DATA LAB

Evidence, not a dashboard fantasy.

Every SocialFetch request is ledgered in SQLite when you run the CLI. This page is the live product view of what the lab measures. CLI: pnpm selftest or pnpm live.

API key presentrun 2026-08-25 · cap 3000 credits

Experiments

#	Question
02	Historical reach Does the ad library return ended ads, or only currently-live ads?
01	Resolution Can we resolve a company's ad presence from name/domain, without hand-entered IDs?
06	Web + SEO/AEO Can we read homepage/pricing well enough to state positioning, and are pages SEO/AEO-extractable?
07	Monitors Are monitors relevant, and what would they cost? (dryRun only)
08	Media hosting What would hosting media cost — so we can refuse to do it?
03	Incremental polling Does routine monitoring cost 1–2 requests per competitor per source?
05	Reliability How often does it fail for no visible reason? Which routes bill on failure?
04	Completeness Is there enough in a returned ad record to classify or group it?

Frozen thresholds

Committed in thresholds.md before the first run. The scorecard is PASS / FAIL / PARTIAL from measured values, not rewritten history.

resolve_meta
Meta (Facebook) advertiser resolution
≥ 90% exact on Facebook
resolve_linkedin
LinkedIn resolution on the B2B sample
≥ 80% exact across b2b_* companies
resolve_tiktok
TikTok resolution on the DTC sample
≥ 80% exact across dtc_* / consumer companies
regional_usable
MY/SG usability
MY/SG cohort complete enough for the beachhead
poll_requests
Incremental polling cost
Median routine poll ≤ 2 requests per competitor per source
completeness
Ad-record completeness
≥ 85% of retrieved ads can attempt message_angle classification
reliability
Unexplained failure rate
Repeated unexplained failures < 5%
web_extraction
Website + pricing extraction
Homepage and pricing usable across most tested sites
unit_economics
Cost per competitor per month (daily polling)
≤ $1.50 / competitor / month