How our results are worked out

The principle

A result you cannot question is a result you cannot trust. Every tool here publishes what it measured, what threshold decided the answer, and how close the runner-up was. Where it is guessing it says so and lowers its own confidence rather than sounding certain.

None of these systems is a science. The categories are editorial conventions stylists use to compare options quickly, and different practitioners draw the lines in different places. What this site can offer is consistency: the same inputs always give the same answer, with the reasoning attached.

Face shape

The photo and camera routes place a 478-point face mesh on the image using MediaPipe Face Landmarker, running as WebAssembly inside your browser. Nothing is uploaded or stored, because no server is involved in the reading at all. From the landmarks the tool takes four distances: face length from hairline to chin, cheekbone width at the widest point, forehead width roughly halfway between the brows and the hairline, and jaw width below the ears. Those become three scale-independent ratios, so the answer does not depend on how far you sat from the camera.

One step sits between the mesh and the figures shown. The mesh takes each width between fixed points rather than at whichever point is widest, which gives a narrower spread than a tape measure, so a raw mesh reading cannot be compared against thresholds written for a tape. Each ratio is converted onto the published scale first. That is why the number under your result can be read against the same thresholds printed in the guides.

Each of the seven shapes has a published ratio centre. The tool takes the weighted distance from your ratios to each one, ranks all seven and reports the nearest, with the gap between first and second place becoming the confidence. The tape route skips the model and takes the four distances directly from you, then runs the identical comparison. The describe route maps three plain answers onto approximate ratios and is capped at medium confidence, because it is an estimate rather than a measurement.

• Known limit: the reading depends on head angle. A face turned or tilted away from the camera compresses one side and shifts the widths.
• Known limit: hair covering the forehead or jaw changes the apparent outline, which is why the instructions ask for hair off the face.
• Known limit: the seven-shape system itself is a convention. Many real faces sit genuinely between two of them, and the tool is designed to say so rather than to force a single answer.

Body shape

The tape route reduces bust, waist and hip to two ratios, bust to hip and waist to hip, compares them against five published shape profiles and reports the nearest, with the same first-to-second gap driving confidence. The result page draws both ratios on a scale marked with the thresholds that decided the call, because the boundaries sit close together and a hundredth either side of one is the reason two calculators give the same person two names.

A tape measure is the only route. Two others were built and withdrawn: one derived measurements from the clothing sizes you buy, one asked how garments sit on you. A US size 8 varies by more than four inches of bust between brands, so neither could produce a ratio worth reporting, and a result that cannot be trusted is worse than no result.

• Known limit: bust stands in for upper-body width because the tool does not collect a shoulder measurement. For someone with narrow shoulders and a full bust, or the reverse, the result will lean.
• Known limit: the five-shape system ignores height and torso length, both of which change what actually works on a person.

Skin undertone

Every individual undertone check circulating online is unreliable on its own. Wrist veins read differently under warm and cool light, plenty of people genuinely suit both metals, and sun response is confounded by everything from ethnicity to how recently you were outdoors.

So this tool runs four checks, weights them and combines them into one call, then shows which checks agreed and which did not. Three of four agreeing is a result to rely on. Two and two means the evidence is mixed rather than proof of a neutral undertone, so it returns neutral as a practical starting point at lower confidence rather than picking a side. The side-by-side route is closer to what an analyst does in person, comparing a warm and a cool version of the same colour near the face three times, and it is the more reliable of the two because it tests what the colour does to your skin.

• Known limit: screen colour varies by device and by brightness setting, so the side-by-side route is indicative rather than calibrated.
• Known limit: makeup, recent sun exposure and coloured indoor lighting all shift apparent undertone. The instructions ask for daylight and a bare face for this reason.

Undertone measured from a photo

The photo route measures skin colour rather than asking about it. Three patches are sampled, on the forehead and along each side of the jaw, never on the cheeks. Le Fur and colleagues found in 1999 that a* is significantly higher on cheeks and b* highest on the forehead, so a cheek sample inflates redness, drags the hue angle down and biases the answer towards cool. Since hue angle is the clinical correlate for erythema, blush, rosacea or post-acne redness would otherwise be read as an undertone.

Each patch is reduced to a per-channel median rather than a mean, because a mean follows the specular highlight that is always present on a forehead. The three patch values are then combined by median again, so one bad patch cannot swing the reading.

