What a photo can’t see.

A phone camera only captures light bouncing off the surface of your skin. It never captures depth.

So every app and every model that scores a selfie is working out the shape of your skin from shadows, ours included. Shadows are a good clue, but they are not a measurement.

We ran a study to decide what Skin Intelligence would be built on, and kept testing after. Twelve systems have now read the same scan photographs — nine general-purpose vision models running our proprietary scoring, and three commercial skin-analysis models.

We also checked those results against dedicated skin-analysis machines that had read the same subject in person, at a separate sitting. These are the kind that sit in a clinic or a salon. They photograph under fixed light, at a fixed distance, across several spectra — ultraviolet and polarised light reach what visible light cannot.

On what a phone camera can see directly — surface clarity, how even your tone looks, how hydrated your skin reads — the closest systems tracked the machines well. Even there, they did not all agree.

On anything that depends on depth — wrinkles and firmness — every system that scored them came back kinder than the machines did, ours included.

That tells you more than a winner would. The gap is built into flat photography, not a flaw in any one system.

We photographed the same subject twice in one sitting — minutes apart, same camera — once under warm indoor light, once in daylight in a different room. The skin itself could not have changed in between. Every service we tried scored them differently. Some moved by a few points. Others by far more.

That is a fact about photography. The question is what an app does about it.

A ceiling you can measure is a ceiling you can design around. Three things in the app come directly from what this study found.

The app will not take the photograph at all until the light on your face is even — one side no brighter or warmer than the other.

Before each of your first five scans, it coaches you through the same ritual — same time, same light, no makeup, ideally right after your evening cleanse.

And we blend two independent models instead of trusting one. One model on its own just sounds confident.

Then we tested the blend, and it held. Read by the models the app ships today, those same photographs came out at nearly the same score under both lights — while one model on its own swung on a single lane. Averaging two independent reads is what absorbed that, and it is why two models read every scan rather than one.

Skin Intelligence is built to track changes, not to hand down a verdict. One scan is one photograph, in one light, on one day, so a single number will always move around a little. Compare it with your own past scores — the trend over a few weeks is the part worth watching.

So treat every number as directional. A scan is a reading, not a measurement.

This is not a clinical study. This is a model and lighting comparison study rather than a laboratory test. It shows a direction, and that was what we needed.

Your score does not build your routine. The routine is built from ingredient science — what genuinely works together — plus what you tell the app: your routine preferences, the pace you set, and the two concerns you lead with. The scan’s job is to suggest those two concerns, and you choose whether to keep them. That is how a score steers the engine — one input among many.