Correction, Publishing, and Retesting

Measurement without action is a dashboard; action without measurement is guesswork. This page covers the action half of the loop: from a confirmed finding to a retested outcome.

From finding to task

Every confirmed finding becomes a remediation task with an owner, a target property, and its grounds — the approved facts it rests on. Typical task shapes:

• Fix the origin: update the trust center, integrations directory, or pricing page that assistants should be citing;
• Publish superseding facts: when a stale third-party source keeps winning retrieval, publish the current facts in more extractable form and request re-crawls where channels exist;
• Add the missing page: many “AI is wrong” cases are actually “we never stated this anywhere crawlable”;
• Contextualize an opinion: publish SLA facts or process facts beside an attributed opinion — never a rebuttal of the reviewer.

Drafts are fact-constrained

Remediation drafts may only assert claims present in the approved fact registry, and they carry their fact references. Anything the registry cannot support gets removed or sent back for approval first. This is what keeps correction from becoming spin.

Publishing belongs to the client — and gets logged

FactRelay never edits your properties and never publishes on your behalf. You review, edit, and publish; we log URL, version, and date. The publishing log matters more than it looks: it is the timestamp that separates “before” from “after,” and the only honest anchor a retest has. Unlogged fixes cannot be evaluated.

Like-for-like retesting

A retest re-runs the same frozen prompt-set version, on the same declared surface, under the same sampling rules as the baseline — nothing else qualifies for comparison. Reports show:

• baseline vs retest n/N per prompt, with sample sizes;
• control prompts (brand-free) to expose background drift;
• the published actions in the window between;
• explicit interpretation limits.

Interpreting results honestly

• Improvement is directional evidence, not proof one page caused it — platforms change constantly beneath us;
• Null results are results: the source may not have been re-crawled, the fix may be insufficient, or the hypothesis was wrong. Null outcomes move to a longer window (30/60 days) with that reasoning stated;
• Regressions get investigated from snapshots, not narrated away.

Cadence: an early retest at ~14 days catches fast-moving surfaces; 30–60 day windows better reflect re-crawl and index cycles. Ongoing operations run this loop monthly or quarterly against a maintained prompt bank.

Customer-facing worked demonstration: public demo case, sections 05–06.