and the figures circulating tend to cite each other rather than a primary study. Treat your AI referral number as a floor rather than a total.

Is the GA4 AI Assistant channel backdated?

No. GA4 processes data daily and does not reprocess history, so sessions from before the channel switched on stay in whichever channel they were originally assigned. There is no year-on-year comparison available and there will not be one until the channel has been running for a year. That is a good reason to record a baseline now rather than later.

How much AI traffic should a small UK business expect?

Very little, in absolute terms. Semrush's clickstream analysis put total ChatGPT referral visits at around a billion a month from November 2025, spread across roughly 170,000 domains, with the top ten domains taking just over 30% of that and Google itself taking 21.6%. The distribution is heavily skewed toward large sites. A local UK business should expect single or double digit sessions per month at this stage, which is enough to watch as a trend and not enough to calculate a conversion rate from.

Do I need a paid AI visibility tool?

Not to answer the basic questions. GA4 and Search Console now cover whether you were shown and whether anyone arrived, for free. What paid tools add is prompt-level monitoring across several assistants, which is genuinely useful at scale and genuinely noisy at small scale, since answers vary substantially between repeat runs of the same prompt. If you are considering one, ask how many prompts it runs, how often, on which engines, and whether it repeats them, before you ask what it costs.

Sources

Every figure comes from the linked source. None of it describes a Stagg Studios client. The clickstream figures are from a US panel, so the absolute numbers will not map onto a UK business directly, and I have used them for scale rather than as a benchmark.

• Google, Default channel group definitions, Google Analytics Help (2026), for the AI Assistants channel definition
• Search Engine Land, Google Analytics adds AI Assistant channel to measure AI traffic (2026)
• Google Search Central, Introducing Search Generative AI performance reports in Search Console (June 2026)
• Search Engine Land, Google Search Console AI performance reports and controls to block your content in AI responses (2026)
• Semrush, ChatGPT traffic analysis: insights from 17 months of clickstream data (2026)

What I would do about it, and what it costs

Here is the part worth sitting with. Both of these reports start from the day they are switched on. Neither is backdated. So the difference between setting this up in August and setting it up next spring is not a few hours of work, it is whether you have a year of history when you actually need it, or nothing to compare against. That is the entire argument for doing it now, and it is a good one.

The setup itself is not hard and I have written it out above so you can do it yourself. The part that people do not keep up is the monthly reading: checking that the channel is still catching what it should, that the impression trend is going somewhere, that the pages appearing in AI answers are the pages you would choose, and that nothing on your site has quietly started blocking the crawlers that feed those answers.

That is what I do. The technical side of being found, handled as one job, plus a monthly read of what the numbers actually say rather than a dashboard nobody opens. It gives you a chance of ranking better and makes it more likely that what Google and the assistants read about you is accurate. I cannot promise you a ranking, a citation or a customer, and I would be wary of anyone who does.

£100 a month, first month free, no contract, and you deal with me. Start here, or send me a note and I will tell you what your current setup is missing.