AI Trafiğini GA4'te Ölçmek ve Dönüşüme Bağlamak
First Monday of the month, and GA4's Traffic acquisition report is up on the screen. At the top of the channel list sits "Direct", carrying a sizeable share of sessions on its own. Someone enlarges the screen and asks: "Who are these people?" Nobody at the table can answer. The next item on the agenda is a review of the GEO budget that has been paid for three months. The return on that budget is in all likelihood sitting inside the exact bucket nobody could just put a name to — but because it does not show up as a line in the report, it cannot be defended.
Here is the issue: traffic from AI assistants is not missing on most sites, it is in the wrong bucket. GA4's default setup gives these visitors no name of their own. And a line item with no name gets treated as if it were not there in the budget meeting. The steps below are a few hours of configuration, and at the end of them the answer to "how many people came from AI, and what did they do" comes out of your own property.
Why the traffic disappears
When a browser moves between sites, it carries a referrer telling the destination where the visitor came from. GA4 reads that and writes a source onto the session. If the referrer arrives empty, GA4 has nothing to write and drops the session into the Direct bucket.
With AI assistants, that information falls away at three separate points:
- Mobile apps. When you tap a source in the ChatGPT, Perplexity or Gemini phone app, the link opens in an in-app browser. Many of these embedded browsers send no referrer. The same person clicking the same link on desktop via
chatgpt.comusually arrives with the source intact. So the size of your channel varies with how your audience splits across devices. - Copy-paste behaviour. The model sometimes gives the URL inside the text, and the user copies the address and pastes it into the address bar. That is direct traffic by definition; no setting rescues it.
- The visit that turns into a brand search. The assistant describes you, the user is convinced but does not click the link; they type your brand name into Google instead. The session is recorded as
Organic Search. The influence came from AI, the credit went to search.
There is no way to solve all three completely. But making visible, as a channel of its own, the portion where the referrer does arrive intact — and it is not a small portion — is entirely within your control.
Setting up an AI assistants channel in GA4
GA4's default channel group does not recognise these domains; it counts some as Referral and some as Organic Search. You have to define your own channel.
The path: Admin → in the property column, Data display → Channel groups → Create new channel group. Give the group a meaningful name (for example "Including AI assistants"), then open a new channel with Add channel. Channel name: AI Assistants. For the condition, pick the Source dimension, set the operator to contains, and add the domains below one by one as OR conditions.
| Domain | Where it comes from |
|---|---|
chatgpt.com | Source links inside ChatGPT answers |
openai.com | ChatGPT's older addresses and OpenAI pages |
perplexity.ai | The source list on Perplexity answers |
gemini.google.com | The Gemini chat interface |
copilot.microsoft.com | Microsoft Copilot |
bing.com/chat | The chat interface inside Bing |
claude.ai | Links inside Claude answers |
you.com | You.com chat and search |
Three warnings. First, order matters: move your new channel up the channel list. GA4 evaluates rules top to bottom; if Organic Search or Referral sits above it, that rule catches the session first. Second, bing.com/chat is a path expression, and the source field usually carries only the domain, so Bing chat can end up in the same bucket as organic Bing. Test it against your own data before adding it, and drop it from the list if it does not separate out. Third, once you create a custom channel group the report does not settle instantly; if the first day looks lower than you expected, give it a few days.
Verifying the setup in half an hour
Do not build the channel group and then wait for the report to fill in on its own; a day of wrong configuration means a month of lost data. Test it with your own hands.
Start on desktop: open an assistant and ask about your brand or a topic you have written about. If your site shows up as a source in the answer, click the link. Then in GA4 open Reports → Realtime and look at the card list to see which source/medium the session arrived with. If you see the domain, the referrer is intact. Now repeat the same visit through the phone app: this time you will most likely see the source arrive empty, meaning the session is counted as direct. The difference between those two attempts lets you explain to management in one sentence why your channel looks smaller than it is.
If you do not appear in the answer at all, there is no click to test; the problem is not measurement but visibility, and that has to be solved first. When the test is done, open the channel group rule once more and check the domains for typos — the most common fault is a single wrong letter that stops the rule matching anything at all.
Opening up the channel with an Explore report
The channel group gives you the total; the real information is in the breakdown. Open Explore → Free form.
- Rows: Session source / medium. Add Landing page as a second row dimension — that pair shows which of your content is pulling traffic from AI.
- Columns: Device category. The gap between mobile and desktop gives you a sense of how much referrer loss you are carrying.
- Metrics: Sessions, Total users, Key events, Session key event rate, Average engagement time.
There is a reason for putting average engagement time on the list: a visitor from AI has usually had part of their question answered already and comes to verify. The time that visit spends on the site is not the same thing as the time from a social media click, and showing management that difference is the fastest way to explain the channel's quality.
Once the exploration is built, save it and fix the name. Rebuilding the report from scratch each month leads you to change a dimension or a date range between one month and the next, and to compare apples with oranges without noticing. Keep the date range on a fixed window too, something like 28 days; months are not equal in length, so "last month" comparisons mislead at low volume.
Where UTM helps and where it does not
UTM parameters work on links whose distribution you control: the link you send in a newsletter, the post you share on social, the address you hand to a partner. Tag those, because you decide the source.
On links the model finds by itself, UTM has no force at all. The assistant quotes your page's canonical address; it does not know the parameter you added and will not invent it. So putting UTM into your sitemap, your llms.txt or your canonical tag to "capture AI traffic" does not work — and it is actively harmful, because it pushes two versions of the same page, with and without parameters, into the indexing queue. The rule is simple: UTM for outbound links, channel groups for inbound ones.
