Prompt
“Best kitchen remodeler in Charlotte, NC”
Answer, as recorded in our run log (our summary, not the engine’s words)
ChatGPT named three independent Charlotte firms rather than defaulting to directories. It justified its picks almost entirely from third-party records: Houzz review counts and awards, NARI awards, an NC contractor license number, and a BBB rating. It cited none of the firms’ own websites.
Three independent firms named, rather than a list of directories.
The justification rested on third-party records. No firm’s own website was cited.
Protocol. ChatGPT temporary chat with no prior conversation, 25 August 2026. The full answer was logged verbatim with a dated screenshot at the time of the run. The answer text is not reproduced on this page; the exhibit gives the prompt as run and what the log records about the answer.
Passes. This entry records one pass. It is a record, not a score: under the method, a cell is scored on three passes and this run would count as one of three.
Annotations below the hairline, and the answer summary above it, are ours. The prompt is as recorded in the run log.
What did the answer lean on?
Records kept by other parties, not the firms’ own pages. ChatGPT justified its picks almost entirely from Houzz review counts and awards, NARI awards, an NC contractor license number, and a BBB rating, and it cited none of the three firms’ own websites. Every one of those details is ChatGPT’s claim, not ours, and we have not checked them. One run does not show why these firms were chosen over others. It does show where this answer looked for its reasons: at what other parties record about a firm, not at what the firm says about itself.
The answer did not send the reader to a directory. Houzz appeared as a record the engine drew on, not as the recommendation. Whether that holds in Charlotte on another day, for a differently worded prompt, or on another engine is exactly what a single run cannot tell you.
What does this entry not show?
It does not show what Perplexity, Gemini, or Google AI Overviews said, what ChatGPT said on a second or third pass, or what a signed-in homeowner in Charlotte would see. It does not show how many people ask this question. It does not reproduce the answer text itself. It is one dated answer, kept here because the site's pitch should be tested against answers that do not help it. The next entry is a client re-measurement with three passes, and it publishes whatever it shows.
What would a reader ask next?
Does one answer naming local firms mean the problem is overstated?
It means one prompt, on one engine, on one day, named local firms. It does not tell you whether the same prompt does so tomorrow, whether the other three engines do, or whether the firms named are the ones a homeowner would be best served by. That is exactly why the method runs three passes per cell across four engines. One answer is a data point; it is published here because a record that hides the inconvenient points is not a record.
Was this run part of a client engagement?
No. It was a standalone run in a market with no client relationship, chosen to test the site's own pitch against a real answer. Client re-measurements are published as separate entries and only with written consent.
Austin Brewer founded AI Visibility Partners and runs every measurement himself. This note covers remodeling, the first trade we measured; notes on other trades follow. All Field Notes