Spotting and evaluating AI traffic in Google Analytics

Why AI traffic is a channel of its own

More and more people no longer research on Google but ask ChatGPT, Perplexity, Gemini or Copilot directly. When these assistants name your website as a source and the user clicks, a visit is created that is neither classic search nor classic social. For you as the operator of an online shop, a law firm or a trade business, this is a new path along which customers arrive. And anything that influences revenue is something you want to be able to measure.

The problem: by default this traffic lands in the most unhelpful places in GA4. Sometimes it shows up as referral, sometimes as direct, sometimes completely invisible. Without deliberate setup you never even see that a noticeable share of your new contacts comes from AI answers. You then judge channels wrongly and pour budget into the wrong measures, because the data foundation lies.

The first step is therefore a mindset, not a click: treat AI assistants as a traffic source of their own, just as you look at Google search, newsletters or Instagram separately. Once you accept that, the rest of the setup almost follows on its own, because you know what you are looking for and which reports you need to build.

Which referrers you need to know

AI visitors often give themselves away through the referring domain. The most important referrers are chatgpt.com and chat.openai.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. Add to that search engines with AI overviews, for example on Bing. Some assistants also use redirect domains, so it pays to check your referral list regularly to catch new patterns early. The landscape changes fast, and fixed lists go stale.

In GA4 you find these domains in the Acquisition report under Traffic acquisition, once you choose Session source or Session referrer as the dimension. Filter by the known domains and you immediately see an initial volume. It is important to account for upper and lower case as well as subdomains, otherwise sessions slip through your net and the numbers look smaller than they actually are.

The referrer problem: why much of it lands as direct

Here comes the honest part that many guides leave out: a large share of AI traffic arrives without a referrer. ChatGPT and other assistants open links partly in their own in-app browsers, via desktop apps or with strict privacy settings. Then the information about where the click came from is missing, and GA4 sorts the session into the Direct channel, the same place where bookmarks and directly typed addresses usually land.

In concrete terms: your referral numbers are almost always a lower bound. The true AI share is higher, you only see the part that dutifully sends along a referrer. Anyone who looks solely at the domain list systematically underestimates the channel. You have to know this blind spot, otherwise you make decisions based on a flattering half of reality.

The countermeasure is a bundle of clues rather than a single clean metric. You combine the visible referrers with striking patterns in the direct traffic: sudden spikes on deep subpages, unusual landing pages without a campaign, new users with no history. No signal alone is proof, but together they form a robust picture of your real AI traffic.

Building a dedicated channel in GA4

So that you do not have to gather AI traffic by hand for every report, you set up a custom channel group. You do this in the admin area under Channel groups, where you create a new group and define a rule: if the source contains chatgpt.com, perplexity.ai, gemini.google.com and others, assign the session to the AI assistants channel. You move this rule up in the order so that it takes effect before other rules classify the session as referral.

The advantage of a dedicated channel group is clarity. Instead of individual domains you see one clean channel that you can compare, like any other, by users, conversions and revenue. A tax adviser can thereby see whether client enquiries come from AI recommendations, a furniture retailer whether AI answers trigger purchases. The channel becomes comparable to Google search or email, and thus finally manageable.

Setting UTM parameters and your own measuring points

Wherever you control the links yourself, you should tag them. If, for example, you place URLs in a specialist article, a public knowledge base or a product data feed that AI systems might pick up, append UTM parameters, for instance utm_source=chatgpt and utm_medium=ai. If a click appears with this parameter, the attribution is unambiguous and even survives the loss of the referrer, because the information sits in the URL itself.

In addition, a dedicated event pays off. In GA4 you can trigger an event via Google Tag Manager as soon as the referrer or a URL parameter points to an AI assistant. This way you collect a clean, permanent metric that works independently of the standard channel rules. That is especially handy when you later want to evaluate goals and conversions specifically for this segment.

Remember to document your UTM convention and use it consistently across the team. Inconsistent spellings like AI, ki and chatgpt-app splinter your channel into many little fragments, and in the end no one sees the overall picture any more. A short, binding list of allowed values prevents this chaos and saves you hours of clean-up work in the reports later.

Evaluate behavior instead of just counting

Counting visitors is the easy part. What they do is more interesting. For your AI channel, compare the engagement rate, the session duration, the pages per session and above all the conversions with other channels. Often an interesting pattern emerges: AI visitors arrive with a very specific question and land deep inside your website, but bounce faster if the answer does not fit immediately. They are more targeted, but also more impatient.

For your assessment this means: raw click counts can deceive. A service provider with few but highly qualified AI enquiries can draw more value from twenty visitors than from two hundred fleeting social clicks. So always look at the chain right up to the conversion, not just at the entry point. A channel with small reach and a high closing rate deserves more attention than its bare visitor count would suggest.

