HOW REAL ESTATE BROKERAGES GET RECOMMENDED BY CHATGPT: SOURCE DATA
Pluspoint, November 2026

If you use this data, please credit Pluspoint and link to the study: https://www.pluspoint.io/blog/chatgpt-real-estate-brokerage-recommendations


1. WHAT THIS IS

This is the data behind the study "How Real Estate Brokerages Get Recommended by ChatGPT: 500 Answers From 50 US Cities". It holds every question we asked, every answer ChatGPT gave, how we coded the companies it named, and the Google Maps data we compared them with. With it you can check what was asked and answered, and rebuild the article's main numbers.

The same tables come twice: as sheets in data.xlsx and as UTF-8 CSV files in the csv folder. The first sheet of data.xlsx lists the tables.


2. HOW THE ANSWERS WERE COLLECTED

This is the method as the article states it.

We ran the test on 9 and 11 November 2026 in the 50 largest US cities, using ChatGPT on the web with web search. Every question named the city, and the only location we set was the United States.

We asked three questions about companies twice each, 300 answers in all:
  C1  "What are the best real estate companies in [city]?"
  C2  "I'm selling my house in [city]. Which real estate company should I list with?"
  B   "Which real estate brokerage in [city] has the best reviews?"
The first two are the best-company questions. B is the best-reviews question.

We added 200 answers about agents from the same cities:
  A1  "Who are the best real estate agents in [city]?"
  A2  "I'm buying a home in [city]. Which real estate agent should I hire?"

We sorted every company named by type, looked up its listings on Google Maps and compared them with the top 20 Maps results for "real estate agency" in each city.

Dates. C1 and C2 were asked twice on 11 November. B was asked once on 9 November and once on 11 November. A1 and A2 were asked twice on 9 November. The exact question text for each answer, with the city and state filled in, is in the answers table. Every answer reports its model as "gpt-6".


3. THE TABLES

answers (500 rows)
  One row per ChatGPT answer: 100 to each of C1, C2 and B, and 200 to A1 and A2.

names (3,506 rows)
  One row per business or person ChatGPT named in an answer, with the study's coding.

maps_top20 (1,000 rows)
  The Google Maps top 20 for "real estate agency" in each city, collected on 11 November. Names are replaced by ids.

companies (620 rows)
  One row per company or team named as a list pick on a best-company question in a city. Names are replaced by ids.

profiles (4,493 rows)
  The Google Maps listings that carry a named company's name in its city: offices, agents and teams. This is the data behind "where the reviews sit". Names are replaced by ids.

city_summary (50 rows)
  The numbers the article's city lookup shows, one row per city.

national (22 rows)
  The national numbers the article's city lookup shows.

figures (54 rows)
  Every figure the article states, with its exact value, its n, a definition and the table and columns that rebuild it.


4. DATA DICTIONARY

Shared values
  question_code: C1, C2, B, A1 or A2, as in section 2.
  question_group: "best companies" (C1 and C2), "best reviews" (B) or "agents" (A1 and A2).
  ask: 1 for the first time a question was asked in a city, 2 for the second time.
  yes / no: a flag. A blank cell means the value does not apply or was not collected.

answers
  answer_id: city, question code and ask, for example "boston-C1-r2". It links to names.
  question_code, question_group, ask: see Shared values.
  collected_date: the date of the answer (UTC).
  collected_at_utc: the date and time of the answer (UTC).
  city, state: the city and state named in the question.
  question_text: the exact question sent to ChatGPT.
  model_reported: the model name the answer reported.
  answer_text: the full answer, as ChatGPT wrote it, in Markdown. Masked as in section 5.
  web_search_count: the number of web searches ChatGPT ran for the answer.
  web_searches: the text of those searches, separated by " | ".
  cited_url_count: the number of links ChatGPT cited in the answer.
  cited_urls: the cited links, one per line. Masked as in section 5.
  cited_domains: the domains of the cited links, each once, separated by " | ".
  cited_url_types: the study's source type for each cited link, one per line, in the same order as cited_urls. Types: ranking_matching (rankings and agent-matching sites), brokerage_site, portal, review_platform, bestof_directory, google_maps, news_business_journal, news_magazine, license_registry, association, forum, wikipedia, other.
  cited_url_page_types: for links to a brokerage, team or agent website, the kind of page, one per line in the same order (homepage, office_or_location, agent_or_team_page, about_awards_press, market_report, reviews_page, blog, listings, service_page, other_page). Blank for other links.

