How we measure

The GDPR profile

The GDPR profile summarises what a provider publishes about your data. It is a tool for your shortlist, not legal advice and not an audit.

For a professional tool we award points: 2 if the provider states that the data is stored in the EU, 1 if the provider cites a certification (ISO 27001, NEN 7510) or GDPR compliance, and 1 if the provider is established in the Netherlands or the EU/EEA. 3 or 4 points is Strong, 2 is Good, 1 is Limited and 0 is Unknown.

A mention only counts if it is not denied: "no GDPR statement found" or "data not in the EU" earns no point.

For a general AI model we look at three things: EU hosting available, no training on your data for business use, and self-hosting. All three, or EU hosting together with no training, is Strong. Any other combination of two is Good, one is Limited, none is Unknown.

Caution is given to a model with a concrete, documented concern, such as an investigation by a regulator or a licence restriction. That outweighs the points.

Unknown means we found none of the three things that earn points. Sometimes the provider publishes something else, such as a SOC 2 report, HIPAA or servers outside the EU; that earns no points and is noted on the safety page. Unknown is not a judgement on the quality of the tool.

Status today across all 964 professional tools: 144 strong, 234 good, 392 limited, 194 unknown.

Elo and Code-Elo

Elo is the quality rating from LMArena (arena.ai). Users ask a question, receive two anonymous answers and choose the better one. From those duels each model gets an Elo score, much like in chess.

Code-Elo comes from the same duels, but only with programming tasks. A model may rank high on ordinary text and lower on code, or the other way round.

Per model family (Claude, ChatGPT, Gemini, …) the best-scoring variant on LMArena counts. We fetch the text and code rankings every morning and keep the history, so you can see the trend. If the official ranking is not reachable, we use a shorter public copy of it. A model that is not in that copy keeps its last real measurement; the model page then shows that date, and in the monthly ranking it appears among the models without a measurement, not among the models that have dropped off LMArena.

A difference of around 20 points is noticeable in practice. Differences below 5 points fall within LMArena's margin of error: in a comparison between two models we show them as equal, and the monthly ranking notes which models lie within that margin of the number 1 spot. Models only share a position when their Elo is exactly the same.

A model that is not listed on LMArena itself gets no Elo of its own; we do not estimate one. Exception: Microsoft Copilot runs on OpenAI's GPT, so for Copilot we show that model's Elo. Copilot therefore does not count in the monthly ranking, otherwise the same model would appear twice.

The editorial score

The editorial score is our own judgement of 10 per model family. It weighs three things: how suitable the model is for professional work, the price including what you need to buy alongside it in practice, and privacy.

No formula calculates the score: it is an editorial judgement, recorded in the data and adjusted when a model, price or term changes. That is why it never stands alone, but always sits alongside the independent Elo.

The model leaderboard is based on Elo; the editorial score sits alongside it. Where we give a recommendation, both carry equal weight: in a comparison of two models (which one scores higher on quality) and in the order of the general assistants per profession, after the GDPR profile. Both are then scaled between the lowest and highest model in the index and each counts for half.

How often we check

Every morning we refresh quality (LMArena), API prices and context windows (OpenRouter), popularity (HuggingFace, GitHub) and community signals (Hacker News).

The editorial agent checks professional tools against the provider's website and terms: every day a fixed batch on features and free trial, every day ten tools with missing data ('Not known'), every Thursday a round of prices, twice a week new tools in a few professions, every Friday the models, and on the first of the month five GDPR profiles again. For the ten tools with missing data, the tools with the most impressions in Google come first; a script fixes that list in advance, the agent does not choose it itself.

The 'Last checked' line on a tool or safety page is the day that tool or model was last reviewed. On a profession page, it is the last day the list or one of the tools on it was reviewed.

Facts come from public sources: the site, documentation and terms of the provider, or an official source. Where we have recorded the source URL, it appears with the tool or model. If something turns out to be wrong, we would be glad to hear about it via the update form on the tool page or via the About page.

Not for sale

Providers cannot pay for a place, a higher score or a better GDPR profile. There are no sponsored listings.

A provider can supply missing details via 'Is this your tool?'. That submission goes into a queue and only changes something once the source confirms it.

If affiliate links ever appear, we will flag them. The order always stays based on data.