Before anything is measured, the sample is corrected for the colour of the light using the whites of the eyes as a reference. The sclera is the least strongly coloured surface reliably present in a face photograph, which is what makes it usable, rather than any claim that it is truly neutral. It is not: Russell and colleagues reported in 2014 that sclera become darker, redder and yellower with age, so the correction carries an error that grows with the age of the person in the photograph, and it is an approximation rather than a calibration. A grey-world estimate is deliberately not used here: a face filling the frame is the exact case where grey-world assumptions fail, because they push the dominant region towards grey and remove the signal being measured. If no usable sclera is found, the correction is skipped and the reported confidence is reduced.

The corrected colour is converted to CIELAB. Depth is reported as the Individual Typology Angle. Undertone is decided by hue angle against a boundary that shifts with lightness, because the warm and cool split does not sit at the same angle on light and deep skin. A reading within four degrees of that boundary is returned as neutral, since four degrees is inside the error of any phone camera. Olive is treated as yellow well ahead of red combined with low chroma, that is muted rather than golden, and this is labelled on the page as a working convention rather than as settled science.

• Known limit: an ordinary photograph contains no reference of known colour, so an absolute skin colour cannot be recovered from it. What this measures is relative, and the tool says so on the page.
• Known limit: the measured separation between warm and cool is roughly sixteen degrees of hue angle, and uncorrected camera error runs around twenty. Without the sclera correction the reading is close to noise, which is why the confidence is cut when the correction cannot be applied.
• Known limit: heavy foundation defeats the method entirely, because it is measuring the foundation.

Colour season

The twelve-season system is used rather than the older four-season one, because four categories are too coarse to be useful and the sub-seasons are where the practical difference lies.

The common failure of online season quizzes is that they judge one vague overall impression. This tool scores four axes separately: temperature (your undertone), depth (how light or dark you read overall), contrast (the gap between hair, skin and eyes) and clarity (whether clear or muted colours suit you). Each of the twelve seasons has a published position on those four axes, and the tool reports the nearest with the gap to the runner-up as its confidence.

Because the four axes are scored separately, you can change one answer and watch the season move. That is deliberate. Most people are sure of three of the four and unsure of one, and being able to see exactly how much the result depended on the shaky one is more honest than a single unexplained verdict.

• Known limit: this is a self-assessment. A trained analyst draping physical fabric in controlled daylight will beat it.
• Known limit: seasons near each other on all four axes will produce low confidence, and that is correct behaviour rather than a fault.

Hairstyle and glasses guidance

These two tools do not measure anything themselves. They take a face shape, from your saved result or from your own selection, and map it to guidance drawn from published styling sources, each cited with the date it was last checked.

The glasses tool adds one calculation of its own. Given your face width in millimetres it works out a target lens width and bridge width, which are the first two numbers printed inside the temple arm of every pair of glasses. This turns frame advice into something you can check against a real product label instead of a vague shape recommendation.

Trend figures, and what is joined to them

Articles quoting search volume take every figure in a single table from one source, named with the date the publisher gives for the data and the date this site last checked it. Mixing keyword tools inside one table produces numbers that look precise and are not, because two tools rarely agree on a rounded global figure for the same term. Growth windows differ between terms and are printed as published rather than converted, since a three-month figure and a year-on-year figure are not the same measurement.

The face shapes shown beside each cut are not from the source and are labelled that way wherever they appear. They are worked out here from one question asked of every cut: what does it do to the outline of a head. Weight removed from the sides narrows the silhouette, which helps a face close to as wide as it is long. A cut finishing above the jaw leaves the jaw as the widest visible line, which works against a face already widest there. A blunt fringe shortens the visible face, which a short face cannot spare. Where a cut is listed against a shape, the usual fix is one adjustment to where the shortest layer falls rather than rejecting the cut.

• Known limit: search volume is a global average and says nothing about the region a reader is in.
• Known limit: popularity is not suitability and the two columns frequently disagree. That disagreement is the reason the table exists rather than a flaw in it.

Data, storage and corrections

Results are saved in your browser's local storage so one tool can hand its answer to the next. Nothing is sent to a server, there is no account, and clearing the results removes them completely. The saved results page shows everything held on the device and clears it in one action.

The full source list behind the written guidance, with the date each source was last checked, is maintained alongside the site. Where a source is later found to be wrong, the page is corrected and the correction is dated on the page rather than removed quietly.