Skip to content
Neoval
Sign in

Blog

Buyers did not leave search. The click did.

· 7 min read · by

A web page of grey text lines with one line highlighted, read on the left by a browser of search results and on the right by an AI answer card that repeats that line; the paths back to the page are broken dashed lines ending in empty circles.
A web page of grey text lines with one line highlighted, read on the left by a browser of search results and on the right by an AI answer card that repeats that line; the paths back to the page are broken dashed lines ending in empty circles.

The story you keep reading is that buyers abandoned Google for ChatGPT. The data says something less dramatic and much more useful: they kept the search habit and added a second one, and the thing that actually disappeared was the click. If you plan around the headline, you will chase a migration that is not happening. If you plan around the click, you will fix the problem you really have.

The switch that did not happen

The assistants are enormous. OpenAI announced 900 million weekly active users on 27 February 2026. Google said on 22 July 2026 that AI Mode had passed a billion monthly users and that its Gemini app was at 950 million monthly users, with daily users tripling in a year.

Search is still bigger, and it is not shrinking. Similarweb, measuring March to May 2026, put search at 3.3 billion average monthly unique visitors worldwide against 655 million for every AI chatbot combined: roughly five to one, even after the chatbots grew 57 percent year over year. The same study found the overlap almost total. Of ChatGPT's 494 million users in that window, 461 million, or 95 percent, also used Google in the same period. Google's own framing on the July earnings call was "an expansionary moment" with queries growing, not falling.

So the honest picture is not migration. It is layering. The same person asks an assistant to narrow the field and then searches to check what it said, or searches first and asks an assistant to explain the results. Two habits, one buyer, and both of them now happen before anyone reaches your site.

What did change: the answer ends the trip

Here is the number that matters. Pew Research Center watched the browsing of 900 US adults across 68,879 Google searches in March 2025. When an AI summary appeared, people clicked a traditional result in 8 percent of visits. When no summary appeared, they clicked in 15 percent: nearly twice as often. Clicking a link inside the summary happened in 1 percent of visits. In that dataset, 18 percent of searches produced a summary at all, and 58 percent of the people in it saw at least one during the month.

Assistants are stingier still. Similarweb found that 6.8 percent of ChatGPT answers in the US included a link to an external source in May 2026. That is a fivefold rise from about 1 percent a year earlier, and it still means that more than nine answers in ten mention no source you can click.

Read those two findings together and the shape of the problem is clear. Your page is still being read. It is being read by a machine, summarised, and delivered without the visit that used to prove it happened. The traffic chart that looks like decline is often a citation chart nobody is drawing.

The visits that do arrive are worth more

The flip side is the part most reporting skips. Adobe Analytics, measuring US retail sites in July 2026, found AI-referred traffic up 62 percent year over year, converting at a rate 60 percent higher than non-AI traffic, and generating 53 percent more revenue per visit. Those visitors spent 59 percent more time on site, bounced 33 percent less, and added to cart 28 percent more often.

That is what a pre-qualified visitor looks like. The assistant did the comparing, the visitor arrived with the decision half made, and the site's job shrank to confirming it. One more detail from Similarweb makes this concrete: 58.8 percent of the referral traffic AI sends back lands on the homepage. Not a product page, not a comparison page, the homepage. Which means your homepage is now a landing page for people who already know what they want and are checking whether you are real.

What this changes for a small site

Three consequences follow, and none of them require you to abandon the work you have already done.

  1. Measure citations, not just rankings. Ranking third on a page nobody scrolls, and being absent from an answer everybody reads, look identical in an analytics dashboard: flat traffic. They are not the same problem and they do not have the same fix.
  2. Write the sentence a machine can repeat.An assistant has to summarise what you do in one line before it will name you. If your homepage never states the category, the audience and the market in plain words, the model borrows someone else's description of you, or skips you.
  3. Check that the crawlers are allowed in at all.A robots.txt that blocks GPTBot or ClaudeBot does not make you invisible in a polite way. It means the only pages an assistant can quote about your category are your competitors' pages and the review sites you are not on.

There is a fourth, and it is the uncomfortable one: most of what gets cited is not on your site. Assistants weight what other people write about you, because anyone can call themselves the best on their own domain. That is slow work, and no technical fix replaces it. What the technical work does is make sure that when somebody finally does write about you, the machine reading that sentence can identify, classify and quote you correctly.

How to check your own numbers

Every figure above is somebody else's aggregate. Yours will differ, and the only way to know how is to ask the engines about your own domain and record what comes back. Our free audit does exactly that for one domain: it puts five buyer prompts to ChatGPT and records, for each one, whether you were cited, who was cited instead, and the source the answer leaned on, with the query, the engine and the date attached. It also checks whether the crawlers can read you at all. Unavailable evidence is marked not observed, never estimated.

The deep audit widens that to twenty prompts across five engines, repeated over three days, so a citation that holds can be told apart from a lucky sample. Either way, the point is the same: stop inferring your AI visibility from a traffic chart that was never measuring it.

Sources