Methodology

How It Works

Every week, we ask the same question to three AI providers: ChatGPT, Claude, and Gemini.

One prompt per category. The same prompt goes to all three providers.

Each provider returns its own ranked list of tools — independently, with no seed list. We don't tell the AI what to recommend. We ask, and we record.

The CLAI Score

CLAI stands for Composite Leaderboard from AI Intelligence. It's a score from 0 to 100 that measures how visible a brand is across all three AI providers.

The score is a weighted combination of four signals:

The percentages are formula weights — Average Position has the most influence on the final score, followed by how many providers mention the brand.

The Providers

We use the best available model from each provider.

What We Show

NR (Not Ranked) means the brand was not returned by that provider. This is not a failure — it's a signal. Some tools are invisible to certain AIs.

Trend arrows show rank changes between the current and previous weekly cycle.

Disagreement highlights appear when providers rank a brand at significantly different positions.

AI Sources

Beside the rankings, each category has an AI Sources page showing what AI engines actually read before recommending software. It is a different instrument, and it asks different engines.

Who we ask. Perplexity and Google AI Overviews — the two engines that hand back the pages they opened while answering. The rankings above come from ChatGPT, Claude and Gemini, which do not. What an engine reads on an AI Sources page does not explain a rank on a leaderboard.

What we ask. The category's existing question — the same one listed below — sent on its own plain call, from the United States, in English. Never inside the ranking prompt: the strict answer format that prompt imposes changes which pages get retrieved.

When an engine returns nothing. There are three different kinds of nothing and we ask again for all of them, up to 3 calls: the call failed, Perplexity answered without searching, or Google showed no AI Overview. Whatever the last attempt produced is what we publish, with the number of attempts printed beside it.

Four rules govern every number on those pages:

• Retrieved, never cited. The engines report the pages they pulled back, not the ones they leaned on. We never claim a page was cited.
• Per engine, never pooled. The two engines read largely different webs, so a combined page count would describe a list neither of them produced. The only cross-engine facts we publish are that both engines read a site and that both named a company.
• Counts, never rates. Every figure is a count of a named thing — “4 of the last 5 checks”, never “80%”. One question per category is a small sample by design, and a percentage computed from it would be false precision.
• Not measured is never a zero. An engine we could not read that week is labelled as ours and enters no count. “Google showed no AI Overview for this question” is a different statement, and a measured one: it is a property of that market, printed with how many times we asked.

What the page is about. One question, not a whole category. It says: when a buyer asks this exact question, here is what each engine retrieved and who each one named. We do not open the retrieved pages, and we make no claim about what any page says.

Brand identity is by domain only. Each check is stamped with its own methodology version, which moves independently of the ranking methodology version above.

Across every category. Version 1.1.0 of this dimension added two things to every check: what each engine did with an individual company — named or not, at what order of first mention, and which of the company's own pages it opened — and the classic results Google listed on the same page as its AI Overview. Both feed the AI Sources index, which reports every category at once. It shows one complete cycle — the newest week in which every category has a settled check, or a week that has closed — and never a blend of two, because a count that spanned two cycles would not be a count of anything anyone could cite. The week it is showing is printed on the page and in its citation line; earlier complete weeks are reproducible from the API.

What We Don't Do

• We don't accept payment for rankings.
• We don't use seed lists or pre-selected brands.
• We don't edit or filter AI responses.
• The data is what the AI returned. Nothing more, nothing less.

The Prompts

Full transparency. Here are the exact prompts we send to each AI provider. The same question text is what Perplexity and Google AI Overviews are asked for the AI Sources pages — on its own separate call, with no answer format imposed.

Update Frequency

Weekly. Rankings and debates run every Sunday; AI Sources runs mid-week on its own timer and is filed under the same week. Historical data is preserved — trend arrows show movement over time, and each week's archive page carries that week's rankings, debate and sources check together.