The CMA advisor returns a comparative market analysis in under 5 seconds. Input a subject property; it pulls 3–6 comps from public listing sources (MLS / Zillow / Redfin / Craigslist) within the radius, condition, and timeframe constraints, and emits a fair / low / high price estimate with each comp sourced and a confidence score on the range. It runs as one of the eight atomic agents in the platform — composable with the buyer-match, listings, and outcomes tools — not a separate product.
TL;DR
- Output:
{fair, low, high, comps[], confidence}with every comp sourced - Input: subject property address or MLS ID
- Speed: under 5 seconds end-to-end
- Sources: MLS (with realtor's credentials), Zillow, Redfin, public sales records
- Composability: feeds the listings tool for valuation tracking and the inbox tool for outreach drafting
- Verification: realtor reviews comps before the estimate is committed; the thought stays with the realtor
How does the CMA advisor pick comps?
The selection is a two-pass match. The first pass filters on hard constraints: property type, bed and bath count, square-footage range (±15% by default, configurable), neighborhood or radius (½-mile default for urban, 2-mile default for suburban), and sale or list date within the prior 6 months. The second pass scores remaining candidates on similarity using a real-estate fine-tuned embedding model (gemma4-realty:v1) — comparing layout, condition signals, lot size, and historical price-per-square-foot trajectory. The top 3–6 by composite score become the comp set.
Filters are tunable per realtor. Some farm areas are tight and need ½-mile + 90 days; others are sparse and need 5-mile + 12 months. The realtor's tunings are saved per market and reused on the next CMA in the same area.
Why does the realtor see the comps before the estimate is final?
McKinsey's productivity framework distinguishes steps from thoughts. Pulling comps and computing a price range are steps — repeatable, rule-driven, and faster when automated. Choosing which comps actually represent the subject property is a thought — it requires local knowledge, condition awareness, and trust calls. The CMA advisor automates the steps and hands the thought to the realtor before locking the estimate.
The realtor can drop or override any comp; the estimate recomputes immediately. The override is logged in agent_audit_log, and over time the realtor's override patterns tune the per-tenant comp-selection weights.
What's the confidence score and how is it calculated?
Confidence is a 0–100 number reflecting two factors: the variance across the comp set's price-per-square-foot, and the embedding-similarity distance between the subject and the chosen comps. Tight clusters of highly similar comps yield 80+ confidence. Wide variance or weak similarity yield 40–60 with a flag suggesting the realtor expand the search radius or the timeframe.
The confidence score is shown alongside the fair / low / high range so the realtor knows when to trust the estimate as-is versus when to rework the comp set manually.
How is this different from a Zillow Zestimate or RPR's automated CMA?
| Tool | Source | Realtor input | Confidence shown | Composability |
|---|---|---|---|---|
| Zillow Zestimate | Zillow's algorithm + listings | None — output is fixed | No | None — standalone consumer page |
| RPR Automated CMA | NAR data feeds | Realtor selects comps manually | No | None — exports to PDF |
| Cloud CMA | MLS + various | Realtor builds comp set | No | Embeds in marketing presentations |
| This platform | MLS + public + per-tenant tunings | Realtor approves and overrides comps before lock | Yes (0–100, embedding-derived) | Composes with listings, inbox, outcomes atomic agents |
The platform's CMA advisor is not a standalone product to replace Cloud CMA or RPR. It exists because the realtor needs CMA inside the same workflow as buyer-matching and outreach drafting — pulling a CMA, drafting a price-reduction message, and logging the outcome should be one continuous flow, not three tools.
Can the CMA be exported as a PDF for buyers or sellers?
Yes. The export uses a per-tenant template (logo, broker info, disclosure language) and is generated server-side. Brokerages with their own template requirements can supply a Markdown / Handlebars template at onboarding; the platform fills it from the structured CMA output.
What if the subject property has no comparable recent sales?
The advisor surfaces this explicitly rather than fabricating an estimate. When the comp set is too small or too dissimilar, the output flags low confidence and suggests three responses: expand the radius, expand the timeframe, or fall back to neighborhood median trends with the absence-of-comps acknowledged. McKinsey's trust scaffolding principle applies: the system says here is what I could find, here is what I cannot, never here is a number, trust it.
How does the CMA tool feed the rest of the buyer pipeline?
CMA output is structured data, not just a PDF. The fair / low / high range writes back to the listings table and is referenced by:
inbox— for drafting price-positioning messages to buyers ("priced $12K below the neighborhood fair value")outcomes— for tracking whether listings priced near the CMA fair-value sell faster than listings priced abovebuyer-matchskill — for filtering buyers whose budget falls inside the fair / low / high band
This is the atomic-agent pattern in practice: one tool's output is another tool's input, all composed by the realtor's skill.
Start a pilot
Free 14-day pilot. The CMA advisor ships in every tier. Request access — no credit card.