names
  answer_id, question_code, question_group, ask, city, state: the answer the name comes from.
  position: the name's number in the answer's list or table. Blank for names mentioned only in prose.
  how_named: "list" (a numbered or bulleted pick), "table" (a row in a table) or "prose only" (mentioned in the text, not as a pick).
  in_study_counts: "yes" for list and table names. The study's figures count these rows only.
  name_as_written: the name as ChatGPT wrote it.
  name_key: the study's normalized form of the name (lower case, generic words such as "realty" and "group" removed). The study uses it to tell whether two answers name the same company in the same city.
  type: the study's coding. "franchise brand" (an office or company under a national franchise system), "national brand" (a national brokerage that is not a franchise system), "independent brokerage", "team", "agent", "commercial firm". The article's "big brands" are franchise and national brands.
  brand: for franchise and national brands, and for teams and agents tied to one, the brand.
  named_as_brand_only: for franchise and national brands, "yes" if the answer named only the brand, "no" if it named a franchisee company or office.
  brokerage_as_written: for agents and teams, the brokerage named beside them, as written.
  quoted_review_count: the largest review count the answer quotes for this name. These are counts ChatGPT quoted, not verified counts.
  quoted_review_platform: the platform named next to that count, or "not stated".
  quoted_rating: the first star rating the answer quotes for this name.
  quoted_sales_figure: "yes" if the text about this name quotes sales volume, transaction sides, homes sold or market share.
  gave_street_address: "yes" if the text about this name gives a street address.
  street_address_as_written: that street address.
  answer_excerpt: the text about this name in the answer, up to 600 characters. Masked as in section 5.
  maps_listing_review_count, maps_listing_rating, maps_listing_category: the Google reviews, rating and Google category of the Maps listing the study matched to this name. For names on the best-company questions, the listing is the office the answer gave an address for, else the only matching listing, else the most-reviewed matching office in the city. Blank when the study did not look the name up or found no match.
  maps_lookup: "11 Nov lookup" (best-company names, and best-reviews names that were also named on a best-company question in the same city) or "9 Nov lookup" (agent-question names and names from the 9 November best-reviews answers).

maps_top20
  city, state: the city searched.
  rank: the listing's position in the Google Maps results, 1 to 20.
  listing_id: a stable id in place of the listing's name. The same listing has the same id in profiles.
  google_category: the listing's Google category.
  brand_type: "franchise brand" or "national brand" if the listing's name contains one of the brands the study codes, else "not a national brand".
  review_count, rating: the listing's Google reviews and star rating.
  named_for_best_reviews: "yes" if its name matches a company or team ChatGPT named on the best-reviews question in that city.
  named_for_best_companies: the same for the best-company questions.

companies
  company_id: a stable id in place of the company's name. It links to profiles.
  city, state: the city where it was named.
  type: the study's coding, as in names (the type it was coded as most often).
  asks_named_best_companies_q1: in how many of the two C1 answers in that city it was a list pick (0, 1 or 2).
  asks_named_best_companies_q2: the same for C2.
  maps_matched: "yes" if the study matched it to a Google Maps listing.
  chosen_listing_review_count, chosen_listing_rating: Google reviews and rating of the listing the study chose for it (same rule as in names).
  office_listings_in_city: the Maps listings carrying its name that the study classed as offices, counted once per street address.
  agent_listings_in_city, team_listings_in_city: the listings classed as an agent's or a team's.
  maps_results_returned: how many results the Maps search for its name returned (at most 20).

profiles
  company_id, company_type: the company, as in companies.
  listing_id: a stable id in place of the listing's name. A listing that carries two companies' names appears once under each.
  listing_kind: "office", "agent" or "team", by the study's rule. A title that is only the company's name, or the name plus a place or office word, is an office. Two or more personal-name words besides the company's name make it an agent. Words such as "team", "group" or "partners" make it a team.
  office_counted_once_per_address: for offices, "yes" for the one listing kept per street address (the most-reviewed), "no" for the others. The star figures use the "yes" rows.
  city, state: the city of the search.
  google_category: the listing's Google category.
  review_count, rating: the listing's Google reviews and star rating.
  stars_1 to stars_5: the number of one-star to five-star reviews on the listing. Blank where Google gave no split, mostly listings with no reviews.

