How to Track AI Search Traffic in Singapore: GA4’s AI Assistants Channel and Its Three Blind Spots

In May 2026, Google Analytics quietly did something it had refused to do for two years: it gave AI chatbot traffic its own channel. No configuration, no regex, no custom channel group. Traffic from ChatGPT, Gemini and a handful of others now lands in a row labelled AI Assistants, and every property owner got it for free.

This was widely reported as the moment AI traffic became measurable. It is not. It is the moment one narrow slice of AI traffic became measurable, and the reporting around it has left a lot of Singapore businesses with a number they trust more than they should.

The honest position is this: the new channel is genuinely useful, it is not what most people think it is, and the two largest categories of AI-influenced traffic are still invisible in it by design. This guide covers the AI-specific measurement problem only. For general GA4 configuration, channel groups and attribution windows, start with our GA4 setup guide for Singapore businesses and the walkthrough of attribution models; this post assumes that groundwork is already done and goes straight at the AI layer sitting on top of it.

First, separate three things people call “AI traffic”

Most measurement arguments dissolve once you notice that three completely different things are being discussed under one label. They arrive through different mechanisms, they show up in different places in your reports, and only one of them is genuinely new.

The new channel addresses the first row only. The second row is explicitly excluded, and the third row is not measurable by any analytics product, at any price, because the causal step happened in a conversation you were never party to. Keeping these three separate is most of the discipline.

What Google actually shipped in the AI Assistants channel

The rule is published in Google’s default channel group documentation, and it is worth reading literally rather than through a summary. A session is assigned to AI Assistants when “the medium exactly matches ‘ai-assistant'”, and Analytics sets that medium itself: “the medium is set to ‘ai-assistant’ and the campaign is set to ‘(ai-assistant)’ if the referrer matches a list of AI Assistants.”

Two consequences follow that are easy to miss.

First, the classification is driven by a list Google maintains, not by a pattern you control. The documented set covers ChatGPT, Gemini, Deepseek, Copilot and Grok. Perplexity — which by most published analyses sends unusually engaged traffic — is not on that list, and its sessions continue to land in Referral. If you sell to a technical or research-led audience, a meaningful share of your genuine assistant traffic is sitting outside the channel that claims to measure it.

Second, because Analytics is overwriting the medium, this is not a passive relabelling. Any custom channel group, regex or UTM convention you previously built around AI referrers now interacts with a medium value you did not set. If you built a custom AI channel in 2025 — most agencies did — check whether it now double-counts or silently drops sessions.

The third consequence is the one that catches people out at reporting time: the channel is forward-only. It began classifying sessions from mid-May 2026 onward and did not reclassify history. A year-on-year comparison that shows AI Assistants going from zero to a real number is not measuring growth in AI traffic. It is measuring the arrival of the channel. Say so in the report, or someone will build a forecast on it.

Blind spot one: AI Overviews and AI Mode count as Organic Search

This is the single most consequential fact in AI measurement, and it comes straight from Google’s own channel definitions rather than from anyone’s speculation. The Organic Search channel is documented as covering users arriving through non-ad links in organic search results, including Google’s AI Overviews and AI Mode. The AI Assistants channel is documented as excluding them.

That is a deliberate design choice, and once you see the logic it is hard to argue with. From Google’s point of view an AI Overview is a feature of the search results page, not a separate product. A click from it is a click from Google Search. Google’s own optimization guidance takes the same line, stating plainly that “from Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”

But for anyone trying to answer “how much of my traffic now comes from AI”, the effect is severe. In a market where StatCounter’s July 2026 data puts Google at 92.46% of Singapore searches against Bing’s 3.37%, the overwhelming majority of AI-mediated discovery is happening inside the one channel that will never break it out. Your AI Assistants row could read 0.4% of sessions while AI answers are involved in a far larger share of your Google traffic, and both statements would be true simultaneously.

There is no GA4 configuration that fixes this. Not a custom channel group, not a regex, not a paid connector — the distinguishing signal is not in the request. What exists instead is a separate, partial view in Search Console, covered further down. Anyone selling you a GA4 setup that “separates your AI Overview traffic” is either misunderstanding the mechanism or misrepresenting it.

Blind spot two: the missing referrer problem

The second gap is technical rather than definitional, and it pushes traffic in the opposite direction — out of AI channels and into Direct.

A visit is only classifiable if the browser sends a Referer header naming where it came from. Two common situations suppress it. Native desktop and mobile assistant apps frequently do not send one at all, and links inside AI answers are often rendered with referrer-suppressing attributes. In both cases GA4 receives a visit with no origin information, and its rule for that case is unambiguous: it becomes Direct.

Cloudflare, which measures this at network scale, states the limitation on its own crawl-to-refer metric explicitly: “traffic referred by Claude’s native app does not include a Referer: header, and we believe that the same holds true for traffic generated from other native apps as well.” Cloudflare goes on to say its own published ratios “may overstate the respective ratios, but it is unclear by how much”. When the company running a large share of the internet’s edge says it cannot size the gap precisely, treat any vendor claiming a precise figure for your site with suspicion.