A time series is useful. Set up a comparison report that shows the development of the AI channel over weeks. Because the channel is young and growing strongly, you spot trends earlier than with established sources. A steadily rising share is a clear signal that you should prepare content more deliberately for AI visibility, before competitors use this head start for themselves.

Simple reporting you look at every month

Build yourself a lean report that you can read without thinking. At a glance it should show four things: the volume of the AI channel, its engagement quality, the conversions and the development over time. Avoid data graveyards with thirty metrics. A report no one understands will not be used, and unused data is worthless. Fewer metrics, but looked at regularly, beat any overloaded dashboard.

Add qualitative spot checks. Look specifically at which pages AI visitors land on, and check whether these pages really answer the presumed question. Often a handful of such glances reveals more than any statistic, for example that a particular guide page is recommended by assistants above average and therefore deserves to be expanded.

Framing the limits of measurement honestly

Be realistic: you will never measure AI traffic to a hundred percent cleanly. Referrers are missing, assistants change their behavior, new providers appear, and privacy mechanisms deliberately obscure origins. Your numbers are a good approximation, not a precision instrument. Whoever accepts this works more calmly and still makes better decisions than someone waiting for the one perfect metric that simply does not exist in this environment.

So the rule is: use the data for directional decisions, not for decimal places. The core question is not whether it was 4.2 or 4.7 percent AI traffic, but whether the channel is growing, whether it converts and whether it deserves your content strategy. You can answer these questions reliably enough with the methods described to act on them.

Also review your setup regularly, at least every quarter. New AI services get new domains, old ones change their redirects. Keep your referrer list, your channel rules and your UTM convention up to date, otherwise your reporting slowly drifts away from reality without you noticing. A short maintenance slot in the calendar saves you from nasty surprises in the numbers.

A worked example: what sits behind the numbers

Suppose you count 480 sessions in one month that you can attribute to the AI channel. At first glance that sounds like little. But look at the behavior: the average session duration is 3:10 minutes, while your overall average is 1:40. The bounce rate is half as high, and 38 of these sessions lead to a contact enquiry. That is almost eight percent conversion from a channel you did not even see before.

Put it in context: if your website generates two enquiries from 100 sessions on average, the AI traffic is at four times that. The reason is simple. People who come across you through an AI answer have already half-solved their question and arrive with clear intent. They are no longer comparing ten providers, they are checking you out as a concrete suggestion.

So it is not enough to judge AI traffic by volume alone. A small channel with high quality can be worth more than a large one with scatter losses. Enter the conversion per channel side by side in your reporting, and you see this difference immediately and make better decisions about where to sharpen your content.

Industry differences: not everyone sees the same thing

How strongly AI traffic shows up for you depends noticeably on your industry. In B2B with services that need explaining, people often ask AI systems for comparisons, definitions and provider lists. Above-average numbers of referrals land here, because the search is complex and an assistant takes work off your hands. Consulting, software and specialist services feel it first.

In local business the picture is different. Anyone looking for a restaurant, a tradesperson or a hotel often gets a direct recommendation from AI systems with name and address, without a click to your site being necessary. The effect is real, but barely visible in analytics. All the more important here are soft signals like more calls or enquiries that trace back to no measurable channel.

In e-commerce, meanwhile, patterns shift toward product research and comparisons. Before you adopt benchmarks from outside sources, better measure your own baseline over two or three months. What is a strong channel in one industry can remain a fringe phenomenon in another, which you should nonetheless keep an eye on.

Common questions that keep coming up

Can I identify individual AI users? No, and that is not your goal either. You work with aggregated patterns, not with people. Anyone trying to de-anonymize individual visitors via AI referrers wastes time and risks privacy problems. Stay at the level of channels, trends and behavior.

Should I check my reporting daily? For most, a fixed monthly rhythm is enough. AI traffic fluctuates strongly from day to day and the numbers are too small to draw daily conclusions from them. A monthly look shows you the trend without you getting lost in noise. Only after bigger changes to your content is a closer interim look worthwhile.

What if the numbers stay very low? That is normal and no cause for concern. A channel does not have to be big to be valuable. Document the baseline, watch the direction over several months and judge the quality. A slowly growing, high-quality channel is a good sign, not a disappointment.

A roadmap for the first three months

In the first month it is only about setting up and collecting. Build your channel, set your measuring points and let everything run without judging yet. Note the state as a starting point, so that you have an honest basis for comparison later. Resist the temptation to draw conclusions from the first few days.

In the second month you start reading. Look at which pages AI visitors head for and how they behave. Compare it with your other channels. Now first hypotheses arise about which content is picked up by AI systems and where you could sharpen things.

In the third month you act deliberately. Improve the pages that already receive referrals, and check in the following month whether anything moves. This turns measurement into a cycle: observe, adjust, measure again. After these three months you have a robust routine that you can keep up permanently with little effort.