Rank the most valuable recurring use cases where an AI agent needs visual evidence, and recommend what Evidence API should support first.
Re-run the analysis specifically for Evidence API — which accepts a text question and returns an existing, authoritative chart, table, map, diagram, or research visual with source, provenance, and context. Include only recurring, text-searchable retrieval of existing public visuals; exclude GUI/computer-use, OCR, uploaded-image analysis, video, inspection, and generation tasks. Rank the top 30 use-case-and-corpus combinations with scores (product fit 40%, recurring demand 25%, visual necessity 20%, source quality 10%, acquisition feasibility 5%), deduplicate by corpus, and give the top 10 corpora to add next in build order.
This report ranks 30 recommendations — one per normalized corpus family — screened from 260 high-relevance candidates, themselves drawn from 9,812 deduplicated pages found by 48 web searches across economics, policy, health, climate, science, energy, transport, education, and research figures. Scores run 0–100; the top score is 90.5. Ranks 1–10 form the recommended build order.
Weighted score = 10 × (0.40 product fit + 0.25 recurring demand + 0.20 visual necessity + 0.10 source quality + 0.05 acquisition feasibility), each component 0–10. Colour marks the corpus sector. Hover or tap a bar for the likely question, the visual needed, and the component scores.
The earlier ranking mixed general computer-vision and visual-agent tasks. This rerun keeps only what Evidence API actually does: a text question in, an existing authoritative visual out — with source URL, provenance, and surrounding context — so an agent can support, verify, compare, or explain an answer. GUI and computer-use work, OCR, invoices, uploaded-image analysis, video, inspection, satellite analysis, AR, geolocation, camera feeds, defect detection, and image generation were all removed. Of 260 screened candidates, 190 met the Evidence API definition and 70 were excluded — mostly raw-data-only retrieval, derived aggregates an agent would have to compute rather than retrieve, and interaction tasks. Evidence API is not a chart generator, an OCR service, an image-analysis model, or a raw-data API: its edge is that a text-only result may quote a number, but it rarely preserves the visual comparison, legend, scenario definitions, confidence intervals, or geographic pattern embedded in the authoritative artifact.
| # | Use case — corpus | Score | Fit | Demand | Visual | Quality | Feas. | Visual needed | Cadence | Source |
|---|
48 broad web searches produced 9,812 batch-level URL-deduplicated pages and 773 extracted candidate records; 260 high-relevance candidates were screened against the exact product constraints, 190 included and 70 excluded with recorded reasons. The final 30 rows deduplicate to one recommendation per normalized corpus family, so many questions from one source cannot crowd out corpus diversity. Text prefixed “Inference:” in the evidence is analyst inference; everything else is a source observation.
Feasibility ranges from open, keyless public access (FRED, World Bank CC BY 4.0, NOAA drought files) to unstated reuse terms (IPCC, CBO, several NASA pages) and registration gates (GBD). Where licensing is null the terms were not established — do not assume either permission or prohibition; confirm reuse terms with the publisher before ingestion.
Observed cadences span real-time (FRED), daily (Drought.gov Climate Engine), weekly (U.S. Drought Monitor), 56-day (FAA charts), twice-yearly (IMF WEO), annual (UNEP, WDI, EJScreen), and 6–7-year cycles (IPCC assessments). Schedule connectors per corpus at its observed cadence; for the 14 rows with no established cadence, start with periodic re-crawl and detect edition changes from page metadata.
The 30 rows span 22 distinct source hosts; no single host backs more than 5 rows. Several evidence pages are mirrors, archives, technical documentation, or secondary descriptions (archive.org, web.archive.org, wiki mirrors, developer write-ups) that point to the named authoritative corpus rather than its primary visual landing page. Flag those rows for primary-source connector validation before production ingestion.
Method: 30 ranked Evidence API use-case-and-corpus recommendations (result of a 260-candidate screen from 9,812 pages across 48 searches), one row per normalized corpus family. Weighted score 0–100 = 10 × (0.40 product fit + 0.25 recurring demand + 0.20 visual necessity + 0.10 source quality + 0.05 acquisition feasibility), components 0–10. Nulls mean the fact was not established. Table trims observed-demand, advantage, and example-search text for space; the chart tooltip and sources carry them. Compiled from the screened research corpus current at analysis time.