# How GovernBench reads AI governance software > Source: GovernBench — https://governbench.com/methodology What we read on each maker’s site, how a punch is struck, how the lists are ordered and how to report an error. ## The readings so far Each tool is read on its maker’s own site against the same three frameworks and its price, then read again on a cycle. ## What we read Only the maker’s own site: product pages, documentation, trust and pricing pages. Never an analyst, a partner or a sales call. For each tool we record what it maps of each framework, the governance work it does, how it is sold and where it runs, each with its source. ## How a punch is struck A framework is struck solid when the site shows ready controls, templates or a mapping for it; outlined when it covers part of it or only names it; dashed when the site shows nothing. A blog post about a framework is not a mapping, and a maker’s own certificate is not one either. ## Order **Featured** tools first, labelled, then each list ordered by what the maker’s site lets us confirm: the frameworks it maps, a published price, documentation and stated limits. Paying never changes the order under the Featured places. ## Not for sale No punch, rank or verdict on GovernBench is for sale. Paid places carry a **Featured** label; our [editorial policy](https://governbench.com/editorial-policy#not-for-sale) sets out the rules. ## Reading again Each tool is read again on its own site on a cycle, and its listing changes when the site does. ## How a punch is struck - **Yes**: Solid: the maker’s site shows controls, templates or a mapping for it. - **Partly**: Outlined: it covers part of it, or names it without a mapping. - **No**: Dashed: the maker’s site shows nothing for it. - **Not stated**: Dotted: we could not read an answer on its site. ## How we count the figures ### In our articles - Every number in an article we write comes from our own records of each tool or from a source the article cites, and is checked against them before the article goes out. - Prices are read from each tool's own pricing page, and each ranking shows the days we read them. - The typical price is the middle value of each tool's cheapest paid plan, so free plans stay out of it. If fewer than 5 tools have a price, we give no typical price. - Counts such as how many tools have a free plan use the same records and rules as our category pages, so an article and its category give the same answer. ### In our price tracker - Once a week we read every tool's own pricing page. No AI model reads the page. - A price change counts only when the same plan, in the same currency and billing period, shows the old price and then the new one, each in two readings at least a day apart. - A move of more than 5 times, up or down, doesn't count: at that size it is almost always a different plan or unit, not a new price. - Each change shows the price as the pricing page prints it, not converted or rounded. - Our price changes page lists the last 13 weeks, newest first. Fixes to our own records are not price changes. ## Who runs GovernBench GovernBench is written and run by [Team GovernBench](https://governbench.com/authors/research). Every listing, guide and report is published under the site's own name, and the team answers for all of it. ## Changes to a listing Changes a maker asks for are made only on a paid listing. A listing that has never been paid for shows what our own data says. Anyone can report a wrong fact with “Report an error” on a listing, at no cost. We fix it once the tool's own website confirms it. Paying buys changes and speed, never accuracy. ## Where we use AI AI models do the reading and the first drafts. A model reads each tool's own website and drafts its listing. Our articles and guides are drafted by a model from our own data. Checks decide what goes live. Every fact on a listing must point to its source on the tool's website, and a sentence that website does not support is cut. In an article, every figure must match our data and every quote must appear word for word in its source. A second reading checks a sample of facts against each tool's own pages. An article that fails a check is not published.