Skip to content

Evidence protocol · Version 0.1

Publish enough context to reproduce the claim—or narrow it.

This protocol is the standard for new public evidence packs. Historic materials do not yet satisfy every field. The index states those gaps instead of silently treating old summaries as raw evidence.

Required fields

Every public result carries its measurement envelope.

The envelope makes the claim smaller, clearer, and harder to misread.

Named property, exact prompt, engine, and model identifier
Capture timestamp, publication timestamp, and collection period
Raw response and screenshot, or an explicit unavailable status
Hash and signature details, including what each mechanism proves
Intervention timeline with deployment anchor and known confounds
Failures, refusals, non-movers, and divergent engine results
Methodology version, sample size, network point, and rerun instructions

Collection sequence

Baseline, intervention, observation, publication.

  1. 01

    Freeze the question

    Define exact prompts, buyer intent, engines, models, network point, and collection window before intervention.

  2. 02

    Capture the source record

    Store raw responses, screenshots, request metadata, hashes, and the site state used during the run.

  3. 03

    Anchor the intervention

    Separate preparation, announcement, deployment, first verified crawl, and subsequent changes.

  4. 04

    Rerun like for like

    Use the same prompt panel and disclose any model, prompt, geography, or retrieval change. A different question is a separate observation.

  5. 05

    Publish the full sequence

    Include upward results, dips, omissions, refusals, inconsistent engines, and metrics that did not move.

  6. 06

    Sign the statement

    Sign the final publisher summary only after the source pack is reconciled. State that the signature authenticates the statement, not its truth.

Measurement boundaries

One scorecard. Separate measures.

GEO is site-side readiness for ingestion, verification, reasoning, and inference in one visit. It is not a synonym for whether an engine mentions or recommends a brand.

GEO readiness

Whether the site is ready for AI ingestion, verification, reasoning, and inference in one visit.

Mentions

Whether an answer names the brand.

Recommendations

Whether an answer advises a user to consider the brand.

Citations

Whether an answer links to or attributes the brand's material.

Sentiment

How the generated answer characterizes the brand.

Crawl

Whether and how often identified AI systems retrieve the site.

Referrals

Visits sent from an AI answer or product.

Signature verification

Integrity is necessary. It is not independent validation.

Ed25519 verification confirms that the payload was signed by the key published by AnswerShare and that the payload has not changed. It cannot confirm a model response, capture date, ranking, or causal effect without the underlying source material.

node verify.mjs https://answershare.com/receipts/lavidge.signed.json
Download verifier Public key

Operating limit

AI outcomes are probabilistic. Engines can diverge, rankings can move, and the same prompt can produce different answers across users, devices, caches, and retrieval windows.

A public evidence pack documents what happened under stated conditions. It does not promise that the result will persist or that the intervention caused every observed change.

Review protocol changes