People are finding businesses through ChatGPT, Perplexity and Google’s AI Overviews, and most analytics setups are quietly filing that traffic as “direct” alongside people who typed the URL from memory. The result is a channel that is genuinely growing and completely invisible in reporting. This guide covers why AI referrals get misattributed, how to separate them in GA4, and what the numbers can and cannot tell you once you do.
Key Takeaways
- Most AI-search traffic lands in GA4 as direct, because chat interfaces frequently strip or never send a referrer.
- Some AI platforms do pass a referrer. Those can be isolated in GA4 today with a custom segment or exploration.
- Traffic cited inside a Google AI Overview usually reports as ordinary organic search, not as a separate channel.
- A sudden rise in direct traffic with no offline campaign behind it is the clearest available signal that AI referrals are growing.
- AI referral volume tends to be low and convert well, because the person arrives pre-qualified by the assistant’s answer.
- No measurement setup is complete yet. Treat what you can capture as directional, and be honest with clients about the gaps.
Why AI Traffic Disappears Into “Direct”
Analytics assigns a source using the referrer, the piece of information a browser passes saying where a visitor came from. When someone clicks a Google result, the browser passes google.com and GA4 files it as organic search.
Chat interfaces break that chain in several ways. Some strip the referrer entirely. Some open links in a way that loses it. Some users copy the URL out of a chat and paste it into a new tab, which produces no referrer at all. Where the referrer is missing, GA4 has one place to put the session, and that is direct.
So “direct” in most accounts is now a mixture of people who genuinely typed the address, people arriving from apps and email clients, and a growing group who were recommended by an AI assistant. Those are very different behaviours sitting in one bucket.
The Two Kinds of AI Traffic, and Why Only One Is Visible
This distinction matters and is often missed.
AI referral traffic is someone using ChatGPT or Perplexity, receiving an answer that cites your site, and clicking through. This is a genuine visit from a distinct source, and it is sometimes trackable.
AI Overview traffic is someone searching Google normally, seeing an AI-generated summary that cites your page, and clicking that citation. That click still comes from a Google search results page, so it reports as ordinary organic search. It does not separate out.
The practical consequence: you cannot currently isolate AI Overview clicks from other organic clicks in GA4. What you can do is watch for the pattern in Search Console, where impressions rise while clicks flatten. That divergence is the fingerprint of AI Overviews answering the question without the user needing to click. Our post on how to appear in Google AI Overviews covers the visibility side.
Separating AI Referrals in GA4
Some platforms do pass a referrer, and those can be isolated today.
Build an exploration
In GA4, open Explore and create a free-form exploration. Add Session source / medium as a dimension and Sessions, Engaged sessions and Conversions as metrics. Then apply a filter on session source containing any of the AI platform domains.
The domains worth including as a starting point: chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com. This list changes as platforms launch and rename, so it needs occasional review rather than being set once.
Create a reusable segment
Rather than rebuilding the filter each time, create a segment on the same condition and save it. It can then be applied across reports and compared with organic and paid over time.
Mark them as a channel group
In Admin > Data display > Channel groups, create a custom channel group with a rule matching those same domains. AI referrals then appear as their own line in acquisition reports rather than being scattered across “referral”.
This is the single highest-value change, because it makes the channel visible in the reports people actually look at rather than only in an exploration someone has to remember to open.
Direct traffic climbing with no explanation?
That is usually the first sign of AI referrals. Separating them takes an afternoon and changes what your reporting is worth.
Reading the Direct Traffic Signal
Since a large share of AI traffic will remain in direct regardless of setup, direct itself becomes a signal worth reading rather than ignoring.
What suggests AI referrals rather than genuine direct visits:
- Direct sessions rising with no offline campaign, no PR and no brand activity to explain it
- Direct traffic landing on deep pages rather than the homepage. Very few people type a long article URL from memory
- New users making up an unusually high share of direct traffic, which contradicts what direct normally means
- Engagement patterns closer to organic search than to returning visitors
None of these is proof. Together they are a reasonable inference, and considerably better than treating direct as a black box.
What Search Console Adds
Search Console does not label AI Overview impressions separately, but the pattern shows up.
Watch for queries where impressions rise while clicks stay flat or fall. That combination usually means your page is being shown, and increasingly summarised, without the user needing to visit. It is the clearest evidence available that AI Overviews are answering on your behalf.
That is not automatically bad. Being cited in an answer has brand value even without a click, and the clicks that do come tend to be from people who wanted more than the summary gave them. But it does mean judging performance on clicks alone will make a page that is doing well look like it is declining.
Server Logs: The Fuller Picture
If you have access to server logs, they show something analytics cannot: the AI crawlers themselves.
GPTBot, PerplexityBot, ClaudeBot and others identify themselves in the user agent string. Their crawl activity tells you which of your pages AI systems are actually reading, which is upstream of whether you get cited at all.
Worth checking two things while you are in there: that those crawlers are not being blocked in robots.txt, and that they are not being blocked at the CDN. Bot protection on Cloudflare and similar services can block AI crawlers by default, which quietly removes you from consideration entirely. It is a common and invisible problem.
