Weekly cadence. Every Monday morning, each of the ten prompts is run against ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Responses are captured and scored.
Five-point rubric. Each response gets scored 3 (named + accurate), 2 (named but incomplete), 1 (mentioned in passing), 0 (not named at all), or −1 (named with incorrect facts — the worst outcome, because the AI is actively giving buyers wrong information).
Competitive context. For each prompt, we also log which other builders and communities are named. If an AI tool consistently puts a competitor ahead of Deer Springs on a buyer-realistic query, that's a signal we feed back into the content library.
What we do with the data. When an AI engine misses Phase 2 or quotes it wrong, we update the content library (schema, FAQ markup, llms.txt, structured features) so the next week's query corrects. The arc from a 20% baseline to a 60%+ target is what the shadow site is designed to measure.
What we don't do. MRKT doesn't warehouse buyer data. Lead data from the registration form goes directly to Honeyfield's existing Mailchimp audience. The only data MRKT stores is public — AI engine responses and competitive mentions. Honeyfield owns the leads; MRKT owns the measurement.