Corrections and disputes

A vendor-neutral protocol for correcting factual errors and disputing measured visibility results, including record fields, re-run rules, and an append-only log format.

Use the correction record schema · Read the measurement methodology

RankRift does not publish measured results or run an intake service, so this page has no correction history to show. It sets out the protocol any publisher of AI-visibility figures should follow, and the record to check one against.

In short

This protocol separates factual corrections, stale-page notices, retractions, and disputes about a measured result. A publisher following it appends dated notes, preserves the prior claim, links the evidence, and re-runs a disputed measurement under its published method.

Who this is for

Brand, communications, and reputation managers evaluating a published result that may misrepresent their company
Vendors whose product appears in an evaluation or comparison and who think a description is factually wrong
Practitioners designing an intake route for broken formulas, stale rows, or unsupported links
Journalists and researchers checking a publisher's correction and versioning method
Readers who want to know what happens to a finding when someone powerful objects to it

When to use this

When a specific number, date, model version, or statement on this site is factually wrong
When a measured result about a brand looks inconsistent with a separately recorded observation
When a page has clearly gone stale
Not when you disagree with what a model said about you — that is a real problem, but a different one, and the section below explains why

What counts as a correction

A correction fixes something that was wrong: a miscalculated figure, a wrong model version, an incorrect run date, a misattributed source, a link that does not support the claim it is attached to, a description of a product that is factually inaccurate, or a statement that was accurate when published and has since become false.

Not everything that changes is a correction. Routine updates — adding a section, refreshing a reference table on schedule, re-verifying a page and finding it unchanged — are recorded through the page's modification and last-verified dates rather than logged as corrections. Reserving the word for actual errors is what keeps the log meaningful.

Material correction — the substance of a claim, figure, or recommendation changes. Noted on the page and logged in the central correction record
Minor correction — typography, a broken link, a formatting error that does not change meaning. Fixed, and noted on the page if it touched anything a reader might have quoted
Stale-page marking — a page that fails re-verification is corrected or visibly marked as no longer verified, never left standing silently
Retraction — a finding that cannot be supported is withdrawn, with the original text and the reason kept visible rather than deleted

Correction-publication requirements

Append a correction in place on the page that carried the error, with its date and a short description of what changed and why. Do not silently overwrite the wrong version; a reader who quoted it needs a traceable record.

Log material corrections in a central record as well. If a corrected figure appeared in a dataset, issue a new dataset version and update its citation string and version note.

If a correction changes a number that shipped with a dataset, the dataset is re-issued rather than edited in place, so existing citations remain traceable.

Disputing a measured result: re-run rather than argue

When a brand disputes a published measurement, the response should be a re-run under the same published method, not a debate about impressions. The re-run record identifies the original run, frozen prompt-set version, current model identifiers, dates, sample size, variance, and raw evidence.

This matters because the most common source of disagreement is not error at all: you have been asking assistants about your own category, probably in your own phrasing, from your own location, in sessions with your own history, and getting a different picture than a sampled prompt set produced on fixed dates. Both observations can be real. Only one of them has a sample size.

Required intake and operator actions

The more specific the challenge, the faster and more useful the resolution. Vague objections take longer and usually end up somewhere less satisfying for everyone.

A publisher should expose an accessible intake route with no purchase requirement. The request form should collect only what is necessary to identify and reproduce the disputed claim.

Intake fields: exact page and figure, alleged error, alternative observation, model or interface, prompt wording, approximate date, and region where relevant
Optional evidence: the challenger's prompt phrasings and a reason the published prompt set may not represent the category
Operator action: compare the request with the raw responses and verify that the published provenance matches the actual run
Operator action: re-execute the relevant prompt-set slice, recording model versions, sample size, dates, variance, and reviewer
Publication record: append the dated outcome to the original page and correction log, including when the original result stands
Integrity rule: do not change the method, prompt set, or sampling to produce a preferred answer for one brand

The outcomes a dispute can have

There are four honest endings, and one of them is that nothing changes. Saying that up front is fairer than implying every dispute is a step toward removal.