What to Report to a Client, Honestly
The temptation is to present AI traffic numbers with more confidence than they deserve. A more defensible framing:
- Report measurable AI referrals as a distinct channel, with the caveat that it is a floor rather than a total
- Report direct traffic trends alongside it as a directional indicator
- Report impressions versus clicks from Search Console as the AI Overview signal
- Say plainly that full attribution is not currently possible, and why
That last point protects everyone. A client told AI traffic is fully measurable will eventually discover it is not, and the credibility cost of that is larger than the discomfort of saying so up front. For the conversion tracking underneath all of this, see our guide to GA4 conversion tracking for lead generation.
UTM Parameters: Useful, With One Limitation
UTM parameters are the reliable way to label traffic you control, and they are worth using consistently wherever you can add them.
The limitation is important: you cannot add a UTM to a link an AI assistant generates from its own index. Where UTMs genuinely help is anywhere you place a URL yourself, such as a directory listing, a partner site, a PDF, an email or a social profile. Tagging those removes a chunk of ambiguity from the direct bucket, which makes whatever remains a cleaner signal.
Keep the convention simple and write it down. Inconsistent tagging, where one campaign uses “newsletter” and another uses “email”, produces fragmented reporting that is worse than no tagging at all.
What to Actually Do With the Data
Separating AI traffic is only worth the effort if it changes a decision. Three that it reasonably can:
Which pages get cited. If AI referrals concentrate on three or four pages, look at what those pages have in common. Usually it is a clear definition early on, a genuine FAQ, and a structure that answers a question directly. That is a repeatable pattern worth applying elsewhere.
Whether the traffic converts. Compare conversion rate for AI referrals against organic and paid. It is commonly higher, because the visitor arrives pre-qualified by the assistant’s answer. If it is much lower, the page is probably being cited for something it does not actually deliver.
Whether to keep investing in AEO. A channel that is growing month on month and converting well justifies continued work. A flat line after six months of effort is worth knowing too. Neither judgement is possible while the traffic sits inside direct.
The honest caveat: these numbers are a floor, not a total. Decisions should be directional rather than precise, and anyone presenting AI attribution as exact is overstating what the tooling currently supports.
Setting a Baseline Before You Optimise
The most common mistake with a new measurement setup is starting to change things immediately, which leaves nothing to compare against.
Before making content or technical changes aimed at AI visibility, record where you are. Capture current direct traffic volume and what share of it lands on deep pages, current AI referral sessions by platform, current impressions and clicks from Search Console for your main queries, and conversion rate by channel. A screenshot and a dated note is enough.
This matters more than usual here because AI search visibility moves for reasons outside your control. Platforms change how they cite sources, and volumes shift without anything on your site changing. Without a baseline, a rise looks like proof your work succeeded and a fall looks like proof it failed, when either could be the platform rather than you.
Review monthly rather than weekly. The numbers are small enough that weekly readings are mostly noise, and monthly comparison against a recorded baseline is the first point at which the trend means anything.
Frequently Asked Questions
Why does ChatGPT traffic show as direct in Google Analytics?
Because chat interfaces often strip or never send a referrer, and many users copy a URL out of a chat rather than clicking it. Without a referrer, GA4 has nowhere to file the session except direct. Some platforms do pass a referrer, which is why a portion of AI traffic is trackable and a portion is not.
Can I track traffic from Google AI Overviews separately?
Not currently. A click on an AI Overview citation still originates from a Google search results page, so it reports as ordinary organic search. The available signal is in Search Console: queries where impressions rise while clicks stay flat generally indicate AI Overviews answering the question without a visit.
How do I set up AI referral tracking in GA4?
Create a custom channel group under Admin, Data display, Channel groups, with a rule matching known AI platform domains such as chatgpt.com, perplexity.ai and copilot.microsoft.com. That makes AI referrals a visible line in acquisition reports rather than something buried in referral traffic or hidden in direct.
Is AI search traffic worth tracking if the volume is low?
Yes, for two reasons. Volume is growing, and a channel nobody measures is a channel nobody optimises. AI referrals also tend to convert well, because the visitor arrives having already been recommended, which makes even small volumes disproportionately valuable.
How can I tell whether AI crawlers are reading my site?
Server logs show them by name: GPTBot, PerplexityBot, ClaudeBot and others identify themselves in the user agent. While checking, confirm they are not blocked in robots.txt or by CDN bot protection. Blocking AI crawlers at the CDN is common, usually unintentional, and removes the site from consideration entirely.
What should I tell a client about AI traffic measurement?
That measurable AI referrals are a floor, not a total; that direct traffic trends are a directional indicator; that impressions versus clicks in Search Console signals AI Overview activity; and that complete attribution is not currently possible. Overstating the precision creates a credibility problem later.
Setting This Up This Week
The practical sequence: build the custom channel group in GA4 first, since it takes minutes and makes the channel visible in reports people already read. Then create a saved segment for deeper analysis. Then check robots.txt and your CDN settings to confirm AI crawlers can actually reach the site.
That last step is the one most likely to find something. Being invisible to AI search because a bot rule blocked the crawler is a considerably bigger problem than imperfect attribution, and it is far easier to fix.