Does AI Recommend Your Real Estate Business? A Repeatable Test Worksheet

A screenshot of an assistant naming an agent can be useful evidence. It cannot tell you how often everyone else sees that agent. This worksheet helps a real estate team preserve enough context to compare its own observations and choose a sensible next investigation.

Google documents that AI Overviews and AI Mode can use different models and techniques, and their responses and links can vary. OpenAI documents that memory can influence search queries. Google also describes personalization in AI Mode. An identical sentence is therefore only one part of a comparable test.

This is our suggested manual observation method, not a provider ranking formula or a statistically representative survey. It uses public questions and an existing interface you are authorized to use. It does not require buying tracking capacity, connecting private client accounts or collecting other people's conversations.

Choose decisions your prospective clients actually need to make, then give each question a stable identifier. Keep a buyer question and a seller question in the panel so a convenient winning example does not become your entire test. The three prompts below are synthetic starting points, not measured search demand or actual customer messages.

Replace the city or property type before freezing your first version. Use a city stated in the prompt rather than assuming the assistant knows your intended market. Avoid inserting your agent name in a discovery question: that tests a different task from finding an agent without already knowing whom to ask for.

Record the product and surface as they actually appear: ChatGPT with search, Google AI Mode or another named interface. Do not combine Gemini, Google AI Mode and an AI Overview into one Google result. Record the visible model label if one is shown; otherwise write unknown instead of inferring a model.

Start each standalone question in a fresh conversation where available, and note that choice. A fresh conversation does not prove that account personalization, location context or other influences are absent. Note the existing personalization and sign-in state you can observe; unknown is an acceptable value. Do not change a shared account's settings just to complete a worksheet.

Write down the date, time zone, language, location wording, prior conversation context and any indication that web search ran. An interface with no visible citations is not automatically a search failure or proof that no web information influenced the answer. Record what was visible and leave the retrieval mechanism unknown when you cannot establish it.

Keep the exact wording with its panel version so an edited question cannot silently replace a previous test. Store only the business-relevant public observation; do not paste client names, addresses, finances or private conversations into test prompts or a public report.

For a manageable practice exercise, you could plan three fresh-conversation runs per question and product on a stated review day, then repeat the same procedure the following week. Three is an editorial example, not a minimum sample size that produces statistical confidence. Use a smaller panel if that is all you can review consistently.

Keep all planned attempts, including no mentions, errors and requests for clarification. Report planned attempts, completed answers and exclusions separately. If you summarize mentions among completed answers, state that denominator and show the unresolved attempts beside it. Do not call that fraction a share of all users or a city-wide ranking.

Do not keep refreshing until your business appears and then save only that answer. Changing the question, city, product, conversation state or sampling procedure creates a new comparison context. Version the change rather than presenting it as improvement under the original method.

Imagine a completely invented SELL-01 record: an answer mentions Example Agent and links to a directory, while your own business is not named. This is teaching material; it is not a test result, a real agent recommendation or a finding about Huntington Beach competition.

The appropriate note is narrow: in this invented answer, the business was not mentioned and a directory was cited. Inspecting the cited page could establish what that page says. It would not establish the assistant's hidden selection logic, how frequently people receive this answer or whether paying the directory would change recommendations.

If a real observation reveals an outdated public phone number, assign verification of that fact to its owner. If the answer is inconsistent without a clear factual error, keep observing before rewriting the site. A citation alone also does not prove that the cited source supports every claim in the answer.

Use a verified access problem to start the AI crawler-access checklist. Use a factual discrepancy to review the public source and its ownership. Use an unanswered client decision to improve the most relevant existing page before creating another similar URL.

Keep a website release log beside the panel, but do not attribute a changed answer to one edit merely because it happened afterward. A mention is also different from a website visit, an inquiry or a qualified consultation. Investigate attributed visits separately with the ChatGPT referral worksheet.

Google says there are no special optimizations required for its AI features beyond the relevant search fundamentals, and eligibility does not guarantee appearance. This worksheet can support a better decision; it cannot guarantee inclusion in ChatGPT, Claude, Gemini, Grok, Perplexity or Google's AI features. Discuss the pages and evidence behind your next decision through our real estate AEO service page.