Quick answer: ask it. On September 17, 2026 we asked ChatGPT to explain why it keeps recommending a Peninsula property manager whose online presence we built. It named five things: reviews written in specific language, one identity repeated across authoritative sites, a narrow specialty, quotable facts on his own website, and recent activity. It put technical markup last. Then we checked its claims against the live pages.
Andrew Guglielmi is a friend, a property manager in San Mateo, and the first client LeadJens ever built for. In June 2026 he appeared zero times on "best property manager in San Mateo" across ChatGPT, Perplexity and Google's AI answers. By late August he was second on that list, and a caller told him ChatGPT had put him first. So I asked the engine doing the recommending.
What did ChatGPT say was driving the recommendation?
Five factors, ranked, with the technical layer at the bottom. Its top two were the review corpus and identity consistency. The reviews mattered because of their language: owners, tenants, brokers and referral partners describing specific situations, with the same words recurring across Google, his site and his profiles. Consistency mattered because his name, license number, firm and service area say the same thing on his site, the state license lookup, the MLS profile and the directories, so every profile plainly described one person.
Third was the lane. Andrew is a named, reachable individual who manages one to about 25 units and industrial condos, and says so. ChatGPT's read: "property management company" favors the big firms, while "a responsive individual who manages my San Mateo condo and can advise me whether to rent or sell it" fits him almost exactly. Fourth was the website, because it publishes facts a model can quote: 77 properties, about 250 doors, since 2016, vacancies typically filled in under two weeks, and the fee. Fifth was corroboration from surfaces we do not control, the license record and transaction history.
Freshness came after that, then crawlability and schema, described as making everything easy to read while creating no credibility on their own. Its closing line: "schema and 'AI optimization' did not manufacture Andrew's authority. His real client history created the authority."
Which of its claims checked out?
Most of the checkable ones. On September 18, 2026 we opened every page cited.
| ChatGPT's claim | What we found on September 18, 2026 |
|---|---|
| The site publishes 77 properties, ~250 doors, since 2016, under two weeks to fill, and the fee | Verified on guglielmigroup.com and its llms.txt |
| The site explicitly allows the AI retrieval crawlers | Verified in robots.txt |
| The license number sits in the structured data and matches the state record | Verified: DRE 01852584 |
| The site's review count lags Google | Verified: the site shows 50; Google shows more |
| One profile link points at a generic page; two profiles carry different email addresses | Verified; on his fix list |
| He is not first on broad Google queries like "best property manager in Burlingame" | Matches our record: second on the San Mateo list, first once on August 20 per a caller, third logged out on August 26 |
That last row is the most useful thing it wrote. Our case study says second, first once, third once, with dates. The model checked Google itself and reported the same shape. A ranking claim needs a platform, a query and a date, and the engine agreed.
What did it get wrong?
Two things, and both are instructive. It read the site's ratingCount: 50 beside reviewCount: 49 as an inconsistency. One reviewer left five stars and no text, which is exactly the case those two fields exist to separate.
It also listed ClaudeBot among the crawlers that matter for being cited. Anthropic's crawler documentation, read September 1, 2026, describes ClaudeBot as collecting content that could contribute to training its models; the agents that fetch and cite a live page are Claude-SearchBot and Claude-User. Of 20 published guides on the topic that day, 18 made the same error, so the model was repeating the field's consensus.
A model explaining its own recommendation is assembling a plausible story from what it can read, with good recall and no access to its own weights. Every sentence is a claim to verify, and the wrong ones become your to-do list.
How do you ask an AI why it recommends you?
Six steps, no tools, twenty minutes.
- Ask the stranger's question first. "Best real estate agent in [your city]" and the nuanced version a real client would ask: "an agent who handles trust sales in Alamo."
- Ask why. "Why did you choose these? What would make you more confident recommending [your name]?"
- Ask for the teardown. "Review [your name]'s online presence and reverse-engineer what drives or blocks a recommendation." Do this logged out, in a fresh session, so your own history is not the answer.
- Open every page it cites. Verify each claim about you against the live surface, the way we did in the table above.
- Sort the results. Where it is wrong about you, fix the page. Where it is right about a competitor, that is your gap.
- Run it on three engines, separately. One model's answer is real and partial, and you cannot tell which part. Different models have different blind spots.
What does this actually tell you about AI visibility?
The five factors are evidence, and evidence is slow. The reviews took Andrew years to earn and a summer to make legible. The identity work was tedious and finished. The lane was a decision. The facts on the page were his numbers, written down. Keep the schema for Google's rich results, and put the words on the page for the models. How the engines gather that local data in the first place is in how agents get recommended by ChatGPT.
If you want the same read on your own name without running the queries yourself, the free AI Visibility Score shows you what the engines say today: get your free AI Visibility Score. It reports visibility, with dates, and promises nothing about leads. The mechanics live in the AI visibility FAQ.
Frequently asked questions
Can you ask ChatGPT why it recommends a business?
Yes. On September 17, 2026 we asked ChatGPT to explain why it keeps recommending a property manager we built for, and it returned a ranked list of 5 factors with its reasoning. Its answer is a story assembled from what it can read on the web, so each claim needs verifying against the live page. In our check on September 18, 2026, most of its checkable claims held and 2 were wrong.
What makes AI recommend one real estate agent over another?
In ChatGPT's own account on September 17, 2026, five things: reviews written in specific language by real clients, one consistent identity across the license record, the MLS profile, the website and directories, a narrow specialty that matches a nuanced question, quotable facts on an owned website, and recent activity. For the client in question, the reviews went from 27 to 50 during the engagement and the site publishes 77 properties under management.
Does schema markup make AI recommend you?
ChatGPT ranked technical crawlability and schema last of 8 factors on September 17, 2026, describing them as making a site easy to read while creating no credibility on their own. The client's site carries full structured data and the model said plainly that the markup did not manufacture his authority. Keep the schema for Google's rich results and put the actual words, numbers and answers on the page for the models.
Is an AI's explanation of its own recommendation accurate?
Partly, and you can measure which part. Of the claims ChatGPT made about one client's presence on September 17, 2026, the ones we could check on September 18 mostly held, and 2 were wrong: it misread a star-only rating in the site's review counts, and it named a training crawler as one that governs whether a page gets cited. Treat the explanation as testimony to verify, and the errors as your fix list.
How do I find out what AI says about me?
Ask a discovery question a stranger would ask, in a fresh session, on ChatGPT, Perplexity and Google's AI Mode separately; then ask each engine why it chose those names and what would make it more confident recommending you; then open every page it cites and verify each claim. The exercise took us about 20 minutes on September 18, 2026 and produced 3 fixes for a site we had built ourselves.
Sources: Jens Hansen's ChatGPT session, September 17, 2026, asking why the model recommends Andrew Guglielmi; live fetches of guglielmigroup.com, its robots.txt and llms.txt, and the SC Properties agent profile, September 18, 2026; the California DRE public license lookup; the LeadJens case study (leadjens.com/case-study); Anthropic's crawler documentation, read September 1, 2026; LeadJens's September 1, 2026 audit of 20 published crawler guides.