94% of Houston's 50,000 Realtors were never named in 2,000 AI answers

How this Houston AI visibility study was run

CiteVantage put 500 real-estate queries covering 40 Houston areas to four AI surfaces between 18 and 22 September 2026, producing 2,000 answers, 1,917 of which passed the geo-clean filter. Every business named and every domain cited was recorded, and the row-level data, query set and analysis code are published under CC BY 4.0.

The surfaces were ChatGPT, Gemini (API-based), the Gemini web app and Perplexity, and 83 answers failed the geo-clean filter: 60 had answered about the wrong place, and 23 further answers were excluded because they named businesses from outside the Houston metro. Names were extracted heuristically, so every structural percentage is published beside a strict-filter sensitivity check.

Queries

500, across 40 Houston areas and 8 intent tiers, plus metro-wide questions

Surfaces

ChatGPT, Gemini (API-based), Gemini (web), Perplexity

Answers

2,000; 1,917 counted

Excluded

60 answers (3%) about the wrong place, and 23 further answers that named businesses from outside the Houston metro

Recorded

Firms named and domains cited, per answer

Fieldwork

18 to 22 September 2026

Validation

394 names labelled, blind to outcome

Extraction

Heuristic; strict filter and a labelled sample published

Licence

CC BY 4.0, reuse with attribution

Prior work

The 150-query study, on which domains engines cite

Most named firms are named once

A labelled sample puts 56.5% of the real Houston agents, teams and brokerages named by four AI surfaces across 500 queries in September 2026 at exactly one mention (95% interval 50.8 to 62.2). As extracted, non-businesses included, the figure is 68.1%, and the median name has one mention.

Concentration is real without being extreme: a Gini coefficient of 0.493 across 10,998 mentions, with the top 1% of firms (47) taking 17.2% and the top 10% (478) taking 47.5%.

Breadth separates the head from the tail. Against everyone else, the top 50 firms average 9.3 areas to 1.3 and 3.42 surfaces to 1.15, and they reach 7.02 intents to 1.42, so the head sits in many conversations at once and the tail in one answer to one question. National brands are a small share, though they turn mentions into breadth more often:

• 148 brand and franchise names took 727 mentions, 6.6%, and 17 of them, 11.5%, reached five or more areas.
• The other 4,634, mostly individual agents and local teams, took 10,271 mentions, 93.4%, yet only 86, 1.9%, reached five or more.

Same engine, two surfaces

On 473 Houston queries both Gemini surfaces answered in September 2026, Gemini (API-based) named 7.8 firms per answer and the Gemini web app 3.9. They returned 3,686 and 1,842 firm mentions respectively, shared only 352 distinct firms, and on roughly 57% of the 434 queries where both named a firm named nobody in common, so a Gemini visibility figure that does not say which surface it measured describes neither.

Those 352 shared firms are 6.8% of the two surfaces' combined list. Query by query, they shared at least one firm on 189 of the 434 queries on which both surfaces named at least one firm, 43.5%. Both answered the same 473 queries in the same five-day window, so the sample cannot explain the gap: both are Gemini, and they return different Houston markets.

The web app also behaves like a local search product in a way the API-based surface does not. Google Business Profile markers, star ratings and opening hours, appeared in 82 of 473 Gemini web answers against 27 of 473 API-based answers.

Queries

473 answered by both Gemini surfaces

Firms per answer

7.8 API-based, 3.9 web

Firms named

3,686 API-based, 1,842 web

Shared

352 firms, 6.8% of the union

No firm in common

245 of 434 queries (56.5%)

Business Profile markers

82 web, 27 API-based, of 473

How many Houston Association of Realtors members are ever named

Four AI surfaces answering 500 Houston real-estate queries in September 2026 returned 4,782 distinct names, of which a labelled sample puts roughly 3,000 (95% interval 2,748 to 3,206) as real agents, teams and brokerages. Against the Houston Association of Realtors' own figure of more than 50,000 members, that is about 6.0% named at least once.

HAR's own figure is the denominator, and the point survives a lower one: at 30,000 active agents only about one in ten would be named, against one in seventeen on HAR's figure. FlyDragon's national report reached a convergent result by a different method, pointing the same way at the same order of magnitude, ours the stricter:

• FlyDragon found 91% of agents effectively invisible in AI answers, and of those on at least one AI surface, only 11% on three or more.
• This study's controlled-subset equivalent is 4.3%: 3.1% of names on three surfaces plus 1.2% on four.

Does this hold in other metros?

This study measured one metro at depth: 500 queries, 40 Houston areas, four AI surfaces, September 2026. We will not claim its percentages hold elsewhere until another metro is measured, though the shape it shares with a national study of cited domains is the part most likely to transfer.

Our 150-Query Real Estate AI Search Study found 83.0% of 1,297 cited domains reaching only one of five engines nationally; this study finds the same pattern at business level in Houston. FlyDragon's market atlas covers many US metros, and our data is consistent with its Houston head: 20 of its 25 ranked Houston agents appear in our data, from a page stamped 23 July 2026, two months before our fieldwork.

