An answer can contain outdated information from learned parameters, retrieved pages, or supplied context. Maintaining current pages helps retrieval, but it does not directly update a deployed model. Test these conditions separately.OpenAI publishes per-model knowledge cutoffs in its models documentation; other providers list theirs in model cards.
Prioritize facts that change
A model may repeat an older price without search. With search, it may still retrieve an outdated page or misread current information. Record the claim and the available sources before assigning a cause.
Prioritize facts that change customer decisions: prices, availability, product features, service areas, ownership, and eligibility. Review the sources you control and request corrections to material third-party errors.
Finding and correcting outdated answers
Ask the tested products about those facts. Repeat the questions under recorded conditions, with and without search where possible. Save the incorrect statement, expected fact, source, model, and date.
Correct the authoritative page and check that it can be retrieved. Then repeat the original test. If errors persist, inspect the sources and context that are visible. There is no guaranteed publication-to-correction interval for generated answers.
Update signals retrieval systems actually use
Use honest publication and modification dates. Change an update date when the content receives a substantive revision, and keep sitemap modification values consistent. A new timestamp alone does not make old information current.
When a page needs an update
A new article can provide current facts for retrieval. Whether a particular answer uses that article needs source evidence. Its publication does not establish inclusion in a model’s training data.
Review stable explanations when the subject changes or a problem is found. Do not rewrite them merely to appear recent. Spend regular maintenance time on pages whose facts change.
A maintenance calendar a small team can sustain
- When facts change: update pricing, product, and availability information as part of the release process.
- Periodic review: check specifications, business facts, and statistics according to how quickly they can become outdated.
- Answer checks: repeat important factual prompts after relevant source or model changes.
- Site review: correct outdated pages, retain accurate explanations, and redirect or remove obsolete pages with appropriate checks.
Assign an owner and review frequency according to how quickly each fact can become wrong. Trigger an immediate review when a product, price, or policy changes.
Use changes in the business to schedule updates
| Change | Pages or records to check | Completion evidence |
|---|---|---|
| Price or fee | Pricing, FAQs, offers, supported markup, feeds | Matching approved values and conditions |
| Service area or hours | Location pages, booking details, business profiles | Operating team confirms the published information |
| Product release or retirement | Product pages, comparisons, documentation, feeds | Current availability and replacement guidance |
| New research result | Article, tables, charts, downloads, citations elsewhere | Matching dataset version and calculation |
Record the factual change and publication time. Update a visible revision date when the material changes. Changing a date while retaining stale information does not complete the task.
In the Example Cooling audit, the fee correction applies to the old FAQ as well as the current service page. A later stale answer should lead to a check of its cited source and retrieval date, where available.