city_summary
  city: the city.
  best_company_answers: best-company answers in the city (4).
  best_company_names: names counted in those answers.
  best_company_share_franchise_pct, _national_pct, _independent_pct, _team_pct, _other_pct: shares of those names by type, in percent. "other" is agents and commercial firms.
  best_reviews_answers: best-reviews answers in the city (2).
  best_reviews_names: names counted in those answers.
  best_reviews_count_franchise, _national, _independent, _team, _other: those names by type, as counts.
  top_cited_domains_answers_of_4: up to three domains cited most in the city's best-company answers, each with the number of answers (out of 4) that cite it. As on the article's page, three AI-built directories are shown as "an AI-built directory", two website hosts as "a brokerage's website", and the agent review site as "an agent review site".
  realtrends_brokerage_ranking_cited: "yes" if a best-company answer cited a RealTrends ranking of brokerages.
  business_journal_cited: "yes" if a best-company answer cited a business journal.
  named_companies_median_google_reviews: the median Google reviews of the listings matched to the companies named on the best-company questions.
  named_companies_with_listing: how many named companies that median rests on.
  median_is_thin_under_10: "yes" if that is fewer than 10.
  maps_top20_median_google_reviews: the median Google reviews of the city's Maps top 20.
  names_in_first_ask: company and team names in the first ask of C1 and C2.
  names_named_again_in_second_ask: how many of those were named again in the second ask.

national
  metric: what the row measures.
  value: the value.
  unit: the unit of the value.
  n: what a percentage is out of.

figures
  figure_id: an id for the figure.
  where_in_article: the part of the article that states it.
  article_text: the article's words, shortened.
  article_value: the value as the article states it.
  exact_value: the unrounded value from the study.
  n: what the value is out of.
  definition: how the value is computed.
  reproduce_with: the table and columns that rebuild it, or why the package cannot.
  rebuilt_in_our_check: "yes" if we rebuilt the value from these tables alone and it matched the article.


5. CHANGES WE MADE TO THE RAW DATA

We changed the raw data in three ways and in no other.

Review software vendors are not named. The study does not name companies that sell review, listings or reputation software. Wherever a cited link, a domain or the answer text names one of them, we replaced it with "[review software vendor]". This applies in answers (answer_text, cited_urls, cited_domains) and in names (answer_excerpt, quoted_review_platform). One real estate review site is shown the way the article shows it: we replaced its name and links with "[agent review site]". Masking was on for this file.
  answers: 705 vendor mentions and 402 review-site mentions replaced.
  names: 154 vendor mentions and 125 review-site mentions replaced.
  Other tables: none needed.
The source types in cited_url_types and the counts in the figures are unchanged.

Google Maps listings carry ids, not names. In maps_top20, companies and profiles, we replaced every listing and company name with a stable id (620 company ids, 5,185 listing ids). This keeps the star data from pointing at a named office. The names table keeps the names ChatGPT wrote, since they are in the answers anyway. It carries the matched listing's review count and rating but no listing id, so it does not link a name to a row in maps_top20 or profiles. Anonymizing was on for this file.

Image links are removed. Answers embed photos through long image links on ChatGPT's image server. They carry no study data and look like access tokens, so we replaced each with "[image link removed]" in answer_text (2,456 links). The image's caption text is kept.

What we left out. The pages ChatGPT's searches returned but the answer did not cite, the raw files from the data provider, the request files and our working notes. The figures that need a name linked to a listing (marked in the figures table) cannot be rebuilt from this package, because the ids rule above withholds that link.


6. LIMITS

This is one assistant over a few days, asked the way a buyer who names the city would ask. A buyer who shares a precise location, or asks for companies nearby without naming a city, may see a different office or a different list.

The Maps lookups stop at 20 results and count only profiles that carry the company's name, so the agents' share of reviews is a floor.

On the best-reviews question, the counts and ratings are the ones ChatGPT quoted, and about half of them name no platform. The other review figures come from our Google Maps lookups.

Three checks rest on a few dozen cases: the office ChatGPT pointed to, the review counts it quoted for companies with several offices, and the teams it named under a brand.

The coding of types and of office, agent and team listings is rule-based. A hand check agreed with it in 36 of 40 names. The weakest line is between a team and an independent brokerage.

All of it shows who ChatGPT names and what it quotes. It shows links, not causes, and nothing here proves that more reviews will get a company named. ChatGPT gives a different answer each time, so a new run will not match these answers word for word.

If you use this data, please credit Pluspoint and link to the study: https://www.pluspoint.io/blog/chatgpt-real-estate-brokerage-recommendations