Tying conversions to the AI session
Counting sessions does not defend a budget on its own; it has to line up with conversions. First make sure your key events are defined: form submission, click to phone number, WhatsApp link, price list download. Mark the relevant event as a key event on the Admin → Events screen.
Then look at the key event column on the AI channel row in your Explore report. There are two traps here.
The first is the visit that does not convert in one session. The user arriving from AI often does not decide on first contact; they read, leave, come back directly a few days later, and fill in the form then. Look at a session-based report and the conversion is written to the second session — to Direct — and the AI channel shows zero. For this, open the Advertising → Attribution → Conversion paths report and select your own channel group. That report shows where AI sits on the path — first touch or intermediate step.
The second is calculating rates on small data. Writing a percentage into the report because one of seven sessions converted produces a "the channel has collapsed" headline next month when a single form goes missing. At that scale, a rate is noise.
Server logs answer a different question
GA4 counts people who run JavaScript in a browser. AI crawlers — GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot — generally never enter that measurement at all. To see them you have to look at hosting or CDN access logs, broken down by user-agent.
Confusing the two is a common mistake. The bot log answers "is my content being crawled"; GA4 answers "are humans arriving". The four combinations read like this:
| Bot log | GA4 AI channel | What it means |
|---|---|---|
| Yes | Yes | The chain works; focus on growing volume |
| Yes | No | You are being crawled but not cited, or not clicked — content and measurement have to be checked separately |
| No | Yes | Missing configuration on the measurement or the logging side |
| No | No | An access block: robots.txt, CDN bot rules, firewall |
That last row is GEO's most common and most silent fault. Nobody closes robots.txt on purpose; a "block AI bots" switch in the CDN panel arrives on by default and goes unnoticed for months.
Cross-checking against the agency report
Say your visibility report is going well: your brand is named in more prompts, cited as a source in more answers. But the AI channel in GA4 is flat. That contradiction says one of three things.
First, there are citations but no clicks — the assistant completed the answer inside the answer, and the user had no need to click. That is not a loss, it is brand contact; but it does not turn into a session. Second, the questions you are named in are informational and carry no commercial intent. Third, and most common, the measurement is broken: the channel group was never built, or your audience is mostly in mobile apps.
Eliminate them in order. If a gap is still there once you have verified the measurement, the problem is not visibility itself but the nature of the questions you are visible in. We took up which metrics actually need tracking in a separate piece. For a setup that has shown no results for three months, our guide walking through the diagnostic order step by step is a better starting point. And if you are wondering how the search results themselves are changing, the effect of AI Overviews on businesses fills in the context.
Working with small volumes
At most small and mid-sized businesses the AI channel is low in volume but high in intent. The visitor has talked the topic through with an assistant, narrowed the options, and come to verify you. Reporting that channel as a percentage misleads.
Look at three things: absolute session count, absolute key event count and their monthly trend. You can add a fourth: which pages the visitors from the AI channel land on. Even at low volume that list shows which of your content the assistants find trustworthy, and puts your next content decision on data rather than guesswork. Collect monthly rather than weekly, and draw the slope over a 90-day window. Walking into a management meeting with a sentence like "the AI channel brought 140 sessions and 9 quote requests this quarter, against 60 sessions and 3 requests last quarter" is far more defensible than a slide full of ratios. To see what those numbers correspond to in demand in your own sector, look at the sector-by-sector scenarios; and to build the measurement habit more broadly, use the framework in customer data metrics for small businesses.
The setup is done once and the report arrives by itself every month. Total time is a few hours: half an hour for the channel group, half an hour for the exploration, half an hour for the verification test, and the rest on reviewing key events. In return you walk into the budget meeting with a line item that has a name. Next time, instead of looking at the Direct bucket in silence, you can say what the channel is called, how many sessions it brought and how many enquiries it turned into — which is exactly where the defence of a budget starts.
Frequently Asked Questions
Does setting up a custom channel group in GA4 fix historical data too?
GA4 does attempt to apply custom channel groups to past data once they are created, but this is not instant and the table can take a few days to settle. Seeing a lower figure than you expected on the first day does not mean the setup is wrong. Even so, the earlier you do it the longer the comparable history you hold, which is why the channel group should be defined at the start of the GEO work rather than at the end.
Can I see all of the traffic coming from AI assistants?
No, seeing all of it is not technically possible. The embedded browsers in mobile apps often do not carry the referrer, no source information survives when a user copies and pastes the address, and someone influenced by an assistant who then runs a brand search on Google is recorded as organic search. What you can measure is a lower bound; the real effect is always larger than what you see in the report, and it is right to say so plainly when you present it to management.
Why do the bot visits in my server logs not show up in GA4 reports?
GA4 measurement depends on JavaScript running in a browser; crawler bots like GPTBot, ClaudeBot or PerplexityBot fetch the page but generally do not run that script, so no session is created. This is not a fault, it is two tools answering different questions: the bot log shows whether your content is being crawled, GA4 shows whether human visitors are arriving. A sound GEO report puts the two side by side and does not use one in place of the other.
Can I tag AI traffic by adding UTM parameters to my pages?
You cannot, because the assistant quotes your page's canonical address and does not know about the parameter you added. UTM only means something on links you distribute yourself: newsletters, social media posts, and addresses you hand to partners. Putting UTM into your sitemap file, your canonical tag or your llms.txt does not capture the traffic, and it does create problems on the indexing side by producing versions of the same page with and without parameters.
If I only get a few AI sessions a month, is the channel a waste of time?
It is not a waste, but it looks like one when you use the wrong metric. At most small businesses this channel is low in volume yet high in intent; the visitor narrowed their options with an assistant and came to verify you. So look at absolute session count, absolute key event count and their monthly trend rather than percentages and rates. Conversion rates calculated on small bases swing hard when a single form goes missing, and make it look as though something real has changed.