How We Evaluate AI Tools & Models
This page is the single reference for how TheBestSI labels its evidence. Every product card links here.
Our principles
- Say what the evidence is. Every assessment is labelled by how it was formed.
- Separate fact from opinion. Statements marked “Editorial view” are our judgment, not verified fact.
- Date what changes. Prices, plans, limits, and policies carry the date they were checked.
- Show our limits. We state what we did not test and what we do not know.
- Commercial relationships never set rankings. See our affiliate disclosure.
The three evidence labels
Requires: our editors tested the product themselves, and a published evaluation record is linked, covering the exact version, the test date, the tasks and number of runs, the scoring, and the limitations. Status today: no product qualifies, so the label is not in use.
Requires: the page lists the primary documentation, papers, or filings it relies on, with dated links, and its claims trace to them. Status today: our product cards do not yet cite sources inline, so none carries this label. Some guides and articles list sources and use it in spirit.
Means: a reasoned assessment from public information such as vendor pages and published research, with limitations stated. It is not our own testing. Status today: every product card carries this label.
The five evaluation dimensions
This is the framework we are building toward. We have not applied weights or published scores, and we will not display a score without a linked evaluation record.
- Domain task fidelity. Does the tool do the job it claims to do, reliably, without constant re-prompting?
- Context and state integrity. How well does it keep track of long documents, many files, and multi-turn sessions?
- Data governance and privacy. What do the vendor’s terms say about retention and training, and which plans differ?
- Workflow ergonomics. Does it fit the editors, command lines, and review processes people already use?
- Cost and pricing transparency. Are prices, limits, and renewals predictable and clearly stated?
Where our evaluations stand today
Rankings and comparisons are published as Editorial Analysis. Model comparisons use specifications and policies as documented by the providers, dated and linked, and we do not name a performance winner we have not tested. When we run our own tests we will publish the method and results on the evaluation records page and link them from the relevant product or comparison.
Commercial firewall
Affiliate or commercial relationships never determine rankings or inclusion. Where an affiliate link is used it is disclosed and carries the appropriate rel attributes.