Start with a reader’s actual decision
A useful review answers a specific question: whether a chat can preserve a character, whether a video option fits a budget, or whether a local interface still sends text to a cloud model. That question determines the sources, observations and comparisons we need.
We do not apply the same feature checklist to every product. A model runner, hosted companion and image generator have different prerequisites and different ways to fail.
Separate sources, observations and suggested checks
We distinguish what a provider documents from what we actually observed. A proposed trial is useful advice, but it is not reported as a completed test.
- Official information: product pages, documentation, pricing, policies and dated release notes, with conditions kept beside the claim.
- Direct observation: a real action under stated conditions, such as opening a public control or using a particular workflow.
- Suggested check: a step a reader could try, clearly presented without an invented result.
- Unresolved claim: a question the available evidence does not settle; we qualify it or leave it out.
State what each test can and cannot prove
A genuine observation needs its date, access level, relevant setup, action and result. We record practical limits such as guest-only access, no paid account or a single desktop session.
Opening a saved transcript does not isolate long-term memory. Finding a delete button does not establish erasure from backups. Reading a privacy policy does not amount to a network or security audit.
Paid access is necessary when the claim depends on a paid feature. Paying for a plan alone does not make the review comprehensive. We do not invent output, speed, privacy or billing results when access was limited.
Read shortlists as starting points, not benchmark scores
Category lists organize products around a use case. Some homepage positions are owner-selected featured commercial placements. Their order is not a measured comparison of output or a substitute for reading the reviews.
We do not give a researched listing a performance score just because its documentation was checked. Any numerical assessment needs an explained method and evidence that supports the dimension being scored. Commercial placement does not provide that evidence.
Recheck material changes and preserve real dates
Prices, plans, interfaces and content rules can change after a review. A revision should update the affected advice and make clear which source or observation it relies on; it should not imply that every feature was retested.
We change a page’s update date for a substantive revision. Original observation dates remain attached to those observations. When a claim is wrong or unsupported, we correct or qualify it rather than renewing a date alone.