8:30 PMResearch
What published research says about how people find businesses now that search engines and AI assistants answer first, and what we measured by asking the AI ourselves. Every figure lists who was measured, when, and where it came from.
Lehvel ResearchPublished Updated Next review
Google is still where most searches start, but fewer of them end in a click: in Pew's panel, people clicked a result in 8% of visits when an AI summary appeared, against 15% when none did.
In Lehvel's early sample, AI assistants named a business in every answer when asked about it by name, and in half of the answers when asked for the best provider of its service.
Nearly half of what the assistants said could not be checked against the businesses' recorded facts. The gap was missing facts more often than wrong ones.
Google says AI features need no special markup or files: the same crawlable, accurate, well-structured pages work for both.
The answer now often appears on the results page itself. When it does, fewer people click through to a website, and the few links inside the answer are rarely clicked.
8% vs 15%
Google users who saw an AI summary clicked a traditional search result in 8% of visits, compared with 15% of visits when no AI summary appeared.
18%
About one in five Google searches (18%) in Pew's panel produced an AI summary in March 2025.
58% lower
When an AI Overview is present, the top-ranking page's average click-through rate is 58% lower.
61%
Organic click-through rate on queries with an AI Overview fell 61%, from 1.76% in June 2024 to 0.61% in September 2025.
58.5%
58.5% of U.S. Google searches in 2024 ended without a click to any website.
2.5 billion
Google says AI Overviews has more than 2.5 billion monthly active users.
Assistants such as ChatGPT are used mostly to get information and advice, and they send a small but fast-growing share of visits to websites. Local recommendations are part of that shift.
34%
34% of U.S. adults had used ChatGPT by early 2025, roughly double the share in 2023, including 58% of adults under 30.
Nearly 80%
Practical guidance, seeking information and writing together account for nearly 80% of ChatGPT conversations.
1.1 billion
AI platforms sent more than 1.1 billion referral visits to websites in June 2025, up 357% from a year earlier.
693%
Traffic from generative AI tools to U.S. retail websites rose 693.4% year over year in the 2025 holiday season.
If fewer people click, what an answer says about a business matters more, and so does whether the answer draws on that business's own information.
35% / 91%
Brands cited in a Google AI Overview earned 35% more organic clicks and 91% more paid clicks than brands that were not cited.
1%
People clicked a link inside a Google AI summary in 1% of visits to results pages that had one.
Up to 40%
Researchers found that changes to how a source presents its content can raise its visibility in generative engine answers by up to 40%.
In September 2026 Lehvel ran a study on four U.S. businesses: three local service businesses and one national retailer. Lehvel Observatory asked ChatGPT, Claude, Gemini and Perplexity about each one, once by name and once as a customer looking for that kind of business, and checked every factual claim in the answers against a record of the business's facts. Seven of Gemini's eight calls failed in that run, so the results come almost entirely from the other three.
Four businesses is a small sample. The figures describe this study only, and each one is published with its method.
Asked about the business by name
Asked for the best provider of its service
13 of 13 vs 6 of 12
When an AI assistant was asked about a business by name, it named the business in 13 of 13 answers. When it was asked for the best provider of that business's service in its area, it named the business in 6 of 12 answers, and for one business in none.
69 · 2 · 64
Of 135 factual claims the answers made about these businesses, 69 matched the business's recorded facts, 2 contradicted them, and 64 could not be checked because no recorded fact covered them.
4 of 4
Comparing each business's name, address and phone across its website, its structured data, its Google profile, registries and major directories produced a high or critical finding for all four businesses.
3 of 4
Three of the four businesses' own websites contradicted themselves on a material fact, such as founding year, hours, ownership, or shipping and return terms.
Two of the conflicts are worth describing. One assistant gave the wrong suite number in a business’s address. Another described a studio that offers several services as bridal-only, and the phrase it drew on was the business’s own tagline. A wrong answer can start on a page the business controls. That is what Lehvel Observatory checks for.
An assistant can only repeat what it can find, and it finds a business in more places than its website: Google, directories, reviews, registries, social profiles and other sites that mention it. When those places disagree, or leave a question unanswered, the answer a customer gets is a guess. The research above points at four things a business controls: whether its information is complete, consistent across every place it appears, current, and clear about what it does, where and for whom.
None of this needs a special AI format. Google’s guidance is that AI Overviews and AI Mode rest on the same foundations as search: pages that can be crawled and indexed, useful text, and structured data that matches what the page says. There are no additional requirements and no special schema for appearing in them. Microsoft’s guide for retailers, From discovery to influence, makes the same point from another side: keep the same values in every feed, in the page’s structured data and on the page people see, and keep the live site working, because an assistant that sends a customer to a broken page loses the sale however well it understood the business.
That is the work Lehvel does: one record of the business, published everywhere it needs to appear, checked against what search engines and AI assistants actually say, and repaired at the source when they disagree. How the record works
Outside research. We used primary sources: research institutes, platform owners, a working paper, a peer-reviewed benchmark, and vendor studies that publish their sample. A person found each figure on its source page on the date shown. Figures we could not confirm at the source were left out, including widely repeated ones. Retail-only findings are labeled as retail.
Lehvel’s own figures.Each comes from a stored run of Lehvel’s systems, with the calculation, sample, dates and exclusions shown beside it. The businesses are not named, and no figure for a single business is published.
Updates. Figures are rechecked at their source at least every 120days, or marked “due for review” on the page. The next Lehvel sample will be larger and will be published beside this one.
What the AI platforms your customers use say about a business, checked against its facts.
One client, named. Every change counted.
Lehvel Labs
Lehvel Labs
Lehvel Labs
One note a week: one job we took off an owner's plate, and what came of it.
My posts reach 17,000+ followers across our channels. The weekly email is called Upstream.
Weekly. Unsubscribe anytime. Privacy