Methodology

How we decide what to build

Every candidate tool is scored out of 50 before any code is written, across five equally weighted factors. The score is shown next to each tool on the index page, so the reasoning is public.

Forty and above gets built this quarter. Thirty to thirty-nine goes to the backlog. Twenty to twenty-nine is built only if it completes a cluster that already exists. Below twenty is not built at all, however entertaining it would be.

What has to be true before a tool ships

• It loads and becomes usable in under 1.5 seconds on a mid-range phone.
• It works on a 360-pixel-wide screen, tested by hand, not by resizing a desktop window.
• It is better than the current top result at something specific and nameable — not “nicer”, but a stated capability the competitor lacks.
• It reaches the whole way through a real task without an account or a payment.
• Its limits are documented on the page before anyone hits them.

A tool that fails any of these does not ship. That is why the count grows slowly. Publishing five thousand pages in a week is possible and it is the fastest known route to a manual penalty.

How benchmarks are measured

Any timing shown in a technical notes section was measured, not estimated. Unless a figure says otherwise, the method is:

• Reference machine: Intel Core i5-6300U at 2.40 GHz, 4 threads, 8 GB RAM, Chrome 151, no extensions, a fresh profile.
• Procedure: ten runs, first two discarded to allow for JIT warm-up, median of the remaining eight reported.
• Inputs: generated to a stated size and shape, described alongside the number.

That reference machine is deliberately unimpressive — a 2015 ultrabook processor. Benchmarks run on the newest laptop the author owns flatter the software and tell you nothing about what the tool will feel like on the device you are actually holding. If a number here is acceptable on an i5-6300U, it is comfortable almost everywhere.

Your machine will still differ. The figures show the order of magnitude and are reproducible by the method above; they are not a promise of a specific millisecond count.

Where external numbers come from

Where a page cites a figure that is not ours — a reading speed, a cryptographic guess rate, a standards document — the source is named inline. Standards are cited by RFC or specification number so you can check them directly. We do not cite “studies show” without saying which study.

How the AI features work

Almost nothing here uses a language model. Formatting, hashing, counting and converting are deterministic problems with correct answers, and a model would make them slower, more expensive and occasionally wrong.

Where a model is genuinely better — rewriting, summarising, classifying — the output is labelled as AI-generated on the page, and the deterministic path is never removed. You will never be forced through a model to finish a job a function could do.

Corrections

If a tool gives a wrong answer, that is the most serious kind of bug this site can have. Report it to support@adeptbay.com with the input you used. Confirmed correctness bugs are fixed before anything else, and the fix is recorded in the changelog with the date and what was wrong.

Pages that carry a benchmark also carry a “last verified” month. Those are re-run quarterly. If a page says it was verified more than six months ago, treat the number with suspicion and tell us.

What we do not do

• We do not publish review scores, star ratings or “best of” rankings we did not measure.
• We do not add structured data claiming ratings or reviews that do not exist. Every rich-result claim on this site corresponds to something visible on the page.
• We do not buy links, exchange links, or place paid content without labelling it.
• We do not generate pages for keyword combinations with no real search demand, even though the tooling to do so exists in this codebase.