A general AI chat tries to shortcut serious market research with training data, and it gets things wrong in ways that look right. The fix is a dedicated Project per submarket, grounded in real published sources, with instructions that block speculation entirely.
Name it after the specific area, not a generic label, so the model never blends two markets together.
Macro consultancy reports, submarket-level data, six months of transaction and rental evidence, supply pipeline data, and developer track records.
Tell the model to answer only from uploaded documents, cite the source and page for every number, flag thin sample sizes, and refuse to speculate.
Once the rules live in the project instructions, a one-line question returns a calculated, cited answer without you engineering anything clever.
You are my research analyst for this submarket. Answer only from the documents I've uploaded. Cite the source and page for every claim. Flag any figure resting on fewer than 10 data points. If a number isn't in the documents, say so and name which source would supply it. Do not invent or estimate numbers that aren't there.
The analyst is only as good as the reports you fed it. Refresh sources every quarter.
Records typically surface 30 to 60 days after closing. The trailing month is usually incomplete.
A two-year-old service charge or fee schedule can already be wrong enough to swing a yield calculation.
A community can score well on paper and still have a real problem nobody wrote down. Visit the area. The project is the prep, not the decision.