On whether any of this reaches a buyer, the evidence cuts both ways. For it:

• Pew Research Center (fieldwork March 2025, published July 2025, n=900, 68,879 searches) found users clicked a result on 8% of visits with an AI summary against 15% without.
• Ahrefs measured a 34.5% lower click-through rate for the top result on keywords with an AI Overview in April 2025.
• Its repeat measurement in February 2026 found 58% lower.
• Bank of America's 2026 Homebuyer Insights Report (June 2026, n=2,000) found 20% of prospective buyers and homeowners had used AI tools for homebuying research in the past year.

Against it, at full strength:

• Similarweb counted 1.13 billion AI referrals against 191 billion from Google in June 2025 across the top 1,000 websites, roughly 0.6%.
• The NAR 2025 Profile of Home Buyers and Sellers (4 November 2025, n=6,103) found 88% of buyers and 91% of sellers used an agent.
• In that NAR profile, 91% of buyers would recommend theirs.
• Portals now sit inside the surfaces, Zillow in ChatGPT since October 2025 and Realtor.com since March 2026.

No study anywhere isolates consumers using AI to choose an agent, so the stakes remain unmeasured.

What a Houston agent can do about this

The 500-query Houston data from four AI surfaces, September 2026, points to six moves: target one area and one intent, measure each surface separately, keep the Google Business Profile accurate for the Gemini web app, be present in the portals and agent referral networks the answers cite, write the state into your copy, and remeasure.

Each rests on a Houston finding:

• Pick one area and one intent: an estimated 82.6% of named Houston agents hold one area, and new-build answers exclude 94.0% of the names in the same area's generic answer, so "best realtor in Sugar Land" and "new construction in Sugar Land" are separate contests.
• Measure each surface on its own, because 83.3% of real agents appear on only one, and even the closest pair, the two Gemini surfaces at 13.7% overlap, share nobody on roughly 57% of queries.
• For the Gemini web app, keep the Google Business Profile accurate and current: star ratings and opening hours appeared in 82 of its 473 answers against 27 on the API-based surface.
• Be present where answers are assembled, since portals and agent referral networks take 43.6% of citations while a firm's own domain appears in roughly one naming answer in ten, so a complete, current presence on HAR, Zillow and Realtor.com does more visible work than another post on your own blog.
• Write "Houston, TX" into the copy people quote you in: queries with a state marker went to the wrong place 1.1% of the time against 4.5% without one.
• Then remeasure, because this is one run in one five-day window and answers move.

In a thin area such as League City the surfaces have named 110 distinct firms in total, a field you can picture. For how ChatGPT in particular picks and cites firms, and which tactics have evidence behind them, see our guide to how to rank in ChatGPT.

Method and limits

CiteVantage wrote 500 Houston real-estate queries across 40 areas and eight intent tiers, put each to four AI surfaces between 18 and 22 September 2026, and recorded every firm named and domain cited in all 2,000 answers. Of those, 60 (3.0%) were excluded for answering about the wrong place, and 23 further answers were excluded because they named businesses from outside the Houston metro, 83 in all. Names were extracted heuristically, and every figure is published with a strict-filter check.

Query design. The eight intent tiers are discovery, sell, first-time, invest, relocate, luxury, new-build and condo, and 20 further metro-wide questions ask about Houston as a whole. The 500 break down like this:

• 38 of the 40 areas carry 12 queries each: four discovery, two sell and one per remaining intent.
• Spring carries 24.
• The generic Houston set is the 20 metro-wide questions.
• Half are search-shaped and half conversational, 250 each, asking the same underlying question the way a person types it into a search box and the way a person asks a chatbot.

Surfaces. ChatGPT, Perplexity and Gemini were queried directly through their web apps, alongside Gemini (API-based). The two Gemini surfaces are reported separately throughout because, as the Gemini section shows, they return different Houston markets.

The geo-clean filter. Every answer was checked for whether it was about Houston at all. The 60 that were not were removed before any other number was computed, and 23 further answers were excluded because they named businesses from outside the Houston metro, so every percentage here runs off the 1,917 geo-clean answers; the wrong-place failures are reported in their own section.

Extraction is heuristic. A script pulled business and agent names from the answer text, mistaking some phrases for names and missing some real ones, so every structural percentage is published as extracted and checked two ways:

• A strict filter deletes 24.2% of names, real firms included, yet still lets some non-businesses through, so it is a stress test whose count cannot be read as a floor.
• The headline figures come from 394 extracted names drawn at random and labelled by an AI reviewer, blind to how often or where each was named, then re-weighted to the full list with bootstrapped 95% intervals; the labels are published so any of them can be disputed.

Who labelled the sample. An AI reviewer labelled a random sample of 394 names blind to how often each name appeared. No person re-checked that sample.

One run, one window. Fieldwork ran 18 to 22 September 2026. AI answers move between runs, so this is a snapshot of that window, though some movement is smaller than it looks: on ChatGPT, 29 of 37 retried geo failures were still wrong two days later. A specific firm's position may have shifted since.

We sell this. CiteVantage sells AI visibility work to real-estate firms, so our commercial interest is direct and declared. This is independent research, commissioned by no one, with no vendor platform underneath it, and the own-website finding above cuts against the easiest thing we could sell. The row-level data, query set, labels and the analysis, sensitivity and validation scripts all ship with the study, so every figure here can be recomputed.

Questions about this Houston AI visibility study