How we source and verify what we publish

1. The vendor's page is the source

Before an entry is written, we fetch the tool's own site and its pricing page and treat that text as the only authority. Where it disagrees with anything else, including what a language model believes about the product, the page wins.

This exists because the alternative failed. Entries written from model recall contained real errors: a proprietary editor described as a free open-source extension, an invented free tier, and a listing that described an entirely different product than the one it named. Reading the vendor's page first is what stopped that.

A page fetch that succeeds but returns an error screen is discarded too. A clean fetch of a bot challenge or a deployment error is more dangerous than a failed one, because the text is real and comes from the right address, so it gets trusted completely. One listing was published describing a hosting platform's error page before that check existed.

Every fetch is pinned to a US exit. Every price here is a dollar figure, so the page we read has to be the page a US visitor sees; an unpinned read once returned euro pricing and reported our correct dollar figure as missing.

2. A model proposes, string matching disposes

A language model reads the fetched page and proposes structured facts. Nothing it proposes is published on its say-so. Each claim is then checked mechanically against the source text, and anything that fails is dropped rather than softened:

• A price tier survives only when the vendor's text contains both that tier's name and its figure. A tier whose figure cannot be confirmed is dropped whole, never published with the number removed.
• A plan limit ("300K tokens daily limit") survives only when the model copies a verbatim span of at least 16 characters from the page showing that cap. This is the most invention-prone field there is.
• A contradiction found during a re-check is reported only with a verbatim quote of at least 20 characters that is then confirmed to occur in the source. A quote that does not appear is discarded, so a model that starts inventing goes quiet instead of accusing correct entries.

The failure direction matters more than the hit rate. A false "this entry is wrong" costs more than a missed catch, because it destroys the signal the checks exist to produce.

3. "Unknown" is a real answer

Where a vendor does not document something, we say so rather than filling the gap. Capability rows render "Not listed", a tier with no published figure renders "Custom", and a tool whose pricing we have not confirmed says exactly that instead of showing a number nobody printed. Nothing on this site computes a figure: no currency conversion, no annualising, no dividing a yearly price by twelve. Every number traces to a page we read.

4. Weekly re-checks, published states

Published entries are re-read against the same vendor pages every week. Each entry shows one of four states and links the exact pages read, and the current tally across 107 tools is:

• Verified (91): the re-read ran on the date shown and the entry still matched. It means that and nothing more: not that we tested the product, and not that nothing has changed since.
• Under review (14): the re-read ran and did not confirm everything. The entry keeps the last figures we did confirm and says so, because inventing a replacement is worse than admitting the gap.
• Unconfirmed (2): the vendor blocks automated reads, so nothing could be checked. That is our side failing, not a finding about the tool, and it never counts against the entry.
• Pending (0): no check has run yet.

These counts are published rather than kept internal. A directory that only ever shows green has either checked nothing or is hiding the misses.

5. Changes are recorded, not quietly fixed

When a tracked figure moves, both values, the date, and the page we read are recorded in the change log, 13 entries so far. Rewriting a published price needs more evidence than reporting it: a fresh read must show the new figure beside that tier's own name, the old figure must no longer be on the page, and a second read must agree. A tier that simply fails to appear in one read is never deleted, because a page that rendered differently is indistinguishable from a discontinued plan.

6. What we do not publish

No ratings, no scores, no testimonials, and no aggregateRating markup. We have not used every tool listed here, so a score would be dishonest, and fabricated review data is penalised by search engines besides. Lists are ordered by disclosed mechanical rules, such as free plan first then lowest entry price, never by preference. Comparison pages state facts on both sides and do not declare a winner.

Entries publish automatically each day, and we say so plainly rather than implying a person hand-tested each one. Affiliate links are disclosed and never decide which tools appear, how they are described, or what they are compared against.

Corrections

If something here is wrong, tell us and we will correct it and re-check the entry. The "See pricing" link on every tool page goes to the vendor's live page, which is the only figure that is ever fully current.

Questions