The short answer Resetting a baseline is not deleting an inconvenient result. It is a versioned decision about whether measurements before and after a material change still represent the same construct under sufficiently comparable conditions.
What this checklist does—and does not do A human selects the proposed historical treatment. The checklist records supplied evidence and surfaces contradictions, missing inputs, and downstream work. It does not inspect a provider's private code, infer an undisclosed change, approve the treatment, certify comparability, silently rewrite history, or attribute a score movement to brand performance.

Classify the change before redrawing the trend

A cosmetic chart edit is not the same as a new measured construct. Begin with field-level old and new evidence, then propose a treatment. The tool checks the proposal; it never chooses for the reviewer.

Continue with documentationEvery material field is unchanged or presentation-only. Preserve the method versions and record the non-substantive change.
Continue with a visible annotationA bounded non-critical change affects interpretation but does not replace the construct or critical population. State the effective date and limit.
Reset, separate, or restateA critical method or population changed, a new construct was introduced, or history can only be made comparable through a disclosed reproducible bridge.
Browser-local transition record

Build the baseline reset and series-break memo

Copy exact wording from method documents, exports, provider notices, and decision minutes. If evidence is unavailable, choose Undisclosed or unresolved instead of guessing.

01 · Record identity

Name the series, change, versions, and accountable reviewers

02 · Old-to-new method matrix

Preserve both versions and classify each relationship

Five rows are critical continuity gates. A bounded, fundamental, undisclosed, or inapplicable critical field conflicts with a continuity treatment until the reviewer resolves it.

Measurement fieldOld method evidenceNew method evidenceRelationship
Measured construct and unitWhat event or quantity the series claims to represent.Critical continuity gate
Formula, denominator, and weightsA stable label can conceal different arithmetic.Critical continuity gate
Prompt panel and taxonomyA changed question population changes the denominator and opportunity set.Critical continuity gate
Engines, products, surfaces, and access laneA provider upgrade or API-to-consumer move can change the instrument.Critical continuity gate
Market, language, account, and session conditionsLocale and session state can alter retrieval and generated answers.
Collection window, cadence, repeats, and aggregationA point estimate and a rolling multi-run index are different observations.Critical continuity gate
Entity matching and aliasesA broader alias rule can manufacture apparent visibility.
Mention, citation, recommendation, and accuracy classificationA linked citation is not the same event as a name in answer text.
Valid, null, failure, refusal, and retry rulesSilent exclusions can change both the numerator and denominator.
Source and URL normalizationCanonical URLs, domains, and source classes can move citation counts.
Competitor universe and category membershipShare metrics can move when the comparison set changes.
Uncertainty, variability, and material-change thresholdA threshold change can relabel noise as movement.
Raw evidence, export, and traceabilityA restatement cannot be reproduced from aggregate screenshots alone.
Version, correction, and revision policyPreserve which rule produced every published value.
03 · Human-proposed treatment

Choose a posture, then supply its required evidence

Historical-restatement evidence gate

Checking a box records that evidence was supplied; it does not independently validate the evidence. All seven are required before this tool can call a proposed restatement ready for human review.

04 · Downstream propagation

Inventory every place the series is reused

Assign an owner, status, and note for all ten output classes. A reset is not complete when one dashboard changes but an export, target, board deck, or contract still presents the old series as continuous.

Output classOwnerUpdate statusAction, location, or evidence
Dashboards
Recurring reports
Executive and board decks
Public benchmarks
Targets and commitments
Alerts and thresholds
Experiments and attribution logs
Contracts and SLAs
Exports and APIs
BI formulas and derived metrics
05 · Generate the review package

Expose conflicts before anyone approves the change

Start with the old and new method evidence. Unassessed fields remain visible as unresolved.

All entered details stay in this browser tab. This page has no upload endpoint and does not persist the record.

Three different changes need three different records

Use the examples to inspect the encoded gates. They are fictional and do not imply a real provider, organization, or measurement result.

Forward-only model updateA provider changes the answer surface and visible model. Preserve the legacy series, label the break, and establish a fresh baseline window.
Separate-series construct changeA mention rate is replaced by a weighted composite. Give each construct its own identifier and do not join the values.
Disclosed restatementA bounded formula correction can be reproduced for every affected period, with originals, bridge limits, propagation, and named approval retained.

Make the discontinuity visible wherever the series appears

A useful notice names the first affected period, the old and new concepts or methods, the cause, whether the effect is one-off or lasting, the estimated magnitude where supportable, and a link to full details. Do not visually join points across a known discontinuity as though the instrument remained unchanged.

Parallel runs have a boundary An old/new overlap can estimate the impact near that window. A linking factor derived from a short overlap may not generalize to earlier history, other models, other prompt populations, or other markets. Preserve the overlap conditions and limit the inference explicitly.

General series-break controls are adapted cautiously from the OECD Data and Metadata Reporting and Presentation Handbook. That handbook addresses official statistics; this independent checklist applies its presentation discipline to supplied AI-visibility records and does not claim OECD endorsement, conformity, or statistical certification.

Need to document one method before comparing versions? Use the AI visibility methodology disclosure generator. If two providers produced different headline numbers, use the separate AI visibility score reconciliation template.

Frequently asked questions

When should an AI visibility baseline be reset?

Consider a forward-only reset or separate series when a material change alters the measured construct, formula or denominator, prompt population, AI products or surfaces, or reporting window. Record the effective date, preserve the earlier values and method, and label the break in every downstream use.

Should historical scores be recalculated after a method change?

Not by default. A disclosed restatement needs preserved originals, reproducible old and new calculations, named period coverage, a documented bridge and its limits, downstream propagation, and authorized human approval. Otherwise preserve the prior series and begin a new baseline or separate series.

Does an AI model update always require a new series?

No. First determine what changed in the observed instrument. A presentation-only label change can preserve continuity; a bounded non-critical change may support annotated continuity; a new model, surface, prompt population, formula, window, or construct may require a visible break. The evidence, not the release-note label, determines the review.

Does “ready for human review” mean approved?

No. It means only that the form contains the fields required by the selected path and no encoded contradiction remains. A qualified human still needs to verify the evidence, authority, statistical reasoning, contractual consequences, and final communication.

Does this page upload or retain the record?

No. Classification, conflict checks, memo generation, copying, and download happen locally in the browser. Closing or refreshing the tab clears unsaved entries.