How we test and score AI tools.

Three things carry every rating: the vendor's own documentation, the vendor's own pricing page, and the public record of what users report.

Pricing is the strictest of the three. We read every price off the vendor's pricing page ourselves and stamp the date we read it, on the page, next to the number. AI pricing moves constantly, so an unstamped price is a worthless price. If you find a stamp older than you are comfortable with, treat the figure as stale and check the vendor.

Capability claims come from the vendor's documentation, compared across every tool in the category on the same criteria, so the comparison is like for like rather than a list of whichever features each vendor markets hardest.

We do not claim hands-on testing we have not done. This is the single most common lie in the "best AI tools" genre, and we would rather lose the comparison than tell it.

Where a first-party screenshot or a hands-on run would genuinely strengthen a claim, the page marks it as pending capture instead of implying we already ran the app. Where a criticism comes from users rather than from us, the page says so and attributes it, so you can weigh it as reported experience rather than as our verdict.

Ratings and skip-it calls are our judgment applied to that evidence. They are opinions, formed in public, on stated grounds, and you are equipped to disagree with them.

We read the threads where practitioners argue, and we link them so you can read them too. We also read them sceptically. A recommendation from an account that turns out to be the product's founder is not a recommendation, and a thread whose top answer carries a discount code is an advertisement.

When we cite community consensus, we say which community and link the thread. When a widely cited "I tested 15 tools" post is affiliate-funded, we say that too, because knowing who paid for a recommendation is usually more useful than the recommendation.

Every tool page names the people who should not buy the tool. This is the load-bearing part of the method, and it is the sentence a commission-funded list is structurally unable to write.

A tool that is right for a solo founder is often wrong for a regulated team, and saying so costs us nothing because we are not paid on your purchase. If a category has no tool we would actually recommend, the page says that instead of crowning a winner.

Tool pages carry a public changelog. Re-verification passes are logged there with their date, including the ones that corrected our own earlier drafts.

Corrections are made on the original page, with a note, never as a quiet edit. If a vendor thinks we have a fact wrong, the fastest route is to email with the evidence. Price and capability errors get fixed within 24 hours.

The rules we will not break, including how sponsorship and partner links are kept away from ratings, are set out in our independence charter.