Thirteen days. One real estate agent with no domain authority, in a market owned by teams with thirty-year head starts. What moved, what did not, and the thing we found underneath.
Anne Sostman is a luxury listing agent in Scottsdale, Paradise Valley and Arcadia. In July 2026 she had a brand new website, a domain rating of 4.6, and one closing. The organic traffic she did have was arriving on tourism queries about Scottsdale weather, not on anything a seller would type.
We were not asked to make a good site better. We were asked whether an unknown agent can be recommended by AI answer engines, and how you would know.
So before touching anything, we measured. A fixed panel of 25 prompts, run against web-grounded AI search, recording for every prompt whether she was named, who was named instead, and which sources the answer was built from. That baseline ran on 27 July 2026.
The second run, on 3 August, produced the number that matters and it is not the one anyone expects. Her site went from being cited on 1 of 7 prompts to 13 of 25, and on that panel it tied for the single most cited domain in the entire result set.
The engines had her. They still would not recommend her.
One prompt made it unmistakable. Asked which Paradise Valley agent handles discreet, off market sales, the engine pulled five of its ten sources from her site, then named no agent at all. Retrieval was solved. Generation was not converting.
An answer engine will read your material long before it will vouch for you. Those are two different problems, and almost everyone measures only the second one.
The exception told us why. She was named on exactly one prompt outside the ones that ask about her by name: is there a real estate agent in Scottsdale with a marketing or advertising background? She was the only agent named, and her About page was the cited source.
Attribute questions convert before ranking questions. An engine will answer "which agent has X" from a claim you substantiate yourself. It will not answer "who is best" without third parties it treats as independent. That is not a content problem. It is a corroboration problem, and it changes what you build next.
We publish the right column because a case study without one is an advertisement. Every figure above is dated and traceable to a specific run or report. Nothing is modelled, projected, or averaged into looking better than it is.
Most AI visibility tools sell you a rank position. That number is close to meaningless: ask the same engine the same question twice and the brands come back in a different order. We record presence rate across repeated runs with a confidence interval instead, and we never blend engines into a single score, because the same brand scores very differently on each one.
The prompt wording is frozen between runs. When we noticed one engine had begun refusing to rank agents at all, we flagged those prompts rather than counting the zeroes against her, because they measure the engine's policy and not her standing.
The panel, the run history and the raw answers are kept as files, not as a dashboard screenshot. Anyone can re-run it and get the same shape.