Similarweb published an industry cut of its AI referral data on September 3 and updated it two weeks later. For the twelve months ending May 2026, finance sites received about 19.1 million visits a month from links inside AI answers on ChatGPT, Gemini, Perplexity, Claude and the rest. That is up 236 percent year over year. Across all industries the total was 770 million visits a month, more than double the prior year.
I read numbers like that with some caution. Similarweb is estimating from a panel, and finance is a wide bucket that includes Bankrate and NerdWallet and every crypto exchange. A credit union in central Texas is a rounding error inside it.
But one line in that report matters more than the headline figure, and it is the line almost everyone skipped. Similarweb notes that some AI visits arrive without referrer data and land in direct traffic. Their counts, they say, are a floor.
I have been staring at that floor in credit union analytics for two years.
Direct is where the unexplained goes
Every credit union I work with has the same shape in its traffic report. Organic search is flat or drifting down. Paid is whatever the budget allows. And direct has been creeping up, quarter after quarter, with nobody able to say why. When I ask, the answer is usually a shrug and something about members bookmarking online banking.
Some of that is true. Online banking logins do inflate direct. But when I pull direct sessions that land on a rates page, an auto loan page, or a branch page rather than the login screen, the pattern changes. Those are people arriving with a specific question already formed. They did not type a rates URL from memory. Something sent them.
Increasingly, that something is an AI assistant. ChatGPT passes a referrer most of the time now. Gemini and the assistants inside mobile apps often do not. Copy a link out of an answer and paste it into a browser, and there is no referrer at all. Analytics files it as direct, and the assistant that started the conversation gets no credit.
The reporting problem underneath it
This is where it stops being a curiosity and starts costing money.
Most credit union marketing reports are still last-click. Whatever touched the member last before the application gets the credit. Paid search is very good at being last. A member asks an assistant which local credit union has the best used auto rate, gets a name, googles that name, clicks the branded ad, and applies. The report says paid search closed a loan. The assistant that made the recommendation is invisible, and the next budget conversation goes accordingly.
I have sat on the other side of that table. For most of my career I was the one presenting the report, and I would have trusted it. The report was not lying. It was measuring the last thing it could see.
Right now, in client work, I am watching a related problem from a different angle. The web analytics count of completed loans and the number of loans actually booked do not match, and they never will, because part of the conversion finishes offline after the website. The fix is not to force the two to agree to the unit. The fix is to know why they differ and by roughly how much, so leadership stops arguing about the gap and starts using the trend. Direct traffic deserves the same treatment. You will never attribute every visit. You can know what is in the bucket.
Three things I would do this week
First, split direct traffic by landing page. Separate online banking and the home page from everything else. What remains is your unexplained direct, and that is the number to watch month over month.
Second, build one segment for known AI referrers. ChatGPT, Perplexity, Gemini, Copilot, Claude. It is a short list and it takes twenty minutes. That segment will be small. It will also be growing faster than anything else on the report, and a small number that triples is a story worth telling a CEO.
Third, stop presenting last-click as the whole truth. Show it, then show the assisted view next to it. Where a channel appears earlier in the path and disappears at the close, say so out loud.
None of this makes the AI search question go away. It does mean you can see whether the work is landing. Right now, for most credit unions, the answer is already yes, and the report says no.