Methodology

Last updated: October 25, 2026

How we evaluate AI products, platforms, and models — and where the limits of that evaluation sit.

Evaluation sequence

1. Documentation review. Capabilities, limits, pricing, data handling, and terms as published by the vendor.
2. Hands-on testing. Where we have access, we run the product against representative tasks and record what it actually does, including failure modes.
3. Cross-checking. Vendor claims are checked against independent benchmarks or reporting where they exist. Where none exists, we say so rather than inferring.
4. Comparison. Products are compared on the dimensions that matter to the stated use case, not on a single aggregate score.

What we do not claim

• We do not claim exhaustive benchmark coverage. A test we did not run is not reported as a result.
• We do not extrapolate from one task to general capability.
• We do not treat vendor-published benchmarks as neutral evidence; they are reported as vendor claims.

Reproducibility

Where an evaluation depends on a specific configuration (model version, plan tier, date of test), the article names it. AI products change quickly; a result is only meaningful alongside the version and date it was produced.

Financial and investment content

Articles discussing markets, portfolios, or investment tooling are research and commentary. They are not investment advice, and they do not constitute a recommendation to buy or sell any security. Where we describe a strategy, we state its assumptions and its historical drawdown rather than presenting returns in isolation.

Corrections

If a method description here does not match what an article actually did, that is an error on our side. Report it via the corrections process.