How rare are your Photofeeler scores?

Enter your Dating scores to estimate where each one ranks, and how rare it is to score that high on all of them at once.

1.3%

About 1 in 77 photos

of Dating photos are estimated to score at least Attractive 9.0 Smart 9.0 Trustworthy 9.0 after 20 votes

Your scores

Enter them as shown on your Photofeeler results. Uncheck a trait to leave it out.

Votes on your test

Pick the closest. With fewer votes, more photos reach a high score by luck, so a high score is a little less rare.

Unofficial estimate using the same account data as our score calculator. Combined ranks assume other photos have similar relationships between traits. Spot a mistake? Email nikita@getdatingphotos.com.

How this is calculatedCommon questions, then the math

What does the big number mean?

It's the model's estimated share of Photofeeler Dating photos that score at least as high as yours on every trait you left checked, after the same number of votes as yours. It isn't a measured population frequency.

Take the example in the table: Attractive 7.8, Smart 8.4 and Trustworthy 6.9 after 20 votes. The big number would read 5.5%, the share of photos that clear all three. The odds line under it says the same thing another way: about 1 in 18.

The rank under each trait is that figure for one trait alone. In the example, 22% of photos score 7.8 or higher on Attractive, so it reads Top 22%. Below about 5.0, the middle of the scale, it counts from the bottom instead, so a 3.0 reads Bottom 30%.

Example at 20 votes
Attractive 7.8Top 22%
Smart 8.4Top 16%
Trustworthy 6.9Top 31%
All three at once5.5% 1 in 18

Where does "Top 10%" come from?

From Photofeeler. Its site code labels six steps from 8.0 up, shown in the panel. A score between steps gets the step below it, so on Photofeeler an 8.5 still reads Top 20%. Photofeeler compares your photo with photos of people your gender and age.

Below 8.0 it only says "Above Average" (6.8 and up) or "Average" (5.0 and up). We extend the numerical pattern between labels: subtract your score from 10 and multiply by 10%. We anchor that rule at 20 votes, the most common test size in our data. That reference count is our modeling choice; Photofeeler doesn't say its labels require 20 votes. So here an 8.5 is Top 15%, and the example's 7.8 is Top 22%.

Photofeeler's labels

8.0
Top 20%
9.0
Top 10%
9.5
Top 5%
9.7
Top 3%
9.8
Top 2%
9.9
Top 1%

Why aren't three Top 10% scores 1 in 1,000?

Because the traits move together. A photo that scores high on one trait tends to score high on the others, Attractive and Smart most of all.

If the traits were unrelated, three 9.0s would be 10% × 10% × 10% = 0.1%. With the correlations fitted across 757 Dating tests of men's photos rated by women, the model gives 1.3%, about 1 in 77. We exclude tests that also allowed male voters. The example works the same way: 5.5% together, against 1.1% if the traits were unrelated.

Scoring 9.0 or higher at 20 votes
TraitsTogetherIf unrelated
Attractive & Smartcorrelation 0.634.1% 1 in 241.0%
Attractive & Trustworthycorrelation 0.382.6% 1 in 391.0%
Smart & Trustworthycorrelation 0.352.4% 1 in 411.0%
All three1.3% 1 in 770.1%

Bars show correlations after ranking each trait and mapping those ranks to a bell curve. This separates how traits move together from their different score distributions.

Does my vote count matter?

A little in this model. With fewer votes, we allow more high scores to arise by chance. After 10 votes, a 9.0 is Top 11% instead of Top 10%. Near the top the effect is larger: a 9.9 goes from Top 1.0% to Top 1.4%, or from 1 in 100 to 1 in 73. These adjustments assume Photofeeler's reported ranges approximate vote-to-vote noise; we haven't verified them with repeated tests of the same photos.

The example is at 20 votes, where ranks follow Photofeeler's scale with no adjustment. At 10 votes its combined estimate moves too little to change the rounded result: 5.5%.

More votes also make your own score steadier. A photo's long-run score is the one unlimited votes would give. Under our noise assumptions, a photo with a long-run 9.0 lands in the chart's range 8 times out of 10: after 10 votes, anywhere from 7.2 to 9.8. Our data covers 3–50 votes per test; results outside that range are extrapolations.

