How the scores work

Star ratings alone are a bad way to rank software. Everything bunches up between 4.2 and 4.8, and a tool with nine glowing reviews outranks one with nine hundred. Head-to-head votes fix that — asking “which of these two is better?” produces sharper separation — but they need volume before they settle.

So every board blends both:

• A trust-weighted average of star ratings, smoothed toward a prior so a handful of reviews cannot spike a score.
• An Elo rating replayed from every head-to-head vote ever cast, the same idea chess uses.

A new tool with one five-star review is not the best AI in the world. Each score starts anchored to a prior of 5.0 carrying the weight of 4 average votes, and real reviews pull it away from there. The more genuine ratings a tool collects, the less the prior matters.

The more someone puts into saying how a tool went, the more it counts.

Articles are the only contribution a person checks before it counts. We check that it holds up and that it is really about using the tool. We do not run it through an AI detector: those are unreliable, and a badge we could not stand behind would be worth less than nothing.

You can rate a tool separately for each thing you used it for — ChatGPT for planning a trip, and again for writing code. That is what the task rankings need. But on the overall ranking it would let anyone who rates ten tasks count ten times, so everything one person says about one tool is combined into a single vote.

Your overall rating — the one not tied to a task — makes up 50% of that vote when you have given one, and your task ratings make up the rest. It is the one rating that answers exactly what the overall ranking asks, so your own specifics should not outvote it. Rating more tasks makes your vote better informed; it does not make it louder.

Articles are the exception: each one counts on its own, because a second article about the same tool is a second piece of work.

These tools change constantly, so a rating fades as it ages: it is worth half as much after 180 days on the overall ranking. On task rankings it fades twice as fast — half after 90 days — because how good a tool is at one particular job can change overnight when a new model ships, while its overall character moves slowly.

A review can be edited for 30 minutes after it is posted. After that it is locked, and a change of mind is added as an update underneath, with the old rating left in view. Changing the stars in an update makes the review recent again; adding words alone does not, because the opinion has not changed. Correcting a locked review, or deleting one at all, goes through a moderator.

Accounts earn influence by contributing. A brand-new account carries a fraction of the weight of someone with a long, well-regarded history, which makes mass-registering accounts an expensive way to move a number.

Head-to-head votes are weighted the same way: a vote from a trusted account moves Elo further, because the K-factor of 24 is scaled by that same multiplier. Each person gets one vote per pair, per task: voting again on the same pair changes your vote rather than adding another.

What people actually say about a tool is the point of this site, so star ratings carry at least half of every score from the very first reviewer. Head-to-head votes carry the rest, and they matter most at the start, when there are only a few ratings to average. As more people rate, their share shrinks further.

With no reviewers yet head-to-head can carry up to 50% of a score. By about 49 reviewers that settles at 15%, with ratings carrying the rest.

That is the most it can carry. How much it actually carries for a given tool depends on how many head-to-head votes that tool has been in: half of the full share at 10 votes, most of it by a few times that, and none for a tool nobody has voted on yet, which is then scored on its ratings alone. Without this, one vote from one person could move a tool fifteen places.

Tools are ranked on the score as shown, to one decimal, so two that both show 5.0 share a place. A difference too small to see does not move anything past anything.

A task like write a blog post has no reviews of its own, so its ranking starts out derived from the ability scores it depends on — writing, reasoning, speed, value, each with a published weight shown on the task page. That works immediately, from data that already exists.

For this, each ability starts from how people rate the tool overall, and ratings of that ability move it from there, weighted by trust and age like every other rating. A tool nobody has rated on coding is not assumed to be average at coding; it is assumed to be about as good as people say it is in general, until they say otherwise. Its overall rating counts as much as 4 ratings of the ability itself.

Then it inverts: as people rate tools for that specific task and vote head-to-head on it, their verdicts progressively replace the derived estimate, down to a floor of 25%. Around 40 ratings and votes a task is mostly community-driven. Each task matures at its own pace.

Every row says which of the two is placing it. The label counts only what was said about that task — ratings of an ability inform where a tool starts, but they are not counted as evidence that anyone used it for the task, because they are not.

An AI company can claim its page and correct its own description, tools and pricing — the details they know best and that go stale otherwise. Every claim is checked by a person before it is granted. They cannot touch their score, their reviews, or the listed strengths and weaknesses.

They cannot rate, review or vote on their own page either — not the stars, not a head-to-head their tool is in, not even marking a flattering review helpful. The database refuses all four, and approving a claim removes any rating, vote or helpful mark that account left on the page before it claimed. Replying to reviews is theirs, and the reply is badged as the provider speaking.

That is not a promise, it is a property of the system. A provider cannot write to their page at all: their edits go into a private draft, and publishing it runs one database routine that can only ever touch four marketing fields, their tools and their pricing. The score, strengths and weaknesses are not writable by any signed-in account, and the ranking job reads none of the fields a provider can edit. Every edit a provider makes is recorded. We take no payment for placement, and there is no mechanism through which we could deliver it.

Every hour the whole thing is rebuilt: Elo replayed from the first vote onward, averages recomputed, every board re-ranked. Nothing is incrementally patched, so a score can always be reproduced from the raw votes. Snapshots are kept, which is where the movement arrows and sparklines come from.