There are roughly 28,000 AI tools listed across the big directories. As far as we can tell, almost none of them have been used by the people listing them.
That is the entire reason this site exists. Here is exactly what we do instead, so you can check our working or tell us it is rubbish.
Getting into a category
A category opens when we can properly test enough tools to make a comparison mean something. Candidates come from what readers ask us, what the engines keep naming, what vendors send in, and what we already pay for ourselves.
Then a commercial filter runs, and we would rather admit it than have you find it: a tool needs an affiliate programme before it can appear on a review or a best-of page. That is what funds the subscriptions. It also means those pages rank a defined set, and every one of them says how big that set was.
The leaderboard ignores this filter completely. More on that below.
How the testing runs
Same brief, same conditions, same week wherever we can manage it. We buy our own subscriptions at list price.
Each category has a fixed set of tasks, published on the category page, usually nine, covering the jobs people actually hand money over for. Every task is scored on its own before any tool meets any other tool, so the one we happened to test on Monday is not judged more harshly than the one we got to on Friday.
Where quality is a matter of taste, we test blind. Voice samples get judged without the listener knowing which tool made which.
The leaderboard, and what it is not
Once a month we put the same buying questions to ChatGPT, Perplexity, Gemini and Copilot, and count who gets named. Every answer is saved word for word, with the date and the engine. The percentages link back to those saved answers so you can read them yourself.
Read this before you quote a number off it
It is a record of what four engines said, on those days, to those questions. It is not a verdict on any tool and it is not a measure of quality.
Engine answers move between runs. They move between accounts. They move between you and the person sitting next to you, because personalisation is doing work you cannot see. One run is one session, not a fact about the world.
Anyone waving an AI answer at you as proof that a company is good or bad is selling something.
Which is why we publish it as a time series. One month tells you almost nothing. Six months of identical questions starts telling you something real.
We also publish the runs where an engine hedged, refused, or named nobody. Dropping those would make the data look tidier and would make it a lie.
Dates on everything
Every score, price and engine answer carries the day we checked it.
Commission terms get read off the provider’s own page, never copied from another website. We learned that the hard way: the first two rates we checked at source were both wrong everywhere else on the internet, and both wrong in the provider’s favour.
Reviews get re-tested when a tool changes materially, when pricing moves, or when the score turns twelve months old. Stale pages get marked stale instead of being left to look current.
Things we will not do
- Sell a ranking position, a score, or a place on the leaderboard.
- Run vendor copy as though we wrote it.
- Borrow someone else’s review score and present it as ours.
- Pull a published review because a vendor got upset.
Corrections
Facts get fixed quickly and the fix is noted on the page. Disagreeing with a verdict is not a fact, and we will tell you that politely. See Corrections.