Published estimates of how much AI traffic arrives referrer-less circulate widely — one analysis of roughly 446,000 visits put it around 70% — but these come from single-vendor panels with undisclosed site mixes, and the share depends heavily on whether your audience uses assistants in a browser or in an app. The direction is well established and the magnitude is not. Use it as a reason to expect undercounting, not as a correction factor to multiply by.

What you can do is bound the problem. Watch Direct as a share of total sessions over time, segmented to landing pages that are not your homepage. Genuinely direct traffic overwhelmingly hits the homepage or a bookmarked page; a rising trend of Direct sessions landing deep on an article is a reasonable proxy for referrer-less discovery, including from assistants. It is an indicator, not a measurement, and should be labelled as such in any report.

Blind spot three: the platform list moves and the definitions move with it

Because the channel depends on a Google-maintained list, its meaning changes without notice when the list changes. A platform added next quarter will shift sessions out of Referral and into AI Assistants, and your chart will show growth that is purely reclassification. This is the same class of problem as the forward-only launch, and it recurs.

The practical defence is to keep your own parallel view. Build a custom channel group, or at minimum an exploration, that captures assistant referrers by source pattern — including Perplexity and any others relevant to your market — and keep it running alongside the default. When the default channel jumps, you can tell whether the underlying traffic moved or only the label did. Annotate the property on the day you make any change; GA4 annotations exist for exactly this and almost nobody uses them.

A measurement setup that survives contact with reality

Here is the configuration we would actually put in place for a Singapore client, in order. None of it requires a paid tool.

1. Leave the default channel group alone. It is now the industry’s shared vocabulary. Report against it so your numbers are comparable to everyone else’s, and add your own views alongside rather than replacing it.

2. Build one custom channel group with a wider AI definition. Include the five documented platforms plus Perplexity and any others your audience uses. Name it something that cannot be confused with the default — “AI (extended definition)” rather than “AI Assistants”. Report both, and explain the delta when someone asks why two numbers differ.

3. Segment by landing page, not just by channel. The useful question is rarely “how much AI traffic do we have”. It is “which pages are assistants sending people to”. That is directly actionable: it tells you which content is being read and cited, which is the same signal our guide to AI brand visibility monitoring tracks from the other end.

4. Measure behaviour, not just volume. AI referral volume is small almost everywhere — published benchmarks across large multi-site samples in 2026 have generally placed it around or below the low single digits as a share of sessions. Volume alone will make it look ignorable. Engagement and conversion rate per session tell you whether the traffic is worth pursuing, and that is a conclusion you should reach from your own data rather than from a benchmark.

5. Set up conversion tracking properly before any of this matters. A channel report without conversions attached is a vanity exercise. If your key events are not yet configured, that is the prerequisite, and it is covered in our guide to the GA4 reports worth checking.

6. Annotate everything. Channel launches, list changes, your own configuration edits. In twelve months this is the difference between a defensible trend and an argument.

What Search Console gives you that GA4 cannot

The one official window into Google’s own AI surfaces is the generative AI performance report in Search Console, which began a limited rollout in June 2026 and which Google’s optimization guidance now points to directly as the way to “measure visibility in generative AI features”.

Its limits are as important as its existence. It reports impressions only — no clicks, no click-through rate, no position, no query data. It covers AI Overviews and AI Mode. Multiple results from the same site in one answer count as a single impression. And it is a breakout of data already inside the Web search type rather than a new pool of traffic, so you cannot add it to your organic numbers without double-counting.

Used properly it answers one question well: is Google’s AI showing my pages at all, and is that rising or falling? That is worth knowing. It does not answer what those impressions are worth, and until click data ships, nobody can tell you that from Google’s own instrumentation.

The number that actually belongs in a board report

If you take one structural idea from this guide, make it this: stop trying to produce a single “AI traffic” figure. It cannot be produced honestly, and the attempt is what leads people to quote a GA4 row that excludes the largest source of AI-mediated visits in their market.

Report three separate lines instead, each with its source named:

• Assistant referral sessions and their conversion rate — from GA4, small but precise, and the only AI number where you can see revenue.
• AI feature impressions — from Search Console, directional only, no clicks available, and only if the report has rolled out to your property.
• Total organic sessions and conversions — unchanged in definition, and still where the money is. This is the line that tells you whether AI answers are actually costing you anything, because if AI Overviews were quietly eating your traffic, this is where it would show.

That third line deserves emphasis, because it is the reality check on the entire category. The most-cited zero-click research of 2026 — SparkToro’s analysis of Similarweb clickstream data covering US Google searches from January to April 2026 — found that 68.01% of Google searches ended without a click. That sounds apocalyptic until you read the same study’s finding that just 0.34% of searches transitioned into AI Mode during that window. Zero-click search is a long-running trend that predates generative AI by years; AI Mode’s share of it was, at that point, close to a rounding error. Both facts come from the same dataset. Anyone quoting the first without the second is selling something.

