Methodology

Where the data comes from

We take specifications from official manufacturer materials: technical data sheets, which are usually PDF files from press rooms, and also price lists, press releases and the national sites of the brands. We do not copy data from other automotive sites, from aggregators or from encyclopaedias. Copying would repeat the errors of other people, and we would have no way to detect those errors.

When a value is missing from the material of the manufacturer, we leave the field empty. An empty field is a decision, not an oversight. We would rather show a dash than a number that we cannot find in a source. The materials that we used for a given car are listed at the bottom of the page of that car.

Measurement convention

A comparison means something only when every car was measured in the same way. We therefore follow three rules, even when a manufacturer reports a value differently.

Range and energy consumption: we use the WLTP homologation figures. These figures come from a laboratory procedure, so real consumption is often higher and real range is often lower. WLTP is still the only measure that is calculated in the same way for every car.

Reference configuration: the data describes the base configuration of a given variant. Manufacturers publish range as a span, and that span depends on wheels, tyres and equipment. Taking the upper value for one car and the lower value for another car would turn a comparison into a lottery.

Mass: we use kerb weight in the DIN sense, which means weight without a driver. Some brands publish weight in running order, and that figure includes a driver of 75 kg. In those cases we subtract 75 kg. Without this correction one car would look heavier than another car, although the whole difference comes from the method of counting.

Model, variant, generation

A model is a nameplate. A model has one page, one address and one set of reviews. A variant is a specific powertrain configuration, and every variant has its own specification.

We create a new variant only when the battery, the range, the power, the drivetrain or the charging parameters change. Equipment packs, wheels and colours do not create a separate variant, because they change none of the numbers that we show.

A new generation of the same model is a new entry with its own address. We do not overwrite the old entry. The current generation keeps the clean model address, and the previous generation gets an address of its own. A person who buys a used car then sees data that really describes that car. A discontinued model stays on the site with a note, and with a link to its successor when we know the successor.

Verification statuses

Every variant has one of three statuses. Verified means that a person went through the full set of fields with the material of the manufacturer open. Automatic means that the data is complete, internally consistent and taken from an official source, but that nobody has reviewed the data by hand yet. To be checked means that the data is incomplete or doubtful and waits in a queue.

We do not publish a percentage figure for data confidence. A number such as 50% reads as a claim that the data is half true, but the number means something else. The number only says how many fields are filled in.

What we use AI for, and what we never use it for

We use language models for work on language. Language models edit the model descriptions, and they translate the interface and the content into further languages. The models run locally, on our own hardware.

We never use AI as a source of technical data. No number in a specification comes from a language model: not the range, not the battery capacity, not the power, not the charging power, not the charging time, not the mass and not the dimensions. Every value is copied from the materials of the manufacturer. Every value passes checks: plausible ranges, consistency between fields, and agreement with the rest of the line-up of that model.

We check the descriptions by sampling, and we also check them automatically. The automatic checks look for numbers that contradict the specification, and for text that a language model returned broken. These checks are not perfect. If you still find an error in a description, please report it. We fix the error at the source, in all languages at once.

Video reviews

Reviews come from public YouTube channels. We match a review to a model by the title of the video. A review is always the opinion of its author, and never our assessment of a car. We do not pay for reviews, we do not commission reviews, and we have no influence on their content. We embed the videos in no-tracking mode.

Matching by title can fail, especially when two generations of a model share one name. When we are not sure, we prefer to attach a review to no model rather than to the wrong model.

Ratings from users

A model rating is a number of stars from 1 to 5. A rating is anonymous and has no written content. On the page we show only the average and the number of ratings. A rating is the subjective opinion of a visitor, and not a measurement. We do not check whether the person who rated the car has ever seen that car, and the average is not our recommendation.

We collect no comments and no nicknames. Publishing what other people write would require moderation, and a project run in spare time cannot provide that moderation. A number of stars is therefore the only thing that a user can publish here.

What happens when the data changes

A manufacturer can change a specification, facelift a car, withdraw a version, or revise a declared range after a new homologation. Manufacturers do not tell third-party sites about such changes. The data here can therefore describe the state before a change. Confirm the figures with a seller before you decide to buy. The Disclaimer describes the full extent of our liability.

If you see a difference between our page and the material of the manufacturer, please report it. Use the form on the model page, or write to [email protected]. Every such report improves the site for everyone.