How the machines decide who gets named.

Two sources: memory and retrieval.

Every assistant works from a blend of the same two things. The first is the model itself: patterns learned from a vast crawl of the web, months or years old. The second is live retrieval: search tools the assistant runs at the moment of your question, pulling current pages to ground its answer.

Both paths run through crawlers. If a crawler read good evidence about you, the model may already associate your name with your market. If your site, profile, and reviews surface well in retrieval, the assistant has current material to cite. An agent invisible to crawlers is invisible to both paths at once, which is why the first check in any serious audit is whether your site even lets them in.

Evidence, corroboration, recency.

When the question is "who should I hire in this neighborhood," the assistant is doing something close to what a careful friend would do: gathering what is written about the candidates and judging which one the record supports. Machines run that judgment on signals.

• Identity clarity. One consistent answer to who you are, where you work, and whom you work with, everywhere your name appears. Contradictions between listings read as uncertainty, and assistants do not stake their one answer on uncertain entities.
• Documented territory. Content that ties you to specific places and specialties: neighborhood pages, market commentary, transaction history. "Serves the greater metro area" ties you to nothing.
• Trust signals. Reviews with recency and substance, an active Google Business Profile, and third-party corroboration of the claims your own site makes.
• Liftability. Answers need sentences to build with. Pages that state facts directly get quoted; pages that circle a topic for eight paragraphs get skipped for a competitor who just said the thing.

Four surfaces, one body of evidence.

ChatGPT answers from its model plus web search. Perplexity retrieves and cites openly, which makes it the easiest assistant to audit. Gemini grounds against Google's index and profile data. And Google AI Overviews sit on top of the ordinary results page, drawing on the same local signals as the map pack below them. Different products, heavily overlapping inputs.

That overlap is the strategic point: you do not run four campaigns. You build one body of evidence, clean and consistent and current, and every surface that reads the web reads it.

More than you would think. Less than a vendor promises.

You control your site and everything on it. You control your profile and how alive it is. You control whether your identity matches across the web, whether your content answers real questions, and whether crawlers can read any of it. Those levers are the whole game, and they are all workable.

You do not control the model's final judgment, and neither does anyone else. That is the honest boundary of this work: build the strongest evidence in your market, measure whether the answers move, and distrust anyone who claims the machine is theirs to command. Our measurement, theprompt battery, exists for exactly that boundary.