AI assistants do send people to websites. Not as many as search, and the volume varies enormously by topic, but the visits are real, they convert unusually well, and almost every analytics setup reports them wrongly.
Two specific misconfigurations account for nearly all of it. Both are fixable in an afternoon.
Where the visits come from
When somebody clicks a citation in an AI answer, you get an ordinary browser request with an ordinary referrer. The hosts you are looking for include:
chatgpt.comandchat.openai.comperplexity.aiclaude.aicopilot.microsoft.comgemini.google.commeta.ai,grok.com,chat.deepseek.com,chat.mistral.aiyou.com,phind.com,poe.com
That is a closed and knowable list, which is what makes this tractable: unlike most channel classification, you are matching against a set of hosts you can enumerate rather than guessing from patterns.
Mistake one: ChatGPT gets filed as a campaign
This is the big one, and it is almost universal.
ChatGPT appends a UTM parameter to the links it hands out. A click arrives looking roughly like this:
https://yoursite.com/page?utm_source=chatgpt.comEvery analytics tool in existence gives campaign parameters priority over the referrer header — correctly, in general, because a UTM tag is an explicit statement of where a link was placed and the referrer is an inference.
But the consequence here is that your single largest AI surface is filed under Campaigns, next to your newsletter and your paid ads, rather than under any AI or referral grouping. Your AI number is missing its biggest contributor, and your campaign report contains traffic from a campaign you never ran.
The fix is to classify AI referrers before the campaign branch, not after. In most tools this is a custom channel group that matches the source against the host list above and wins over the default campaign rule. If your tool cannot reorder that, you can at least stop trusting the AI number it shows you.
Mistake two: Gemini gets filed as Google search
Subtler, and it bites everyone who writes their own classification.
The usual test for organic search is a referrer host containing google.. That pattern matches gemini.google.com perfectly. So every visit from Google's standalone assistant lands in your organic search bucket, inflating a number you make SEO decisions with and deflating one you are trying to learn about.
Gemini is not the search engine. Match AI hosts first, then search.
The thing that genuinely cannot be measured
There is a category here that no configuration recovers, and any tool claiming otherwise is inventing a number.
When Google shows an AI Overview at the top of a results page and somebody clicks a link inside it, the click carries google.com as its referrer — exactly like an ordinary organic click. There is no parameter, no distinguishing header, nothing. The same is true of Bing's Copilot answers within a results page.
You cannot separate "clicked a blue link" from "clicked a citation inside the AI summary above the blue links". Not with a clever rule, not with a heuristic, not with anything. The signal does not exist on your side of the connection.
This matters because it is the largest AI surface of all, and it is precisely the one that is invisible. Anyone who reports a confident AI traffic number without mentioning this is reporting a number that excludes most of the phenomenon. The unmeasurable part deserves its own explanation.
What good looks like
Once classification is right, three things are worth watching monthly:
AI as a share of all visits. Small for most sites today, growing, and the trend is the signal rather than the level.
Which assistant, and which landing page. These are wildly uneven. One or two pages usually account for most AI referrals, and they are rarely the pages you would have guessed.
Behaviour against other channels. AI referrals typically arrive further along in intent than search visitors, because the assistant has already answered the general question and the person clicked through for something specific. If your AI traffic converts better than organic, that is normal, and it is an argument for writing more of whatever earned the citation.
The other half of the loop — what the crawlers took in order to produce those citations — is crawler visibility, and the two together are the only way to see the full exchange.
Common questions
Why does ChatGPT traffic show up as a campaign in Google Analytics?
Because ChatGPT appends utm_source=chatgpt.com to the links it hands out, and every analytics tool gives campaign parameters priority over the referrer. GA4 therefore files it under Campaigns rather than as a referral or AI source. The fix is a custom channel group that matches the AI hosts before the campaign rule is applied — otherwise your largest AI surface is counted as marketing you never ran.
Can I see traffic from Google's AI Overviews separately?
No. A click on a citation inside an AI Overview sends the same google.com referrer as a click on an ordinary organic result, with no parameter or header to tell them apart. The distinction does not exist in the data your server receives, so no tool can report it honestly. Any product showing you an AI Overviews traffic figure is estimating, and should say so.
Does AI referral traffic convert better than search?
Frequently, yes, and there is a straightforward reason: the assistant has already answered the general question, so a person who clicks through is after something specific the answer did not cover. They arrive further along than a typical search visitor. Volume is much lower, so the useful comparison is conversion rate rather than total conversions.