On this page
We studied 1,000 AI answers to common real-estate hiring questions in 50 U.S. markets. Specific agents or teams appeared in 88.8% of the answers, while the full lists returned by pairs of tools overlapped by 7.31% on average. This article describes the study design, the observed differences between tools and market sizes, and the limits of these findings.
The results show what four AI tools returned on October 11–12, 2026. They do not rate or recommend any person, team, or brokerage, and they do not predict future answers. AI answers are one part of a buyer's or seller's research process. We report the pattern so agents can assess the information a prospective client may encounter.
These figures describe this study's October 11–12, 2026 run, not every future answer.
How we ran the study
We studied 50 U.S. markets: 12 large markets, 20 mid-sized markets, and 18 small markets. The list covered all four U.S. Census regions: 4 markets in the Northeast, 9 in the Midwest, 20 in the South, and 17 in the West. We used this mix to compare large metros with smaller local markets. It was not a ranking of markets or a sample of every place in the country.
For every market, we asked each tool the same five questions. We retained the placeholders below so that the wording is clear and reproducible across markets:
- Who is the best real estate agent in {City}, {ST}?
- Who should I hire to sell my home in {City}, {ST}?
- Top realtors in {City}, {ST}
- Best listing agent in {City}, {ST} to sell a house
- Which real estate agent should I use to buy a home in {City}, {ST}?
We asked ChatGPT, Gemini, Perplexity, and Google's AI Mode from October 11 to October 12, 2026. The design produced 1,000 answers: 50 markets × 5 questions × 4 tools. All 1,000 planned answers returned. We used the same prompt set for each tool and market, but the study does not establish why a tool selected a particular answer.
We counted a named person separately from a named team or brokerage when an answer gave a specific name. We counted each source type once per answer, even when several links came from that type. For each city and question, we checked whether two tools named anyone in common; to compare full lists, we divided shared names by every unique name in both lists. We also recorded first-name consistency, meaning that at least two tools began with the same named person or organization.
Answers can vary by person, day, and location. Our automatic counting is conservative: it undercounts rather than invents names or source matches. It can miss an unusually written name and cannot identify why a tool chose a result. The data describes this run, not every future answer.
How many agents AI names
Across all answers, AI named an average of 2.68 individual agents and 1.96 teams or brokerages per answer. Together, those measures equal 4.63 named people or organizations in the average response. Some answers contained one or two suggestions, while others contained longer lists of individuals, teams, and organizations. This difference in list length is relevant when comparing tools, but it does not make a longer list more useful.
In 88.8% of all returned answers, at least one specific agent or team appeared. This finding does not mean that the named person was the best fit for every buyer or seller. It shows that direct hiring questions regularly produced names rather than only general guidance. A hiring answer may therefore contribute to a prospect's initial shortlist before that person visits an agent's website.
The remaining 11.2% of answers did not contain a specific name we could count. Some of these responses focused on choosing criteria, and some offered broad guidance. That variation is a reason not to treat an AI tool as a directory with stable results. Question wording and the particular tool both matter.
Where the tools agree and differ
Agreement requires more than one measure. In 82.8% of city-and-question combinations, at least two tools had one named person or organization in common. Their full named lists, however, had only 7.31% average pairwise overlap. The tools often shared one name somewhere in an answer but usually did not produce the same shortlist.
The first position was less consistent. In only 35.2% of city-and-question combinations did at least two tools place the same person or organization first. A screenshot of one answer is therefore weak evidence of a stable pattern. A different tool, question, or day may return a different leading name.
We recommend measuring several questions across several tools. Review the sources as well: missing, unclear, or contradictory public information gives an agent a specific issue to examine. Appearing in one tool but not another is a signal for further review, not a promise or a failure. This distinction matters because the study observes answers; it does not identify their causes.
What sources AI cites
Agent or team websites were the most common source type, appearing in 71.6% of answers. Zillow appeared in 56.9%, FastExpert in 43.9%, other directories in 30.8%, Homes.com in 24.0%, and HomeLight in 21.7%. Realtor.com appeared in 18.6% of answers. Google Business sources appeared in 14.8%, local news in 11.8%, brokerage sites in 9.5%, and Yelp in 6.9%.
