AI attribution is hard because the recommendation and the visit are different events. Someone asks ChatGPT which company to hire, sees your name, closes the tab, and searches that name on Google two days later. Analytics records an organic visit. The AI answer that started it gets nothing.
Similarweb measured that gap with its opted-in U.S. desktop panel from July through December 2025, across finance, travel and beauty. People who were recommended a brand in ChatGPT were 2.5 times more likely to visit that brand within seven days than people who were recommended a competitor. The study left out anyone who had already visited the brand in the prior four weeks, or who had typed the brand into the prompt (Similarweb, reported in detail by Search Engine Land).
The visit is real. The source your report names is often the second step, not the first.
Most of those visits look like ordinary search
Similarweb found that 55.9% of the visits that followed an AI recommendation arrived through search. Visits with no AI recommendation arrived through search 40.4% of the time. Direct traffic was the reverse: 19.9% after an AI recommendation, against 38.8% for other visits.
That is the attribution problem in one comparison. The person remembered the name and searched it. Google Analytics and Search Console then file the session under organic or branded search. An "AI referral" report stays almost empty, so a last-click model treats the channel as if it barely exists.
The same visitors also stayed longer once they arrived: about 12 pages and 11.8 minutes, against 6.5 pages and 5.6 minutes for other visits to the same kinds of sites. They had already narrowed the choice in the chat. The click was the end of the research, not the start.
Last click was built for a shorter path
Last-click attribution is fine when someone sees an ad or a result and converts on that visit. AI search inserts a step that never hits your site.
A typical path now looks like this:
- The person asks an AI tool a buying question.
- The answer names two or three businesses. There may be no link, or they do not click it.
- Days later they search the name, open the site, and call or fill in a form.
The form submission is credited to branded organic, or to direct if they typed the URL. Both labels describe how they arrived. Neither describes why they knew the name.
This does not make Search Console useless. It means a report that only counts AI referral sessions will understate the work. The same caution applies to SEO. A page can be the reason an AI answer mentions you, and the conversion can still land on a different channel a week later.
What to track instead
There is no clean AI column you can add to GA4 and trust. Google folds AI Overview and AI Mode activity into normal web search in Search Console, so you will not get a filter that says "this click came from an AI answer." We cover that limit in the measurement section of AEO explained. Use a small set of signals together.
- A fixed prompt list. Write 10 to 20 questions a customer would actually ask. Once a month, check ChatGPT, Perplexity, Gemini and Google's AI answers. Record who is named. Screenshots beat memory, because the answer changes.
- Branded queries in Search Console. Filter queries that contain the business name. If mentions in AI answers rise and branded impressions rise in the same months, that is a clue, not proof.
- Direct traffic and calls. For a local business, the number that matters is the call or the form, tied to the landing page. A citation that never produces an enquiry is visibility, not a lead.
- Referral hostnames, as a minority. Still look for chat.openai.com, chatgpt.com, perplexity.ai and similar hosts. Treat them as the visible slice, not the whole channel.
On a Florida law firm, the same SEO and answer-page work that lifted organic clicks 90.6% in six months also put 61 pages into AI answers. We did not assign those citations a dollar value in GA4. We kept the Search Console trend, the citation log and the calls from the Google Business Profile side by side. That is the honest version of AI reporting. The numbers are in the law firm case study.
What not to conclude
- A branded-search bump is not automatically AI. A review, a news mention or a seasonal rush can do the same thing. Similarweb showed an association for people who had just been recommended, on U.S. desktop, in three industries. It did not prove that every later branded visit on your site came from ChatGPT.
- Referral traffic is the small piece. If leadership only looks at the AI referral line, they will cut the work that creates the recommendation and keep the work that receives the click.
- You still have to be readable. An answer cannot cite a page the crawler cannot fetch. Check that before you argue about attribution. Our AI crawler checker tests the common bots.
A practical monthly check
- Run the same 15 prompts. Note whether you are named, and whether a competitor is named instead.
- Export branded queries from Search Console for that month and the month before.
- Count calls and form leads. Separate them from rankings.
- Write one sentence: mentions up or down, branded demand up or down, leads up or down. If the three disagree, say so.
That sentence is more useful than a last-click channel labeled "AI" with three sessions in it.
Questions we get asked
What is AI attribution?
It is the attempt to connect an AI recommendation to a later visit, search or lead. The recommendation often happens off your site, so standard analytics never sees it.
Why do AI referrals look so small?
Because many people do not click the answer. They remember the name and come back through Google or by typing the address. In the Similarweb panel, 55.9% of the visits after a ChatGPT recommendation arrived through search.
Should we stop reporting organic clicks?
No. Keep organic clicks, impressions and leads. Add a citation log and branded-query trend next to them, and stop treating the AI referral report as the score for the channel.