Entity Consistency Sweep (W5)

extraction-audit checks whether your own pages are machine-readable. This checks whether the rest of the web describes you correctly and consistently — because on the parametric side, a wrong category or stale fact repeated across trusted surfaces is what the model's next training snapshot will encode. The sweep finds those, ranks them by damage, and turns them into remediation work.

Program context: AI and Automation/Corroboration Layer Playbook/. Rationale: parametric-vs-retrieval-framing memory.

When to use

After a CEK exists, and on a quarterly cadence, to find where third parties misdescribe the client. Also whenever a baseline measurement shows the model asserting a wrong fact/category and you need to trace it to live sources.

Prerequisite

The client's Canonical Entity Kit (ground truth). If none, run canonical-entity-kit first — you cannot judge consistency without the SSOT.

Workflow

Phase 1 — Harvest descriptions (how the web/LLMs describe the client)

• LLM layer: run/read a Gumshoe report — how models describe the client and which sources they cite. Note category, facts asserted, and associations.
• Structured web: DataForSEO business_data_business_listings_search + serp_organic_live_advanced for brand queries → knowledge-panel / listing framing.
• Key third-party profiles: scrape (firecrawl/brightdata) Crunchbase, PitchBook, G2, Capterra, LinkedIn, Wikidata/Wikipedia, top industry directories. For each, capture: category filed under, one-line description, canonical facts stated, comparison set / entities it's placed beside.

Phase 2 — Diff each surface against the CEK

For every surface score: category match (Y/N), each canonical fact match (Y/N/stale), comparison-set match, descriptor drift. Assign severity:

• High — wrong category or a false/stale canonical fact on a high-authority, crawlable surface (feeds corpus and Google graph).
• Med — drift on a mid-authority surface, or right facts but wrong comparison set.
• Low — cosmetic / low-authority / uncrawlable surface. Weight severity by surface authority × extractability (an uncrawlable wrong page is lower priority — the corpus can't read it anyway; confirm with bot-crawl-check on borderline cases).

Phase 3 — Classify remediation

Per finding: self-serve edit (we control/can edit) · submit correction (form/request, e.g. PitchBook/Crunchbase) · outreach to fix (editorial — route to journalist-outreach) · monitor (uneditable, low severity) · Wikidata fix (route to wiki-scout). Map each to acquisition ease.

Phase 4 — Output

• [Client]-consistency-sweep-[date].csv — W5 remediation rows, tracker-compatible (surface, current framing, CEK-correct framing, severity, remediation type, owner, ease). Drop straight into the corroboration tracker.
• [Client]-consistency-brief-[date].md — findings ranked by severity×authority×extractability, plus a consistency score = % of audited surfaces matching the CEK on category + core facts. Track it over time.

Caveats

• LLM output ≠ live web. A wrong description in a model's answer may be historical corpus — fix the live sources, then re-measure over training cycles; don't expect an instant flip. The live-surface fixes are the lever; the LLM output is the lagging indicator.
• Some surfaces are uneditable or slow (PitchBook, third-party editorial) — correction-submit or outreach, and accept lag.
• Consistency score is directional (surface set is a sample, not the whole corpus) — keep the audited surface list stable release-to-release so the trend is comparable.
• Prioritize crawlable + high-authority first; that's where the parametric payoff concentrates.