Are real estate websites getting agents recommended by AI?

We asked whether selected residential real estate businesses appeared in AI answers to local questions that did not include their names. The panel contained 25 publicly attributed websites for each of Agent Elite, AgentFire, Lofty, Real Estate Webmasters, and Sierra Interactive. Each site received the same 12 frozen question templates, localized to its market.

The reported panel used web-enabled OpenAI, Anthropic, and Perplexity APIs. With 125 websites, 12 questions, and three panels, the design produced 4,500 prompt units. Nineteen answers were unavailable, leaving 4,481 usable answers. All Gemini rows, including corrected-pilot rows, were excluded from these reported results.

The panel recorded 570 name mentions (12.7%), of which 417 were reviewed recommendations (9.3%) and 153 were mention-only decisions. Separately, 227 usable answers cited or linked the audited domain (5.1%). These overlapping outcomes cannot be added into a total visibility count.

The same question templates and panel configurations were used across these five cohorts. The selected businesses and markets were not matched. Read this as a descriptive table of the observed sample, with each rate calculated over that cohort's usable answers.

OpenAI contributed 1,500 usable answers, Anthropic 1,489, and Perplexity 1,492. A panel's recommendation rate must use its own eligible count. Model and retrieval configuration can affect the answer, so API results should not be presented as consumer-app market share or a promise of what a logged-in user will see.

All 570 candidate name matches in the current cohort received a recorded review decision: 417 recommended and 153 mention-only. This is why the reviewed recommendation total is smaller than the raw name-match total.

A site-level zero-name-mention statement is permitted only when all 36 expected answers were eligible. Across the five cohorts, 115 sites had complete panels and 63 of those complete sites had no name mention. Sites with unavailable answers are excluded from that complete-panel zero count. Absence in these 36 answers does not establish absence across all AI queries.

The expanded November 11 attempt ledger records 5,811 non-branded prompt units: 1,500 each for OpenAI, Anthropic, and Perplexity, plus 1,311 for Gemini. That broader execution ledger is not the denominator for the reported result rates. The three-panel result remains 4,500 prompt units and 4,481 usable answers.

Attempts describe workload. Eligible answers describe the scored result set. Mixing those counts would make a recommendation rate impossible to interpret and would silently reintroduce excluded Gemini activity.

Our interpretation is that an audit should distinguish being named, being recommended, and supplying a cited source. Each outcome suggests a different next question about evidence and content. This observation does not identify a universal formula or establish that changing website providers would improve an individual business's result.

Start with a fixed set of locally relevant consumer questions and inspect the actual answers and sources. Use original research, documented expertise, clear identity, accessible text, and internal links to make useful evidence easy to find. Google says traditional SEO foundations remain relevant to AI Overviews and AI Mode and requires no special AI markup for inclusion. Other answer systems need their own observed tests.