Note also that this is US clickstream data, not Singapore. There is no published Singapore-specific study of AI search click behaviour that we have been able to verify, so treat every Singapore percentage you see quoted on this topic as an extrapolation from a foreign panel — including any you might be tempted to quote yourself.

What this means for a Singapore business

Three practical conclusions, given the market structure.

Your exposure is concentrated. With Google at roughly 92% of Singapore search, there is no meaningful second channel to hedge into. Whatever Google does to its results page happens to essentially all of your search demand at once. That argues for measuring the total organic line closely and continuously, not for chasing the AI Assistants row.

Small numbers can still be worth having. Assistant referrals are a small share of sessions for most Singapore sites. But the sessions are self-selected — someone read a synthesised answer, decided they wanted your source specifically, and clicked. Judge them on conversion rate per session against your other channels, using your own data. If they convert better, the volume argument is irrelevant; if they do not, no benchmark should persuade you otherwise.

The work that improves this is ordinary work. Google’s own guidance is explicit that structured data “isn’t required for generative AI search, and there’s no special schema.org markup you need to add”, and that eligibility means being indexed and eligible for a snippet. The practical implication is that there is no separate AI channel to buy your way into. What earns citations is content that is genuinely worth citing, on a site that can be crawled — which is why our AI SEO strategy guide spends most of its length on fundamentals, and why technical SEO basics matter more here, not less.

Where to go next

This post deliberately covers measurement only. For the surfaces themselves, read our explainer on what AI Overviews are and what actually changed, which carries the click-behaviour research in detail. For the practical side of earning citations, see how to show up in AI Overviews and how to get cited by ChatGPT. The cluster’s starting point is our AI SEO guide for Singapore. If you would rather have this configured and reported on properly than build it yourself, that is part of what our AI SEO service does, and you can see how we report on outcomes in our client case studies.

Frequently asked questions

Does GA4 show Google AI Mode as a referrer?

No. Google’s own channel documentation classifies clicks from AI Overviews and AI Mode as Organic Search, because they are clicks from a Google search results page. The AI Assistants channel explicitly excludes them. No GA4 configuration can separate them, because the distinguishing signal is not sent to your site. Search Console’s generative AI performance report is the only official view of those surfaces, and it reports impressions only.

Why is my AI Assistants channel showing almost no traffic?

Most likely because it is genuinely small, and because three things are excluded from it: Google’s own AI surfaces, platforms not on Google’s list such as Perplexity, and any visit that arrives without a referrer header — which includes much native-app traffic and lands in Direct instead. A near-zero AI Assistants row is normal and is not by itself evidence that AI search is irrelevant to you.

When did GA4 add the AI Assistants channel, and does it include old data?

Google confirmed the channel in May 2026 and it reached broad availability across properties the following month. It is forward-only: sessions before that point were not reclassified and remain in Referral. Any year-on-year chart of this channel will show growth that is mostly an artefact of the launch, so annotate it.

Can I estimate how much traffic I lose to AI Overviews?

Not from your own analytics, and be wary of tools claiming to calculate it. What you can do is watch total organic sessions and conversions for the queries and pages that matter to you, and use Search Console to see whether impressions in AI features are rising. Published click-behaviour research exists but is drawn from US panels with differing methods, so it indicates direction rather than giving you a number to apply to your site.

Should I pay for an AI visibility tracking tool?

They answer a different question from analytics: whether assistants mention your brand for prompts you choose, not how much traffic you get. That is genuinely useful for competitive benchmarking and useless as a traffic measure. Decide based on whether you will act on prompt-level visibility data. Our guide to AI brand visibility monitoring covers how to read the output, and how to evaluate AI SEO software covers the buying process.

Is Perplexity traffic included in the AI Assistants channel?

Not according to the documented platform list, which names ChatGPT, Gemini, Deepseek, Copilot and Grok. Perplexity sessions land in Referral. If Perplexity matters to your audience, build a custom channel group that includes it and report that alongside the default rather than instead of it.

The short version

GA4’s AI Assistants channel is a real improvement and a narrow one. It measures clicks from five named assistants, forward-only, and excludes the AI surfaces inside Google that account for most AI-mediated discovery in a market where Google holds over nine in ten searches. Search Console’s generative AI report covers those surfaces but reports impressions with no clicks. Neither sees visits that arrive without a referrer.

The response is not a better dashboard. It is a more honest one: three separate lines, each labelled with where it came from and what it cannot tell you, and a continued focus on the total organic number that would reveal a real problem if one existed. Everything else is estimation dressed as measurement.

If you want this set up, reported on monthly, and interpreted by someone who will tell you when a number is soft, that is what our AI SEO service is for. Get in touch and we will start with what your analytics is already telling you.