Two rounds, more than 5,200 AI answers: about 19 in 20 HAR members never named

What we found

In Round 2 (13 to 16 October 2026), ChatGPT, Perplexity and Gemini gave 3,368 counted answers to 1,130 Houston-area real estate questions and named about 2,527 real firms. Against HAR's 50,000-plus members that is 5.1%, so 94.9% were never named. Round 1, on four surfaces in September 2026, found 94.0% across 1,917 counted answers.

Five findings carry the report:

• 94.9% of HAR members were never named. About 2,527 real firms were named across the Houston neighbourhood and suburb studies, which is 5.1% of 50,000-plus members. The figure is likely a floor, because Perplexity's short answers are under-extracted.
• The engines rarely name the same firm. On 381 neighbourhood questions where all three engines named someone, they named a common firm on 6.8%.
• 12 of 23 suburbs have one dominant local firm. In each, a single local firm appears in 40% or more of the answers.
• Most named firms stay in one place. 70.3% of real firms named in the neighbourhood study appear in one of 15 neighbourhoods, and 73.2% in the suburb study appear in one of 23 suburbs.
• The Tampa Bay pilot is more local still. 85.9% of the 184 real firms named appear in one of 3 towns. With three towns it cannot be compared with Houston.

Round 1 and Round 2 side by side

Round 1 put 500 questions to four AI surfaces on 18 to 22 September 2026 and counted 1,917 of 2,000 answers. Round 2 put 1,190 questions to three engines on 13 to 16 October 2026 and counted 3,548 of 3,570. The surface counts, question sets and name extractors differ, so the rounds are read side by side and no trend is claimed.

What changed between the rounds, so nobody compares the wrong numbers:

• Surfaces. Round 1 used ChatGPT, Perplexity, the Gemini web app and Gemini through its API. Round 2 used ChatGPT, Perplexity and Gemini, each through its public web app.
• Questions. Round 1 covered 500 questions in 40 areas. Round 2 covered 1,190 questions in 41 areas, split into neighbourhood, suburb and Tampa Bay studies.
• Extraction and labels. Round 1 used its own extractor and 394 names labelled blind in September. Round 2 used a new extractor and 765 names labelled by an AI reviewer, with a person re-checking a sample.

One sentence holds across both: about 7 in 10 to 8 in 10 named firms appear in one area only, with published figures running from 70.3% to 82.6%. Because the extractors differ, we do not rank the rounds against each other. The full Round 1 write-up is at the 500-Query Houston study.

Round 1

September 2026, 4 surfaces, 2,000 answers, 1,917 counted

Round 2

13 October 2026 to 16 October 2026, 3 engines, 3,390 Houston answers, 3,368 counted; a separate 180-answer Tampa Bay pilot is not in the totals

Removed

22 answers, because they did not match the question asked

Licence

CC BY 4.0, reuse with attribution

Inside this edition

1. 01

   Houston neighbourhoodsShare of 381 questions where a firm was named by all three engines.
   Read the report

   6.8%
2. 02

   Houston suburbsSuburbs where one local firm appears in 40% or more of the town's answers.

   12 of 23
3. 03

   Tampa Bay pilotShare of firms named in only one of 3 towns. A pilot, not comparable with Houston.

   85.9%

Houston neighbourhoods: three engines, rarely one firm

On 381 neighbourhood questions where ChatGPT, Perplexity and Gemini all named at least one firm (13 to 16 October 2026), the three engines named a common firm on only 6.8%. On 37.3% of those questions no two engines shared any firm. The study covers 15 areas, 450 questions and 1,339 counted answers.

Of the real firms named, 70.3% appear in a single neighbourhood and 70.5% on a single engine. Concentration is moderate: the 117 most-named firms, the top 10%, take 59.7% of mentions, and the leading local firm in the median neighbourhood appears in 27.8% of answers.

The detail is on the neighbourhoods page.

Houston suburbs: one local firm often dominates

In 12 of 23 Houston suburbs, one local firm appeared in 40% or more of AI answers (680 questions, 2,029 counted answers, 13 to 16 October 2026). The median suburb leader appeared in 40.4% of answers, from 16.7% in Jersey Village to 61.4% in Missouri City. National brands were set aside when choosing each leader.

