Retrieved, then named: a field observation of purpose-built pages for two local businesses

When language models answer a buyer question from web search results handed to them, with an instruction to name relevant businesses, does a local business get named when its own page is in those results? And does building a page get it into them? We watched both on two Kansas City businesses: our own agency and one client. Six pages targeted buyer questions for which baseline searches had not returned the business's site. Six control questions got no page. A scanner asked each question repeatedly. Each time it ran one web search (Tavily) and handed the same results to ChatGPT and Claude through their APIs, with an instruction to name relevant businesses the results support. It recorded whether the target page was in the results and whether each answer named the business. What we saw in 72 fresh follow-up searches, October 12 to 26: Retrieved, then named. A target page was in the results for 31 searches. Both models named the business in 30 of those 31; across the 62 answers, 61 did. Not retrieved, not named. On the agency's 41 searches without its page in the results, neither model named it, or mentioned it at all. Retrieval was uneven. Four pages were retrieved; two never were. One comparison failed our own criterion. The client's control question started returning his site before the pages went live, and later through one of the new pages. What this is and is not. These are API answers generated from shared Tavily results under an instruction to name relevant businesses. Retrieval was less consistent than naming once a target page was supplied. The study does not reproduce the ChatGPT or Claude app and does not establish that publishing the pages caused the observed changes. In this pipeline, a target page being in the supplied results was strongly associated with the business being named. Conflict of interest. Lefty Media Co sells AI visibility work to local businesses. Four of the six pages are on the author's own agency site, which accounts for 60 of the 72 follow-up treated searches. The client pays Lefty Media Co. The predictions were recorded internally on 2026-10-06, not in a public registry. The database queries, paper reading, answer audit and drafting were done with Claude (Anthropic), run as Claude Code; draft feedback from Perplexity and Gemini was AI-assisted, not independent peer review. Deposit. The zip holds the observation export (805 answers), the analysis scripts, the answer audit and the author's spot-check, copies of the six pages as served on 2026-10-27 with a fetch log, the calendar event record, the database prediction text, a one-line public summary of every logged site change, a README, and a SHA-256 manifest. Author's research page: leftymediaco.com/research