See how ChatGPT, Perplexity and Google AI Overviews describe you today
Free AI Visibility AuditWe publish nothing without a source link, we do not republish other people's tables, and we say which of these entries a person checked.
Prices, context windows, release dates and deprecations are read from the provider's own catalogue, documentation or announcement, and every published entry links to it. We read other trackers to notice what we have missed, and that is all we do with them: a lead from one opens a task to go and check the provider, and our software refuses to publish it as anything else.
Model catalogues are re-read every three hours, provider pricing pages every six, and the cross-checks daily. Each source records when it last answered and whether it answered cleanly, and a source that has been failing shows as failing rather than as quiet.
When it happened, how precisely we know that, when we detected it, and when we last re-read the source and found it unchanged. They are kept apart on purpose: a re-check that finds nothing new updates the verification stamp and does not touch the page's modified date, because telling a search engine a page changed when it did not is a claim that gets discounted, and deserves to be.
A title reading "(September 2026)" means the page was verified in September 2026, not that it was generated then. If a page has not been successfully re-checked inside the current month, the month does not roll and the page says when it was last confirmed.
Only a price change read from a structured catalogue, and only when the move is within a plausible range. A change larger than fivefold, a move to or from free, a page whose content hash shifted, and anything written from a news feed all wait for a person. A hash tells us a page differs; it does not tell us what differs, and a footer year rolling over produces the same signal as a price cut.
A model page is created as soon as the model appears in a provider's catalogue. Whether it is offered to search engines is a separate decision, made on whether anybody searches the model's name. Pages that do not clear that bar stay useful, stay linked and carry noindex, because a folder full of pages nobody searches for makes the pages people do search for harder to find.
A score is labelled verified when we read it on the benchmark's own published results, and provider claim when the only source is the company that made the model. Both are shown. Dropping everything we cannot independently verify would tell a reader less than showing it and saying which is which, and it would quietly favour the labs that publish least. Where sources disagree the row is marked disputed and both figures stay.
A model page reports eight categories, weighted: agentic 22%, coding 20%, reasoning 17%, multimodal 12%, knowledge 12%, multilingual 7%, instruction following 5%, maths 5%. They are weighted for what a SaaS team buys a model to do, which is why agentic and coding lead and maths does not: a model that wins a maths olympiad and cannot use a tool is no use on a Tuesday. Argue with the weights, they are ours rather than anybody's standard.
Where a model has no published evidence in a category, the category is marked not measured and left out of the headline figure, which is then re-weighted across what remains. Scoring an absence as nought would rank a model that was never tested below one that was tested and did badly, and we have no basis for that. Every page shows how much of the weighting its score actually rests on, next to the score.
Two sources today. Artificial Analysis run their own evaluations and publish the primary metrics, which is where MMLU-Pro, GPQA, LiveCodeBench, AIME, MATH-500 and Humanity's Last Exam come from; their data is used under their free API terms, which require attribution, and every page showing one of their figures credits them. SWE-bench supplies its own leaderboard, including whether it has checked a submission's logs, which is what decides whether we label a row verified or a submitter's claim. We do not import anybody's composite index: that is their weighting of the same tests, and taking it alongside the tests would count one piece of evidence twice.
The leaderboard lists models we hold published benchmark results for, and nothing else. An untested model does not appear at the bottom, because appearing at the bottom is a claim that it was measured and found wanting. This is also why the ranking is small next to the number of models tracked: we would rather publish a short ranking that is checkable than a long one that is mostly inference.
Both are published on every model page and neither feeds the ranking. A table built from prices and context lengths sorts cleanly, updates itself and answers a different question from the one a leaderboard is asked. Where a benchmark score does not exist, the honest output is an empty row, not a proxy.
A wrong number is corrected on the page it is on, and the change is logged as its own entry with the date. Nothing is quietly edited: a tracker whose history can be rewritten is a tracker with no history.
Found something wrong? Tell us and we will check it against the source the same day.
The same discipline behind this tracker is what we use to audit whether AI answers cite you: named sources, dated evidence, and a number you can check.