Skip to content
AEO Free test

Study · October 27, 2026

What Claude read to recommend local businesses in 10 US cities

We asked Claude, with web search on, 80 questions like "What is the best dentist in Denver?": 8 kinds of local business in 10 US cities. We saved every answer, every search and every page it read.

Valor Digital, October 27, 2026 · Open data in CSV and JSON (CC BY 4.0) · Brazilian half, in Portuguese

Main numbers

More answers read Yelp than any business's own website

  • Yelp was read in 74 of the 80 answers and cited as a source in 40.
  • Business websites were read in 23 of 80 answers. They were 4.5% of the pages read (37 of 825) and 5.7% of the sources cited (16 of 283).
  • None of the 10 restaurant answers read a restaurant's website. In Brazil, none of the 10 either: 0 of 20 across the whole study.
  • 80 of 80 answers read at least one directory, list or review site. Those pages were 72.5% of the 825 pages read and 66.4% of the 283 sources cited.
  • 79 of 80 answers named businesses or professionals, 8.9 names per answer on average.
  • 79 of 80 answers opened with a caveat, and 78 of those named businesses anyway. 30 opened with the same words: "There's no single objective 'best'".
  • No answer read a Google or Google Maps page. Social media barely showed up: 1 of 80 answers read a social page.

Source for every number on this page: Valor Digital, 80 questions to Claude Haiku 5.5 with web search on October 27, 2026. The counts are in the open data.

Where the answers came from

What kind of site the AI read and cited

Pages read are the results the search returned to the model. Sources cited are the ones it linked to parts of the answer, plus links written in the text.

Answers that read or cited each type of site, and each type's share of pages read and sources cited
Type of siteAnswers (of 80)Pages read (825)Sources cited (283)
Directories, lists and review sites8072.5%66.4%
News and blogs6016.6%24.4%
Business websites234.5%5.7%
Other (Wikipedia, jobs, government, associations)366.3%3.5%
Social networks10.1%0%
Maps (Google Maps and others)00%0%

Top directories

The 8 directories, lists and review sites read most often

Business websites stay unnamed: our rule is not to expose the small businesses that show up in the answers. Directories and platforms are named.

Answers (of 80) in which each site was read and in which it was cited as a source
SiteDomainRead inCited in
Yelpyelp.com7440
Tripadvisortripadvisor.com150
Trust Analyticatrustanalytica.org136
Clutchclutch.co116
Best Lawyersbestlawyers.com97
Best Law Firmsbestlawfirms.com95
OpenTableopentable.com94
HomeLighthomelight.com83

Wikipedia, which is not a directory, was read in 34 answers and cited in 5. Across the US half, the AI read pages from 212 different domains.

By business type

How it varied by business type

Each business type had 10 answers, one per city. That is too few to compare types with each other; the table shows what happened, not a trend.

By business type: answers that named businesses, average names per answer and answers that read any business's website
Business typeNamed businessesNames per answerRead a business website
dentist10 of 109.12 of 10
accountant10 of 1011.21 of 10
lawyer9 of 106.08 of 10
med spa10 of 107.57 of 10
gym10 of 108.51 of 10
real estate agency10 of 1011.63 of 10
pet store10 of 108.21 of 10
restaurant10 of 109.30 of 10

Method

One question per business type and city, no system prompt

On October 27, 2026, from 14:01 UTC to 14:08 UTC, we asked one question for each business type in each city, written the way a customer would: "What is the best dentist in Denver?", "What is the best gym in Austin?". The message was the question alone.

  • Cities: New York, Los Angeles, Chicago, Houston, Phoenix, Austin, Denver, Seattle, Miami and Boston.
  • Business types: dentist, accountant, lawyer, med spa, gym, real estate agency, pet store and restaurant.
  • Assistant: Claude Haiku 5.5 (claude-haiku-5-5) through the Anthropic API, with Anthropic's server-side web search tool, up to 2 searches per question and the approximate location set to the city in the question. The model ran 85 searches: 75 answers with one and 5 with two.
  • Businesses named: a second request to the same model, without search, separated business names from platform names in each answer. Only names written in the answer count, compared without accents or capitals and as whole words. Social networks, directories, government bodies and newspapers do not count as businesses.
  • Site type: each domain read was reviewed with the title and address of its pages and put on a list. The review was done by a Valor Digital AI agent, domain by domain, without human review. A "top 10" list published on a blog or a newspaper counts as news or blog; the same list on a site that lives off listing businesses counts as a directory.
  • Caveat: we looked for reviewed phrases near the start of the text, such as "there's no single best" or "it depends on what you need".

We keep the full record of every answer: the request sent, the API response, the time, the searches, the pages read and the citations. We do not publish the answer texts because they contain business names.

Limits

What this study does not show

  • Only Claude was asked. ChatGPT, Gemini, Perplexity and Google's AI answers were left out.
  • The model is Haiku 5.5 through the API, with up to 2 searches. The Claude app can use another model, search more and answer differently.
  • One answer per question, at a single time. Answers change from one time to the next; this study does not measure how much.
  • Claude's search is not Google's: the absence of Google Maps applies to this assistant, not to Google's AI.
  • 10 answers per business type is too few to compare types with each other.
  • Sites were classified by an AI agent, without human review.

Open data

Download the data and cite the source

The files are under the Creative Commons Attribution 4.0 license (CC BY 4.0): you can copy, adapt and republish them, including commercially, as long as you credit the source. No business is identified in the files.

CSV

answers-us.csv

One row per answer: city, business type, question, time, searches, whether it named businesses and how many, whether it opened with a caveat, and the pages read and sources cited of each type of site.

JSON

study-us.json

The totals on this page: site types, top directories, business types, searches, method and limits.

How to cite

Valor Digital. What Claude read to recommend local businesses in 10 US cities. Study of 80 questions to Claude Haiku 5.5 with web search, run on October 27, 2026. Available at: https://aeo.valor.digital/en/study/. License CC BY 4.0.

The other half of the study, with 80 questions in 10 Brazilian cities, is in Portuguese: O que o Claude leu para indicar empresas em 10 cidades do Brasil.

Free test

What about your business, in your city?

We ask the same kind of questions with your business type and city and show whether you show up.