Suburbs are more settled than neighbourhoods. All three engines named the suburb leader in 19 of 23 suburbs, and on 590 questions where every engine named someone, a common firm appeared on 19.7%, against 6.8% in neighbourhoods. Of the real firms named, 73.2% appear in one suburb and 55.4% were named exactly once.

Agent referral networks supply 26.8% of the citations in the suburb study and portals 39.7%.

Tampa Bay pilot: three towns, not comparable

In the Tampa Bay pilot (Brandon, Clearwater and Lithia, Florida; 60 questions, 180 answers, 14 to 16 October 2026), 85.9% of the 184 real firms named appeared in one of the 3 towns. Every extracted name was labelled, so these are census values. With three towns, the pilot cannot be compared with the 15 and 23 areas in Houston.

68.5% of the real firms were named by one engine only. The leading firm in Lithia appeared in 58.3% of that town's answers, and the median town leader in 36.7%. We ran the pilot to test the method outside Texas, and it is not evidence about Florida as a whole.

How many Houston Realtors are ever named

Across the Houston neighbourhood and suburb studies (13 to 16 October 2026), ChatGPT, Perplexity and Gemini named about 2,527 real firms, 5.1% of HAR's 50,000-plus members, leaving 94.9% never named. The count comes from 1,339 and 2,029 counted answers and a labelled sample of names. It is likely a floor.

Only 10.2% of names appear in both Houston studies, which is why the two counts are combined and not added. The count is a floor for two reasons. Perplexity's answers run about 390 characters at the median against 2,727 for ChatGPT, so fewer of its names are extracted. And one run per question samples the engines once.

The comparison is approximate in one more way. HAR counts individual members, while the engines also name teams and brokerages, so 5.1% is an approximate ratio of named firms to members. There is no list of firms that were never named, and we will not publish one.

How the research was run

CiteVantage wrote 1,190 questions with no brand names and put each once to ChatGPT, Perplexity and Gemini between 13 to 16 October 2026, giving 3,570 answers, of which 3,548 were counted. 22 answers were removed because they did not match the question asked. Names are extracted by published rules, and the row-level data is released under CC BY 4.0.

Design. 41 areas were covered. Each of the 37 Houston-area areas carries 30 conversational wordings (20 core, 10 extended), and Richmond, Texas and the three Florida towns carry the 20 core wordings. Every question names its state.

Fieldwork. Each engine was used through its public web app, in a real browser, with a new conversation for every question. The model behind each Gemini answer is not recorded.

Removals. 22 answers were removed from every count and every published file because they did not match the question asked: 11 in neighbourhoods, 11 in suburbs and none in Tampa Bay. No answer was about the wrong state.

Extraction. Names come from list items, table rows, business-profile cards, person names and firm names ending in words such as Team, Group or Realty. Each name must pass a plausibility check, and variants such as "Jill Smith" and "The Jill Smith Team" count once. A name is counted at most once per answer, wherever it sits in the text.

Sensitivity. A strict filter removes 18.6% of extracted names in neighbourhoods, and every structural figure is published beside its strict-filter value.

Limits of this research

Round 2 asked each of 1,190 questions once per engine in 13 to 16 October 2026, so a repeat would give different lists. Names are extracted by rules and a labelled sample estimates how many are real firms. The figures describe what three engines returned in that window, and say nothing about how many people asked or what any of them did next.

• One run per question. Answers move between runs. This is a snapshot, and a specific firm's position may have shifted since.
• Heuristic extraction. The rules mistake some phrases for names and miss some real ones. About 4.7% of sampled neighbourhood names were not businesses.
• Perplexity is under-extracted. Its short answers carry fewer names, so counts that pool engines lean low.
• Named is not recommended. An engine naming a firm is not an endorsement, and the research does not measure leads, traffic or revenue.
• The denominator is approximate. HAR members are individuals, and the engines also name teams and brokerages.
• Tampa Bay is a pilot. Three towns and 60 questions do not support a comparison with Houston.

Does the same question get the same answer?

Mostly not, as far as the firms named go. We put a random 100 of the October questions to ChatGPT and Gemini again, twice. 63 ChatGPT questions and 61 Gemini questions have all three clean answers, and only those are counted here.

For the average ChatGPT question, between 26% and 38% of the firms that either of two runs named were named by both runs. For Gemini the range is 10% to 15%. Across the same pairs, 3% to 19% of ChatGPT questions and 39% to 53% of Gemini questions came back with no firm in common.

