Why ChatGPT repeats the same agents

1. The short answer

1. ChatGPT names the agents whose own websites it can read and quote. On "agent who knows 55+ communities near Sun City West", 28 ChatGPT answers cited agent-owned sites 134 times and portals (Realtor.com, 55places, Seniorly, a team page) 5 times. No Zillow, no Google, no Yelp. It picks from the pages it retrieved, and the pages it retrieved were personal agent sites with the agent's name, licence, phone, brokerage, specialty and FAQ in plain text.
2. The four it repeats are the four whose sites were built for exactly that. Paulette Ezell, Jeanette Hochstatter, Debbie Pastori and Bill Fleck all run an "AgentEdge Visibility Site" (AgentEdgeAI): a static site with RealEstateAgent, LocalBusiness, Person and FAQPage structured data, the ADRE licence number on the page, knowsAbout "55+ active adult community buying", an llms.txt, a noscript copy of the whole bio, and community guide pages. The other two sites ChatGPT keeps citing, Jan Cotten (a 90,000-word legacy site with 298 mentions of "Sun City West") and Hometown AZ (rich JSON-LD, 4-question FAQ), are the same shape by other means. ChatGPT is not "choosing" these people. It is choosing these pages, and the people come attached.
3. It does not change because the retrieval set does not change. The same six to eight pages come back for the same question every cycle, so the same names come out. Across five cycles (Oct 2 to 10) ChatGPT named Paulette Ezell in 22 of 28 answers, Jeanette Hochstatter 19, Jan Cotten 18, Debbie Pastori 15, Kevin and Annette Wurgler 13, Mike Reppy 12. Bill Fleck appeared once and Sharon Rowlson once on this phrasing; Pete's logged-out screenshot got Fleck first on a different phrasing, which is normal churn (see the measurement rules, 7.9% same-top-business on re-runs for local queries).
4. Google AI Mode is a different machine and names a different list. Its 29 answers cited google.com 146 times (Business Profile and Maps cards), then U.S. News' agent directory, Yelp's "best 10" page, FastExpert, Zillow's agent directory, and only then agent sites (retirearizonastyle.com, justjarl.com, retireaz-resortstyle.com). Its most-named people were Sharon Mason, Sharon Rowlson, Jarl Kubat and Arthur Hasher. Pete's note "Google sends people to a directory for 55+" is exactly this. Zero of the ChatGPT four were in Google's top four, and vice versa, which matches the published 82.6% single-engine figure.
5. Bing's organic ranking is not the explanation. Bing's top results for the same question (read 2026-10-12) are realtor.com and redfin listing pages, a Zillow agent page, an AgentPronto directory and Susie Waggoner's site. None of the AgentEdge sites rank. ChatGPT's own retrieval, with its own query rewrites, is selecting them.

2. The evidence from the answers (the Sun City West question, Oct 2 to 10)

2.1 Who gets named, per engine

2.2 What each engine cited

ChatGPT, 139 citations: pauletteezell.com 25, jancotten.com 23, jeanettehochstatter.com 17, pastorihomegroup.com 16, workwithwurglers.com 9, mikereppy.net 9, hometownaz.com 9, billfleck.com 3, allanharshteamre.com 3, realtor.com 2, seniorly.com 1, thepoolergroup.com 1, 55places.com 1. Almost every citation is a home page; the exceptions are pauletteezell.com/communities, /about, /buyers and pastorihomegroup.com/sun-city-west-guide. Bright Data reported web_search_triggered: false and no query fan-out for these runs, so the rewrites ChatGPT used are not in the data.

Google AI Mode, 206 citations: google.com 146 (opaque searchviewer cards), realestate.usnews.com 11, yelp.com 8, buyaztoday.com 7 (a "Best Realtors in Sun City West" team page with ten licence numbers on it), retirearizonastyle.com 5, fastexpert.com 5, justjarl.com 5, retireaz-resortstyle.com 4, 55places.com 4, realtor.com 4, exitrealtysuncityaz.com 3, zillow.com 2.

2.3 How ChatGPT describes a pick

A typical answer (Oct 10, 2:58 PM) lists five agents, each as: name, brokerage and office ("HomeSmart Nichols · Sun City West"), one sentence of specialty, designations if the site states them ("Holds SRES and ABR designations"), a phone number, and a link labelled "Website and contact". Every one of those fields is on the cited page in plain text and in the RealEstateAgent JSON-LD. Where the page states a number ("320+ closed transactions and more than 100 client reviews" for the Wurglers) ChatGPT repeats it verbatim. It never cites a review platform for these agents; the "reviews" it mentions are the counts the sites claim.

2b. The panel, 12 phrasings, 2026-10-12 (60 ChatGPT answers)

What the 12 extra phrasings (60 ChatGPT answers, October 12) add to §1:

• The four hold on the Sun City West 55+ shape of question ("agent who knows 55+ communities near Sun City West": Hochstatter 4, Wurgler 4, Cotten 4, Ezell 3, Pastori 2 of 5; "which agents should I contact first": Ezell 3, Hochstatter 3, Fleck 2, Pastori 2 of 5).
• They do not hold on other phrasings. "Sun City or PebbleCreek" returned PebbleCreek specialists from Realtor.com profiles (Judy Coventry, Geri Johnson, the Hefferns) plus Rowlson; "best realtor for Sun City West" returned Mike Reppy 5 of 5; the Sun City Grand, PebbleCreek, Trilogy, inherited-house and quick-close prompts named none of the eight.
• Realtor.com is ChatGPT's fallback source when no agent site matches the question: linked in 28 of 60 discovery answers, Zillow in 13. So a complete Realtor.com profile with the specialty and city stated is the second lever on ChatGPT, after the own site.
• The "who is" check passed for all eight (5 of 5 identified with licence and phone), but the answers repeat whatever the first source says: Ezell's site prints a licence validThrough of 2026-05-31, so 5 of 5 answers tell the buyer to verify her status; a cached ADRE read does the same to Caden Leali 5 of 5. A wrong date on your own page becomes a warning in every answer.
• Bright Data's ChatGPT output lost its citations on 2026-10-09; sources above were recovered from the links in the answer markdown.

