Photofeeler score calculator

How the formula works

Your score starts with the raw votes you receive: No, Somewhat, Yes, Very. Not every vote counts the same, though. Photofeeler's VoterStyles algorithm reweights each vote based on the voter's history, summarized as a single number from 0 to 10 called their lean (0 = habitual flatterer, 10 = harsh critic).

What each piece does:

• Base value ($A$) is what each button contributes to z when cast by a neutral voter (lean = 5).
• Lean correction ($B$) sets how much the voter's lean shifts that base value. A "Very" from a harsh critic moves z noticeably more than the same lean shift applied to a "No".
• Trust weight ($\Phi(W_0)$) is the share of a vote that survives Photofeeler's quality filter. Low-quality and suspiciously generous votes lose most of their weight.
• The final score ($\text{median}$) runs the weighted z-score through the standard normal CDF $\Phi$ and multiplies by 10. Because $\Phi$ is an S-curve (not the bell itself), $z = 0$ lands exactly at 5.0 and the extremes saturate near 0 and 10.
• The 90% range ($q_p$) narrows as the vote count $n$ grows. Its width at any given z also depends on the bell-curve density $\phi(z)$ and a global dispersion constant $C$.

Finding the weights

We fit these constants on 473 Dating-category tests, covering 25,098 votes across Smart, Trustworthy and Attractive. A Nelder-Mead optimizer tuned them to match both the reported median and the five quantiles of the 90% range at the same time. The fitted formula then reproduces Photofeeler's reported medians to within about 1.7% mean absolute percentage error.

What we learned

The fitted numbers already tell a few stories. Three values matter most. The base score is what a single vote of that type would produce on its own if cast by a neutral voter (lean = 5). The kept weight ($\Phi(W_0)$) is the fraction of that vote that survives the trust filter. The lean slope ($B$) is the per-lean-point shift each button adds to the z-score. A bigger slope means the same click counts very differently depending on the voter's history. Switch traits to see how they compare:

• "Very" votes argue hard, but get throttled hard. A single "Very" pulls toward a flat 10.0, yet the trust filter keeps only 26–33% of its weight by default. The algorithm assumes anyone clicking "Very" is probably too generous, so its full weight only returns when a habitual harsh critic casts it.
• "No" and "Somewhat" pass through almost unfiltered. On Attractive and Trustworthy, kept weights sit near 100% (94–98% on "Somewhat", 100% on "No"). Smart is the slight exception, keeping 84% of a "No" and 90% of a "Somewhat".
• Harsh critics move the score more than flatterers do. Look at the lean slope row above: on every trait, "Very" is roughly four to five times more sensitive to voter history than "No" (0.39 vs 0.09 on Attractive, 0.39 vs 0.08 on Smart, 0.31 vs 0.06 on Trustworthy). A concrete example on Attractive: a three-point lean jump from neutral (5) to harsh critic (8) adds 0.39 × 3 = 1.17 to a "Very"'s z-contribution, but only 0.09 × 3 = 0.28 to a flatterer's "No". That asymmetry is why a "Very" from someone who usually clicks "No" lifts your score noticeably more than a "No" from a habitual "Very" voter drags it back.
• "Somewhat" barely moves the needle. Its base score sits near 5 on all three traits, so extra "Somewhat" votes mostly pull the weighted mean toward the middle.

Why a high Smart score is easier than a high Attractive one

The three traits don't share the same score distribution. Across the 472 photos we have medians for, the share scoring 9.0 or higher is 18.6% on Smart, 14.2% on Trustworthy and 7.0% on Attractive. Standout Smart scores are almost three times as common as standout Attractive ones.

• Smart rewards legible signals. A clean background, good grooming and a fitted suit or blazer read as "put-together" within a second, and voters act on that cue fast. The bar is essentially categorical: you either look the part or you don't. Photos that clear it tend to land in the 8–9+ range rather than sit in the middle.
• Trustworthy has the highest floor but a thinner top. Its mean is the highest of the three (7.10, versus 6.73 on Smart and 6.74 on Attractive), so most photos drift into the 7s. The 9+ tail is thinner because voters reserve high Trust for warmth cues: a relaxed jaw, real eye contact, a natural smile. Outfit changes can't manufacture that.
• Attractive is the hardest tail to reach: only 7.0% of photos clear 9.0. The score depends on things you can't change on shoot day (bone structure, skin, body composition) plus things you can (lighting, grooming, wardrobe). Progress usually comes from stacking small wins, not flipping one switch.

There's also a visible coupling between traits. Among the 97 photos scoring 8.0 or higher on Attractive, the mean Smart score is 8.79 but the mean Trustworthy score is only 7.94, a gap of about 0.85. Voters cool on Trust once a photo looks physically striking, reading the subject as posed or edgy. That's why, once you optimize for looks, Smart climbs alongside Attractive while Trust lags behind.

What's in a Photofeeler test record

Behind every result page is a JSON object with the photo, your request settings, the full vote breakdown and the notes voters left. Most of it shows up on screen in some form; a few fields (the per-voter leans, the quantiles) only become legible once you know what they mean.