The original result was wrong. The figure is corrected, the page carries a dated correction note, and it is logged as a material correction
The re-run confirms the original result. The original figure stands, and the confirming run is published alongside it with its own dates and sample size
The result has changed since publication. Both figures stand as dated observations — the original was accurate on its run dates, and the newer run is added rather than replacing it. Model behavior moving over time is a finding, not an error
The prompt set was genuinely unrepresentative of the category. The set is revised and versioned, the method changelog records why, and every affected result is re-run — not only the one that was disputed

A measured visibility figure is a statement about what a model output contained on given dates. It is never a judgement about the quality of a product or a company, and it should not be read or reported as one.

A factual error is not the same as disliking what a model said

These may arrive through the same route but need different responses. A factual error means the publisher measured or reported something incorrectly: a wrong figure, model version, date, or product description. The correction record applies.

The other case is a correctly recorded assistant output that a subject dislikes. That may be a real problem, but it is not a reporting defect, and changing the study page would not change what an assistant says.

If an assistant is stating something false about your brand, the useful work is upstream: document it, identify the sources the answer is likely drawing on, correct those sources, and re-measure to see whether it moved. That sequence is a playbook rather than a correction request, and the risk pillar covers it — along with when the situation is serious enough to involve qualified counsel. Nothing here is legal advice.

Protocol integrity rules

These rules make the protocol auditable regardless of who requests a change.

Do not remove an accurate result merely because a subject dislikes it
Do not alter placement, ordering, or inclusion in exchange for payment, advertising, sponsorship, or another commercial relationship
Do not run a bespoke, flattering prompt set for one brand and present it as comparable to the published study
Do not suppress a reproducible finding because it is commercially inconvenient
Do not delete correction history; append and version it
Do not treat a legal threat as measurement evidence; route legal questions to qualified counsel and keep the factual record explicit

The correction log

A correction log starts with its first real correction and keeps entries in reverse chronological order. Never seed it with plausible events; an empty log is honest, and an invented one is not.

Required fields: correction date, affected page or dataset, prior claim, corrected claim, reason code, and supporting source
Research fields: original run identifier, prompt-set version, model identifiers, re-run dates, and replacement dataset version
Dispute fields: requester relationship, whether a re-run occurred, and whether it changed or confirmed the original result
Retraction fields: withdrawn claim, reason, date, and durable record location

Each affected page should also carry its own dated correction note, so the central log is never the only record of an amendment.

Frequently asked questions

How do I dispute a result about my brand?

Use the publisher's documented intake route and identify the page, figure, alleged error, alternative observation, model, prompt wording, approximate date, and relevant region. A conforming process checks the raw evidence, re-runs the relevant prompt-set slice, and records the outcome without requiring a purchase.

What if the re-run confirms the original result?

The original figure remains, and the confirming run is recorded beside it with its own dates, model identifiers, sample size, variance, and raw provenance. A few personal chat sessions are not directly comparable to a sampled prompt set run under fixed conditions.

Can I pay to have a result removed or changed?

No. Placement, inclusion, and removal are not for sale in any direction, and no commercial relationship changes what a measurement found. Sponsors are ineligible to appear in any study they sponsor, and an advertiser has no more standing here than any other reader. The only routes to changing a published result are demonstrating a factual error or a re-run producing a different measured outcome.

An assistant is saying something false about our company. Can you fix that?

Not directly, and neither can anyone else selling you a tool — no product controls what a model outputs. What helps is documenting the false statement with dates and model versions, identifying the sources the answer appears to draw on, correcting those sources at origin, and re-measuring to see whether it moved. The correction playbook on the risk pillar sequences that work. If the statement is damaging, this is general information rather than legal advice and you should consult qualified counsel.

Why does this page not show a correction history?

Because RankRift does not publish measured results or datasets, there are no result corrections to log. A publisher following this protocol should link its actual append-only correction history and keep dated notes on each affected page; invented entries are not evidence of maturity.

What response-time claim should an operator make?

Publish a measured acknowledgement interval and a separate resolution or re-run interval based on actual operating data. Do not invent a service level. If evidence is incomplete, send a dated status update within the stated window rather than implying that a re-run occurred.

Apply the correction protocol

Record the disputed claim, original provenance, re-run conditions, outcome, and dated correction note.

Read the methodology · Use the research specification · Read the editorial standards