Why AI assistants name some local professionals and ignore others: what 3,209 cited pages have in common

Mark van RyswykBuilding the professional trust graph for an AI-first world ·  · 8 min read
Why AI assistants name some local professionals and ignore others: what 3,209 cited pages have in common
Between July 30 and September 8, 2026, we ran 11,367 local hiring prompts through two AI assistants, ChatGPT and Google AI, and recorded every source each one cited. The prompts covered 103 California localities and 84 real estate specialties, phrased the way a buyer or seller actually asks: "I'm a first-time buyer in Burlingame, CA. Recommend a high-net-worth privacy real estate agent to work with." Across 22,268 assistant runs we logged 192,711 citations. We then fetched a sample of the cited pages, and a control set of pages that ranked in Google's top ten for the same locality and specialty but were never cited, and compared them.This article reports what the two assistants actually lean on, how much it overlaps with Google search, and what distinguished the professional pages that got cited from the ones that didn't.

Three things we did not expect

The two assistants source their answers from different places. ChatGPT's citations were 54% individual professional and brokerage websites and 43% Google Maps business listings. Directories such as Zillow and Yelp made up 3%. Google AI inverted that: 49% of its citations pointed to Google's own result panels, 21% to directories, 12% to redirect links that resolve mostly to Zillow, Yelp and FastExpert, and 16% to independent professional sites.Being cited has little to do with ranking in Google. Of the pages ChatGPT cited, 7% appeared in Google's top ten for the equivalent search and 17% appeared anywhere in the top 45. For Google AI the figures were 22% and 41%. Most of what either assistant cites, Google search does not put on page one.Page formatting did not separate cited pages from ignored ones. Question-shaped headings, a stated publication date, JSON-LD schema and FAQ Page markup all appeared at the same rate, or slightly more often, on pages that were never cited. What did separate them was whether the page could be fetched, and whether it named a licensed person.

How the study was run

Prompts: 11,367 prompts generated from a locality-by-specialty grid, 103 California localities (Bay Area and Sacramento heavy, with Los Angeles, Orange County and San Diego represented) and 84 specialties. Each prompt was run once per assistant.Assistants: ChatGPT and Google AI, accessed through Catchouse's prompt pipeline. Each run's cited sources were logged with URL and position.Cited set: after removing Google's own Maps listings and result panels, 14,665 distinct pages were cited directly and a further 12,825 were cited through Google AI redirect links. We followed every sampled redirect to its destination. The fetch sample was 5,327 distinct cited pages: 3,761 cited directly and 1,566 recovered from redirects.Control set: Google top-ten organic results, collected in the same window for the same locality and specialty grid, that no assistant cited. 6,399 pages.Fetch: a plain HTTP GET with a non-browser user agent, no JavaScript, 12-second timeout. Pages that refused the research user agent were requested a second time with standard browser headers, still without JavaScript. A page counted as reached only if it returned HTTP 200.Coding: a script coded each page for body text present (at least 200 words in the raw response), named author with a stated credential, street addresses or ZIP codes in the text, question-shaped headings, a full date stated in the text, and JSON-LD schema. For the four judgment-based attributes, 396 reached pages were independently re-coded by a review pass using a language model reading the saved page text. Script and reviewer disagreed on 7% of pages for question headings and 9% for stated dates, so we report those two from the full sample. They disagreed on 36% for author-with-credential and 32% for street-level detail, almost entirely the script over-coding, so we report those two from the reviewed sample only.Classification: pages were sorted into directories and platforms (Zillow, Yelp, Fast Expert, Indeed, LinkedIn, Reddit and similar), brokerage sites, and independent professional pages. The comparisons below are for independent professional pages, because that is the situation a working agent is in.

What separated cited professional pages from ignored ones

Independent professional pages: 2,853 cited, 1,973 control.
Why AI assistants name some local professionals and ignore others: what 3,209 cited pages have in common
The credential gap is the finding. A cited professional page was almost twice as likely to name a specific person and state their license number, real estate agent designation or similar credential. That difference holds at 26 points on the reviewed sample and is far outside sampling noise.Reachability is the other one, and it is smaller than we assumed. Cited pages were somewhat more likely to answer a plain request and less likely to sit behind a bot wall. But once a page was reached, text was almost always there: fewer than one in 200 pages in either set was a JavaScript shell with no body text. The "your page is empty to a crawler" problem is real for the pages it affects, and it affects fewer professional pages than the industry conversation suggests.Street-level detail showed a five-point difference on the reviewed sample, which is within noise at that sample size. The script's larger gap on the full sample was over-coding, mostly office addresses in footers.

What did not hold up

Three things we believed going in did not survive the data.Structured data did not matter on its own. Cited and ignored pages carried JSON-LD at the same rate, and FAQPage markup, the schema most often recommended for AI visibility, was no more common on cited pages. Markup describes text; it does not substitute for a named, credentialed person being in the text.Question-shaped headings and stated dates ran the wrong way. Ignored pages were slightly more likely to have both. The likely reason is that dated, question-headed pages are blog posts and news items, while cited pages were disproportionately profile and about pages that are undated by design. The assistants were answering "who," and they cited pages about a person.Fetchability was not the whole story. Directory pages on Zillow and Yelp blocked our crawler 59% of the time and were cited anyway, heavily by Google AI. Those platforms reach the assistants through other routes. An individual professional does not have those routes, which is why reachability still matters for them, but it is a condition, not a ranking factor.

Limits of this study

The sample is California residential real estate in one summer. We think the mechanics apply to other local professions because they are properties of retrieval rather than of the trade, but we have not tested that and will not quote numbers for other professions until we have.These are associations in one window from two assistants that change their behavior without notice. The control set is Google's top ten for a search phrased as a keyword, not for the conversational prompt itself; it is the closest equivalent, not the identical query.The review pass that validated the coding was a language model reading page text, not a human panel. The script-versus-reviewer disagreement rates above are the honest measure of how much to trust each row.

Questions we get asked

Does this mean search ranking is irrelevant?No. Classic search still sends most of the traffic a professional can measure. But the overlap between what Google ranks and what an assistant cites is small, 7% for ChatGPT, and that means a page can be invisible in search and still be the one an assistant names, or the reverse.Why does ChatGPT cite Google Maps so much?Because for a "who should I hire" question in a specific place, a business listing with a name, an address and reviews is the most direct evidence available. Forty-three percent of ChatGPT's citations in our data were Maps listings. A current, complete Google Business Profile is doing more work in ChatGPT answers than most professionals realize.Why doesn't my social media help?The expertise is inside the video or the image, and what reaches the model is text. Facebook and Instagram together were under 3% of citations across both assistants, and those were mostly profile pages, not posts.Do I need my own website?You need a page under your own name that a plain request can fetch, with your license or credential stated on it. Whether that is a site you own or a profile on a platform that renders on the server matters less than whether the page names you and says what you are licensed to do.Can I check this myself?Partially. Request your own page without a browser and read what comes back. If your name and your license are not in that response, neither assistant has read them either.

References