Open evidence template · v1.0

Record AI-answer evidence before you summarize it.

A general-purpose observation ledger for preserving prompts, answer context, citation occurrences, source roles, verification states, and reviewer decisions in a reproducible format.

Boundary: one ledger row is a dated observation, not a stable ranking. The template does not query an AI system, judge factual accuracy, infer causation, or claim that a source influenced an answer.

Download the ledger package

The package keeps the blank working template, a clearly synthetic example, and structural validation rules separate. All example entities, prompts, URLs, and answer text are fictional.

Blank CSV templateHeaders only, ready for a controlled observation run. Synthetic example CSVThree fictional rows demonstrating occurrence-level annotation. JSON SchemaField types, enums, formats, and minimum requirements.

Why an occurrence-level ledger?

AI-answer research often jumps from screenshots or exports to totals. That shortcut makes it difficult to reproduce the prompt, distinguish mentions from citations, understand what a source supported, or revisit a judgment when a taxonomy changes. The ledger preserves the smallest useful unit: one citation occurrence within one dated answer observation.

Multiple citations in one answer become separate rows sharing the same observation_id. A mention with no visible citation can still be recorded as an observation summary elsewhere, but it should not be invented as a citation row. If a cited URL appears twice in one answer, retain both positions unless the research question explicitly uses unique URLs.

Collection workflow

1 · Freeze the prompt set

Assign stable prompt IDs and versions before collection. Preserve exact language, intended audience, task intent, and market.

2 · Capture observable context

Record product, answer surface, timestamp, market, signed-in state when relevant, and a privacy-safe answer excerpt.

3 · Preserve each citation

Store the exact observed URL, link label, citation position, mention status, and the claim or passage it appears to support.

4 · Annotate source use

Apply the versioned source-role taxonomy to the answer context. Keep ambiguity as needs_review.

5 · Verify separately

Later, record HTTP state, final URL, publisher canonical, and access constraints without overwriting the observed URL.

6 · Review before aggregation

Validate required fields, reconcile a sample between reviewers, and publish denominators with every summary.

Field groups

GroupCore fieldsReason
Observation identityobservation_id, prompt_id, prompt_versionConnect every row to one controlled prompt run.
Run contextanswer_system, answer_surface, run_timestamp, marketState where and when the variable answer appeared.
Citation occurrencecitation_id, citation_position, observed_url, link_labelPreserve the exact visible source occurrence.
Answer usemention_present, citation_present, answer_evidence_span, claim_summaryDocument what the answer visibly used the source to support.
Role annotationprimary_role, secondary_roles, publisher_relationship, taxonomy_versionSeparate source responsibility from source quality.
URL verificationverification_status, verified_final_url, publisher_canonical, verification_timestampKeep later network findings distinct from raw evidence.
Reviewreviewer_id, review_status, notes, is_syntheticMake uncertainty, examples, and later corrections explicit.

Roles and controlled values

The template uses the companion seven-role taxonomy: entity, definition, evidence, comparison, implementation, validation, and action. Use a semicolon-delimited list for secondary roles in CSV. Publisher relationship is one of owned, partner, independent, unknown, or not_applicable.

Verification status is not_checked, reachable, redirected, access_limited, not_found, or error. A challenge page, login wall, regional restriction, timeout, or rate limit belongs under access_limited or error; it is not automatic proof that the citation is invalid.

Quality checks before analysis

  1. Every citation_present=true row has an absolute HTTP(S) observed_url and positive citation position.
  2. Rows sharing an observation_id agree on prompt, system, surface, timestamp, and market.
  3. Every primary and secondary role uses the taxonomy version recorded on the row.
  4. Observed URLs remain unchanged; normalization and network verification use separate fields.
  5. Synthetic fixtures are labeled is_synthetic=true and never mixed into measured totals.
  6. Privacy-sensitive or copyrighted answer text is minimized; store a short evidence span rather than unnecessary full output.
  7. Ambiguous rows remain needs_review until a reviewer resolves or excludes them.

Report bounded findings

Useful summaries include observations collected, observations containing the tracked entity, citation occurrences, unique observed URLs, normalized URL keys, unique publisher hostnames, and source-role coverage. These are different denominators. Report the prompt-set version, systems, product surfaces, markets, collection window, reviewer method, missing data, and sample size alongside every chart.

What the ledger cannot prove

It cannot establish stable rank position, model training inclusion, publisher authority, factual correctness, search demand, commercial impact, or why an answer system selected a source. Repeated observations can describe change in the measured sample; they do not reveal an undisclosed model's causal mechanism.

Apply the AI answer source-role taxonomy, normalize exported URLs with the citation URL protocol, or review broader evidence-led AI visibility resources from Corank.

Version 1.0 · Published 2026-08-12 · Update the schema version when required fields, controlled values, or meanings change.