How we analyze
a face.
Published in full, including the parts that are approximate. If you can read this and decide the method is wrong, we have saved you $9.99 — and that is a better outcome than a refund.
Step one — six measurements
Everything we tell you descends from six numbers taken off a single front-facing photograph. No other input affects the analysis. Your answers to the questionnaire change which recommendations are emphasised, not what the measurements are.
Step two — classification
The six numbers place you in one of seven categories. The order of operations matters, because length-to-width alone misclassifies about a third of faces:
- Length to width ratio. Below 1.2 goes to the balanced branch (round, square). Above 1.5 goes to the long branch (oblong). Between 1.2 and 1.5 stays open.
- Jaw angle. A hard corner inside the balanced branch means square; a continuous curve means round.
- Forehead-to-jaw ratio. Above 1.15 means heart; below 0.9 means triangle; between means oval or oblong.
- Cheekbone dominance. If the cheekbones are the widest point and both ends are narrow, the shape is diamond — which overrides the length check.
Where two branches produce different answers, the report returns both with the closer one first and applies the correction from the neighbouring shape on the point where they disagree. That is why a report sometimes says "oval, leaning oblong" instead of a single word.
Step three — the rule library
Two hundred and forty constraints, versioned, in five families:
- Hair (72). Length ranges by shape and hair type, fade heights, fringe behaviour, growth-out intervals.
- Colour (48). Palette families by undertone and contrast level, plus the four colours that fight each combination.
- Fit (64). Shoulder seam tolerance, trouser break, jacket hem, sleeve length, by build.
- Context (36). Formality mapping — what a garment reads as in a given room.
- Exclusions (20). What not to recommend, and why. This family exists because a recommendation without a matching exclusion is how generic advice happens.
Version 1.4 is current. The changelog is public on request — email us and we will send it.
Where the method stops
Being explicit about this is the difference between a tool and a horoscope.
- Asymmetry. A two-dimensional photograph of a three-dimensional face flattens left-right differences. We can see large asymmetry, not small.
- A beard. Facial hair rewrites the jaw outline. Above about 5 mm the measurement describes the beard.
- Angle and light. A photograph from above lengthens the forehead; light from below removes the jaw shadow the angle estimate depends on.
- Hairline shape. A receding hairline changes the effective face length, and we estimate rather than measure it.
- Ethnic variation in soft tissue. The rule library was built primarily on European and North American reference data and is thinner for other groups. We are expanding it and we would rather say so than imply universal calibration.
What a person still does
On Premium, a human reads the generated report before it is sent. Their job is narrow: catch a misclassification, remove any recommendation that contradicts the measurements, and spot the case where the library has no good answer. They are not rewriting the document, and they are not adding taste.
We mention this because "reviewed by a human" is often marketing language. In our case it means roughly five minutes of checking, which is enough to catch errors and not enough to make the report personal in a way software cannot.
Method questions
How accurate is the measurement?
Within a few percent on a good photograph — front-facing, eye level, even light, no facial hair over 5 mm. A photograph taken from an angle or in poor light widens the error band, which is why the report states a confidence figure rather than pretending to certainty.
What happens when two shapes are close?
The classification returns the closer one and flags the boundary. The recommendation then uses the correction from the neighbouring shape where the two disagree, because a marginal correction in the wrong direction is cheaper than a large one.
Is this the same method every report uses?
Yes. The rule library is fixed and versioned. When it changes, the version number changes and we say so on this page. We do not adjust the rules per customer.