These shares describe answers with at least one source from each type. They do not sum to 100% because an answer can cite several source types. They also do not establish that a source caused a recommendation. A source may support a fact, provide a list, or offer a reader additional context.
In practice, start with the information an agent controls: a clear website, a specific local focus, real service details, and a simple contact path. Then check the major profiles and directories that repeat those details. Add independent local evidence a buyer or seller can evaluate. The purpose is a record people can understand, not a way to manipulate one answer.
How market size relates to results
The basic naming pattern held across market sizes. Specific names appeared in 89.2% of small-market answers, 88.3% of mid-sized-market answers, and 89.2% of large-market answers. The average number of named people or organizations was 4.34 in small markets, 4.81 in mid-sized markets, and 4.78 in large markets. These differences do not show a market-size effect on whether an answer named someone.
Small markets had the highest average list overlap at 9.41%. Mid-sized markets averaged 6.49%, and large markets averaged 5.48%. This may reflect smaller pools or more repeated local sources, but the study does not show why the pattern occurred. First-name consistency also declined as markets became larger: 38.9% in small markets, 35.0% in mid-sized markets, and 30.0% in large markets.
Agent or team websites appeared in 73.1% of small-market answers, 68.5% of mid-sized-market answers, and 74.6% of large-market answers. Zillow appeared in 62.2%, 55.8%, and 50.8%, respectively. The shares differ, but the evidence is within the limits of one study run and does not make a single site or answer sufficient to explain visibility. We have not tested the underlying causes of these differences.
Per-tool differences
The four tools returned every planned answer, but their outputs differed. Google's AI Mode named a specific person or organization in 85.2% of its answers and averaged 4.98 names per answer. ChatGPT named someone in 88.0% and averaged 3.12 names. Gemini did so in 98.8% and averaged 8.09 names. Perplexity named someone in 83.2% and averaged 2.35 names.
Their source patterns differed as well. Agent or team websites appeared in 56.4% of Google's AI Mode answers, 75.2% of ChatGPT answers, 64.0% of Gemini answers, and 90.8% of Perplexity answers. Zillow appeared in 38.4%, 77.6%, 30.0%, and 81.6% of those answers. A source type can be common in the aggregate while appearing much less often in an individual tool.
We would not use these differences to name a winning tool. Longer lists are not necessarily better answers, and source links do not fully explain a response. This matters because an agent needs a broad, accurate public record. A review limited to one tool cannot establish a stable pattern across tools or questions.
What we recommend
AI answers are another place where buyers and sellers may begin comparing agents. In our study, names appeared often, but the answer varied across tools and questions. We recommend building for the person conducting the research: explain where you work, the client situations you handle, how you work, and the evidence supporting those statements. Keep the same basics accurate across a website and the places people use to verify them.
We also examined whether local income or home prices change these patterns; they mostly do not. Recasi Local applies this approach across an agent's public web presence. We identify market questions, support the public record with evidence, and clarify local expertise. Learn more about AI SEO for real estate agents or review Recasi Local pricing.
A review limited to one tool cannot establish a stable pattern.
Do not treat this report as a promise that any tool will name an agent. Use it to review what a buyer or seller can actually find today. A clear, honest public record gives people better material for a decision, whether they encounter it through AI, search, a referral, or an agent's site.
Results by market
Northeast
Midwest
South
West
Frequently asked questions
Does this study show the best agent in each market?
No. We did not rank people or organizations. We measured how often four AI tools named a specific person or organization and the sources that appeared with their answers.
Why do the tools name different people?
Each tool can use different information and can respond differently to the same wording. Answers can also change by person, day, and location. A single answer is therefore not a stable ranking.
What does 7.31% overlap mean?
It is the average share of names two tools had in common for the same city and question. The shared names were usually a small portion of all names on both lists. It measures list overlap, not answer quality.
Do source links prove why AI recommended someone?
No. A source can support a fact or provide more context. The links show what appeared with an answer; they do not reveal the full reason a tool chose a name.
What should an agent do with these results?
Review the information buyers and sellers can find about you. Make your website, profiles, and local evidence accurate, specific, and consistent. Then measure more than one question and more than one tool.