ChatGPT, October and repeat 1

27% average overlap; no firm in common on 18% of questions; the same firm named first on 41%

ChatGPT, repeat 1 and repeat 2

38% average overlap; no firm in common on 3% of questions; the same firm named first on 46%

ChatGPT, October and repeat 2

26% average overlap; no firm in common on 19% of questions; the same firm named first on 33%

Gemini, October and repeat 1

10% average overlap; no firm in common on 53% of questions; the same firm named first on 18%

Gemini, repeat 1 and repeat 2

13% average overlap; no firm in common on 49% of questions; the same firm named first on 16%

Gemini, October and repeat 2

15% average overlap; no firm in common on 39% of questions; the same firm named first on 13%

The repeats were run on 29 and 30 October, with a few of the second repeat's answers added on 5 November, so they sit about two weeks after the October answers. ChatGPT's two repeats overlapped more with each other (38%) than either did with October (27% and 26%), which suggests some change over time on top of ordinary run-to-run variation. Gemini shows no such pattern.

One answer is therefore a weak guide to which firms an engine will name for a question, which is why this report counts firms across 3,368 answers and not within any one. It also means asking more questions, or the same ones again, would name more firms, so the never-named share describes the questions we asked.

Limits: Perplexity was not repeated. Both engines stopped before all 100 questions had two repeats (ChatGPT's temporary chat began reading memory, and Gemini returned rate-limit errors), so we kept only questions with three clean answers. 2 ChatGPT questions were dropped because an answer did not match the question asked. Every run used the same setup as October, so these figures describe stability on that setup, and two repeats cannot fully separate run-to-run variation from change over time.

Download houston-repeat-run-names.csv (the firms each run named) and repeat_run_score.py, which recomputes every figure in this section from that file.

How names were labelled

An AI reviewer labelled 765 sampled names across the three studies as real firm, other business or not a business, without seeing how often each name was returned. A person then re-checked 31 names from a 100-name draw of those labels and agreed with every one, and a second AI pass with web lookups for doubtful names re-labelled all 100 and matched the first on 99. Every figure about firms uses labelled samples, and every figure about what engines returned uses names as extracted.

The labelled counts are 1,118 real firms in neighbourhoods (95% interval 1,088 to 1,144) and the Tampa Bay census of 184. Samples were stratified by whether a name appeared in one area or several, then weighted back to the full list with a 5,000-resample bootstrap. We call these labels, not hand checks, because a person re-checked only a sample. All labels are published so any of them can be disputed.

How this compares with other research

Three outside studies frame this report. SparkToro found across 2,961 runs that the same brand list came back less than 1 in 100 times. BrightEdge found the same brands on ChatGPT and Google's two AI surfaces for 17% of queries. Pitenin's census of 4,776 Bali venues found 85.6% never recommended. None measures Houston Realtors.

• SparkToro and Gumshoe ran 12 prompts 2,961 times on ChatGPT, Claude and Google's AI surfaces. The same brand list came back less than 1 in 100 times and the same order about 1 in 1,000. The authors say it is not peer reviewed. It supports our one-run limit.
• BrightEdge reported that platforms disagreed on brand recommendations for 61.9% of queries, with the same brands on all three for 17%. It is vendor data with no query count published. Our 6.8% counts a common firm, not an identical list, so the two are not directly comparable.
• Pitenin, arXiv 2608.07069 (submitted 7 August 2026) used 4,776 venues as a denominator and 2,208 answers to 96 queries, and found 85.6% of venues never recommended. It is the closest design to ours, in a different market.

Download the data

Each Round 2 study ships its row-level answers, named-firm list, question set and validation sample under CC BY 4.0, covering 3,548 counted answers collected 13 to 16 October 2026. The extraction and analysis scripts ship with them, so every percentage on this page can be recomputed. There is never a list of firms that were not named.

• Houston neighbourhoods: answers (CSV), named firms (JSON), questions (JSON), validation sample (CSV).
• Houston suburbs: answers (CSV), named firms (JSON), questions (JSON), validation sample (CSV).
• Tampa Bay pilot: answers (CSV), named firms (JSON), questions (JSON), validation sample (CSV).

Free to republish with attribution and a link back. Absence from this data is not a finding about any business.

Cite this report

Free to republish under CC BY 4.0, including commercially, with attribution and a link back.

Questions about this report