We manage more than 100 GA4 properties across our client base, we went and looked at what AI assistants were actually sending to each of them. The result we expected was that AI referrals convert well, they do. The result we didn’t expect was that on most of those sites, the standard analytics setup couldn’t see the traffic at all. If you’re making decisions about AI search from your default reports, you’re probably working from numbers that are missing your best visitors.
Content Overview
What we did, and how we counted it
For each property we pulled the full channel breakdown for 2026 so far, then isolated AI and LLM traffic two ways. First, the way GA4 reports it today through its own channel grouping. Second, by applying our own Regex that matches the actual source of the visit against the AI assistants people use, such as ChatGPT, Perplexity, Claude, Copilot, and Gemini.
Everything below is expressed as rates and multiples rather than raw figures, and grouped so no single client is identifiable. We’re comfortable sharing the pattern, we’re not going to publish any of our client’s traffic. Where we compare AI traffic to a site’s performance, we always compare each site to its own average, never one client to another, because a dental practice and a home builder convert at completely different levels for reasons that have nothing to do with AI.
The default reporting was hiding about half of it
Across the properties we measured, GA4’s built-in setup captured only about half the AI visitors our own rule found. Put the other way, the default was under-counting AI traffic by roughly two to one.
On a good number of sites it was worse than that. Some properties showed no AI channel whatsoever, so every visit an assistant sent was quietly filed under referral, organic, or unassigned, and nobody looking at the reports would know AI had sent anyone. The sites hit hardest were in construction and education, though it showed up across the board. Even where a default AI channel did exist, it under-counted, in some cases by as much as ten times on a single site, and commonly by three or four times.
The traffic doesn’t disappear when GA4 can’t label it. It just lands in a bucket you’re not watching, mixed in with everything else, so you can’t tell how it behaves or whether it’s worth anything.
The traffic it hid was converting the best
This traffic is worth a lot, AI-referred visitors out-converted the site average on roughly four in five of the properties we looked at.
On B2C sites, AI visitors converted at around 1.6 times the site average. On B2B sites, they converted at around 2.4 times the site average. B2C sites convert at higher rates overall, so in absolute terms their AI numbers are bigger. But the relative lift was stronger for B2B, which is the opposite of what you might assume. The longer and more considered the purchase, the more a recommendation from an AI seems to be worth by the time the person arrives. A drawn-out B2B decision is exactly the kind of question people hand to an assistant, and they turn up having already made up their mind.
By industry the lift was largest in manufacturing and industrial businesses, at around three times site average, followed by professional and financial services at roughly double. Education and home building sat near 1.7 times, and general retail and ecommerce around 1.6 times, though retail was more mixed than the rest.
The pattern is not universal. Hospitality, food, and impulse-buy retail barely moved, sitting close to their site average and dropping below it on several sites. The single property with the highest share of AI traffic in our whole client base is one where AI converts slightly below its own average, which is a useful reminder that volume and value are different things. And in one sector the blended figure looked flat until we noticed a single high-baseline site was dragging the average around. Once we looked site by site, AI had actually beaten the average on every property in that group. So the honest read is that AI referral traffic converts above average for most businesses, strongly for considered and B2B-leaning ones, and not much at all for low-consideration impulse purchases.
Why AI traffic converts higher than the rest
The discrepancy comes from three things happening at once:
- Who these people are. Someone who takes a buying question to ChatGPT or another LLM is already deep in the decision rather than idly browsing, so the traffic is self-selected for intent before it ever reaches you. It’s the same as comparing a first time user compared to someone who’s been on the site multiple times, of course the second is more likely.
- What the AI does with that intent. When a model reads across the options and hands back a short list with you on it, that recommendation carries weight, the same way a referral from someone you trust does. The visitor arrives with their guard already down, because something they rely on has vouched for you.
- Where they are in the buying process. They’ve handed the research, the comparison, and the shortlisting to the model, so by the time they land they aren’t weighing you up, they’re ready to act. Most channels deliver people who still have all of that ahead of them.
It’s about capitalising on buyer decision momentum. They want something, they now trust you’ve got it, they’ve just been told you’re the best or yours is the best for a bunch of reasons and they land on a site that (hopefully) leads them straight to it. If you can effectively keep that momentum going and that traffic converts better than anything else.
That is also why B2B pulls ahead of B2C. The purchases people research hardest are the considered, higher-stakes ones, a long B2B decision far more than a quick consumer buy. Those are exactly the questions that get taken to an AI and worked through in detail, so more of the buying happens before the click. The bigger the decision, the more vetting it gets, and the warmer the person is when they finally arrive. A B2C impulse buy skips most of that, which is likely why the lift is real there too but smaller (although you’d need 100 eCommerce clients specifically to test that theory, which we don’t have just yet).
How to make your AI traffic visible
You can fix this yourself in GA4, and it takes under a minute. In Admin, under Data display, open Channel groups and either create a new group or edit your custom one. Add a channel, name it something like AI Assistant / LLM, and set its condition to match the session source against the AI hostnames you care about, such as chatgpt.com, perplexity.ai, claude.ai, copilot, gemini, and the others your visitors use. Move that channel to the top of the list so it claims AI traffic before organic or referral can absorb it.
Here’s a regex you can paste in straight away to get started (you might need to add more later):
“chatgpt|openai|gemini\.google|bard\.google|copilot|perplexity|claude\.ai|you\.com|poe\.com|deepseek|mistral\.ai|grok|x\.ai|meta\.ai|phind|writesonic|chatsonic|character\.ai|huggingface”
Once that’s in place it applies across most GA4 reports, and you can finally see AI as its own channel: how many people it sends, what they do, and whether they convert. You can’t manage a channel you can’t see, and right now most businesses can’t see the most important one.
Who should set this up now
If you sell considered or higher-value services, or you’ve started any work to get recommended by AI, do this now. You’re almost certainly already receiving this traffic, and every week without proper tracking is a week of your highest-intent visitors landing in a bucket you’re ignoring.
If you’re a low-consideration or impulse-buy business, or your traffic is very low, the payoff is smaller, but there’s still a good reason to set it up. It costs almost nothing, and it gives you a baseline so you can see the trend forming instead of guessing at it later. Set it up, leave it running, and check back in a quarter.
Seeing it is only step one
Making the traffic visible tells you it’s arriving and that it converts well. It doesn’t do anything about the page it lands on, and a pre-sold visitor still bounces if that page makes them start deciding all over again. Once you can see this traffic, the next question is whether your site is built to convert it, which is a job in itself. We wrote about how to build a landing experience that converts a pre-sold AI visitor separately.
If you’d rather not work through your own analytics, we’re happy to take a look at what AI is doing on your site and whether it’s worth acting on.
That’s the kind of thing we set up as part of our generative engine optimisation work. Let’s talk.
Jack Headford
Account Manager & Operations Lead