Rank by number of votes
Votes9.09.9
10Top 11%Top 1.4%
20Top 10%Top 1.0%
40Top 9.7%Top 0.79%
80Top 9.5%Top 0.68%
Where a long-run 9.0 lands
  • 10 votes7.2–9.8
  • 20 votes7.8–9.6
  • 40 votes8.2–9.5
  • 80 votes8.5–9.4

Each bar covers 8 out of 10 results under the model's assumptions. The dot marks the long-run 9.0.

How sure is this?

Ranks for one trait follow Photofeeler's own scale closely. The combined figure is a rough estimate.

Photofeeler doesn't publish its formulas, so ours are approximate. Its comparison group includes photos of people your gender and age. Our account's tests are selected, related photos, so their correlations may differ from that wider population.

  • One trait, 20 votesAnchoredMatches all six of Photofeeler's labels exactly. Interpolation between them and the reference vote count are our assumptions.
  • One trait, 10, 40 or 80 votesModeledA symmetric noise approximation fitted to the reported range widths of 757 tests. Treating those widths as an 80% noise interval is an assumption, not measured repeat-test coverage.
  • Traits combinedRoughResampling the 54 activation weeks puts the model's three-9.0 estimate between 1 in 55 and 1 in 120 in the middle 95% of resamples. Weeks are a rough grouping because we lack photo or shoot IDs. This range doesn't capture selection bias or uncertainty about the model itself.

Does it work for women's photos?

Mostly. Single-trait ranks at 20 votes should hold for everyone, because Photofeeler scores each gender on its own scale. Other vote counts and combined results are less certain, because the vote-luck model and the correlations come only from men's photos rated by women.

Show me the mathNine short steps, each with its formula

One trait

  1. Each trait's score sk (k is Attractive, Smart or Trustworthy) becomes a position zk on a standard bell curve, where 0 is the middle: 5.0 is 0, 8.0 is 0.84, 9.0 is 1.28 and 9.9 is 2.33.

    zk=Φ−1 ⁣(sk10)z_k = \Phi^{-1}\!\left(\dfrac{s_k}{10}\right)

  2. After 20 votes, a trait's share Pk, the fraction of photos scoring sk or higher, comes straight from Photofeeler's scale.

    Pk(20)=1−sk10P_k(20) = 1 - \dfrac{s_k}{10}

Vote count

  1. Each photo has a long-run position θ, the one unlimited votes would give. Across photos it follows a bell curve.

    θ∼N(0, 0.972)\theta \sim \mathcal N\big(0,\ 0.97^2\big)

  2. We assume that after N votes, measured z scatters symmetrically around θ, more so toward the top. The spread is fitted to the outer range endpoints of 757 tests, interpreted as a nominal 10–90% interval. Photofeeler doesn't document those percentile levels.

    σz(θ,N)=0.97 e0.4477504725224432 θN\sigma_z(\theta, N) = \dfrac{0.97\, e^{0.4477504725224432\,\theta}}{\sqrt{N}}

  3. Averaging over all photos gives Q, the model's share of photos that reach z after N votes.

    Q(z,N)=Eθ ⁣[Φ ⁣(θ−zσz(θ,N))]Q(z, N) = \mathbb{E}_{\theta}\!\left[\Phi\!\left(\dfrac{\theta - z}{\sigma_z(\theta, N)}\right)\right]

  4. A trait's share at N votes moves, on the bell-curve scale, by as much as Q moves between 20 and N votes.

    qk(N)=Φ−1(Q(zk,N))Pk(N)=Φ(qk(N)−qk(20)−zk)\begin{gathered}q_k(N) = \Phi^{-1}\big(Q(z_k, N)\big) \\[2pt] P_k(N) = \Phi\big(q_k(N) - q_k(20) - z_k\big)\end{gathered}

Several traits

  1. Each share Pk becomes a cutoff uk on a standard bell curve.

    uk=Φ−1(1−Pk)u_k = \Phi^{-1}(1 - P_k)

  2. The result is the chance that correlated bell-curve values clear every checked cutoff at once (a Gaussian copula).

    P=Pr⁡(Xk≥uk for every k)X∼N(0, R)\begin{gathered}P = \Pr\big(X_k \ge u_k \text{ for every } k\big) \\[2pt] \mathbf X \sim \mathcal N(\mathbf 0,\, R)\end{gathered}

  3. R holds the measured trait correlations, in the order Attractive, Smart, Trustworthy.

    R=(10.630.380.6310.350.380.351)R = \begin{pmatrix} 1 & 0.63 & 0.38 \\ 0.63 & 1 & 0.35 \\ 0.38 & 0.35 & 1 \end{pmatrix}