3. What the recommended profiles have in common

Measured on the six AgentEdge sites plus Hometown AZ and Jan Cotten on 2026-10-12.

The pattern: one page that answers "who is this person, is she licensed, where, what does she specialise in, how do I reach her" in the first screen, in text and in schema, with the 55+ community names spelled out. That is the whole trick. It is also §2 of the AI Presence Score design (Consistency plus Presence plus Depth) implemented on one surface.

4. What the recommended profiles do NOT have

This is the part that matters for Pete's "profiles, not production" thesis, and it is counter-intuitive:

• Reviews on portals are mostly absent or dead. Ezell: no Zillow profile, 0 on Realtor.com and Homes.com. Fleck: 0 on Zillow, Realtor.com and Homes.com. Pastori: 1 Zillow review from 2016, 0 Realtor.com ratings. Hochstatter is the exception (Zillow 13, Realtor.com 6, both 2024). ChatGPT named Ezell 22 times with zero portal reviews.
• Portal data is stale for five of eight. 2025 brokerage moves left Long Realty or Home Realty on Zillow, Realtor.com, Yelp, Birdeye, Yahoo and MapQuest for Fleck, Hochstatter, Pastori, Rowlson (another team's bio on Yelp) and Chavez (Mesa on every portal). ChatGPT did not care, because it did not read those pages.
• Production is not the signal. Pastori has 408 Zillow sales; Ezell has 30 closed sales on her own site. ChatGPT named Ezell more often. Fleck is AgentEdge's own case study and was "not the highest-producing agent in Sun City West".
• Google Business Profile does not drive ChatGPT here, although it drives Google AI Mode. Ezell has no GBP found; Fleck's GBP website field still points at his Home Realty site.
• No third-party "best agent" list carries the ChatGPT four. None of the eight appear in the U.S. News Sun City West top entries; only Rowlson is on FastExpert's page; Yelp's best-10 lists Rowlson and Pastori. The ChatGPT picks did not come from lists.

So on ChatGPT in October 2026 the own-site signal dominates everything in Factors 2 to 5 of the design doc. On Google AI Mode the opposite: GBP, Yelp, U.S. News and FastExpert decide, and agent sites come second.

5. The two engines want different things

Implication for the Record Layer: an agent record page is built the way ChatGPT wants (static, schema, licence, specialty, FAQ). To move Google AI Mode it needs the agent's GBP and a directory presence as well, which the record cannot supply on its own.

6. What this says about the eight agents Pete sent

Two of the eight (Romero, Leali) are Realty ONE Group Glendale agents.

7. The recommendation tool: what another agent should copy

In order of effect on ChatGPT, then Google:

1. A personal domain with one readable page. Name, licence number, brokerage with street address, cell and office phone, markets, specialties with the community names spelled out, years, closed-sales count, three reviews with dates. Static HTML; the same text in RealEstateAgent + LocalBusiness + FAQPage JSON-LD; an llms.txt. This is what the Parallax agent record is (the Parallax record page); the difference from AgentEdge is that theirs sits on the agent's own domain and ours sits on parallaxlistings.com, which on 2026-10-12 had 0 citations.
2. Five FAQ questions in the buyer's words. "Does X specialize in 55+ communities?", "Is X licensed in Arizona?", "What areas does X serve?", "How do I contact X?", "How experienced is X?". ChatGPT's answers are built from these answers.
3. One community guide per community the agent actually sells in, with the community name in the title. Fleck's 11 guides and Rowlson's 33 community pages are the two deepest sets and both agents are recommended.
4. Fix the portal record after a brokerage move (Zillow, Realtor.com, Homes.com, Yelp, GBP website field). It does not move ChatGPT today, but it is what Google reads, and it is Factor 1 of the score.
5. A claimed Google Business Profile with reviews, for Google AI Mode. Rowlson (38) and Romero (31) are the Google picks; Ezell (none) is invisible there.
6. Get onto the directory pages Google cites: U.S. News (fed by HomeLight), FastExpert, Yelp's best-10. These are the "directory" Pete saw.
7. YouTube with community names in titles moves Google AI Overview (Romero, Leali), not ChatGPT.

8. Caveats

• One question, one phrasing, two engines, ten days. The panel adds 12 phrasings and the 8 per-agent checks; see the panel dataset for those numbers.
• Bright Data's ChatGPT is logged out and US-located. Pete's logged-in ChatGPT (IMG_1766) had a chat history about 55+ homes, which personalises the answer.
• Google AI Overview (Pete's third screenshot) is not an Elmo engine here; AI Mode is. They cite differently (Overview leaned on YouTube).
• Correlation on four agents. The AgentEdge sites are also the newest, most text-dense, most specialty-explicit pages in the market; which of those properties does the work is not separable from this sample. Pete's Step 0 comparison group is the way to separate them.