Every Curated answer is more than the text in the chat. Knowing how to read the full output — sources, agent trail, confidence — is the difference between using Curated as a chatbot and using it as decision support.
The four parts of a Curated answer
1. The answer text
The natural-language response in the chat panel. This is what you'd read out loud in a meeting.
2. The map state
In parallel with the text, the map updates to reflect the answer — new layers, filtered features, visual emphasis on the parcels or polygons the answer references. The map and the text are answering the same question; switching focus between them is part of using the tool.
3. The sources
Every metric in the answer is cited inline. Hover or click a number to see:
- Provider — Esri GeoEnrichment, Esri Places, Dwellsy IQ, Hubexo, your CRM, your ArcGIS hosted layer, web search
- Dataset / variable — which specific source within the provider
- Vintage — the as-of date for the data
If a number doesn't show a source, that's a bug — flag it via the report-an-issue button.
4. The agent trail
Below the answer (or in the chat history view), the agent trail shows which specialist agents ran which tools to produce the answer. You'll see steps like:
- GIS agent: geocoded the address
- Spatial analysis agent: ran a 1-mile drive-time isochrone
- RAG agent: retrieved relevant Tapestry segments
- Synthesis agent: assembled the response
This is the methodology trail. It's what makes a Curated answer defensible at the committee — you can show exactly how the answer was produced.
Confidence flags
When an answer rests on thin data — a sparse dataset, a contested measurement, a partial-failure in an upstream agent — Curated flags it explicitly:
- High confidence — multiple sources agree, data is fresh, the question is within Curated's wheelhouse
- Medium confidence — some uncertainty in sources or assumptions; the answer text usually says why
- Low confidence / partial — the data is thin or an agent hit a failure; the answer is hedged
Curated will not manufacture a confident answer when the data doesn't support one. If you don't see a confidence flag and the data looks ambiguous, ask a follow-up.
Following up
Most analytical workflows aren't one prompt — they're a conversation. After the first answer:
- Refine — "narrow this to just the trade area south of I-30"
- Compare — "now do the same for Austin and rank both"
- Source-swap — "use our internal portfolio layer instead of the Esri Places competitor data"
- Export — "make this a PlaceStory" or "export this as a PDF"
The conversation history persists. You can pick up tomorrow where you left off.
Related
- Refining Results with Follow-up Prompts
- Spaces, Artifacts, and Maps
- Voice Mode