DEV Community: Антон Темербеков

Who gets rated higher: age, age gap and the reciprocity myth - from 24 million ratings

Антон Темербеков — Sat, 12 Sep 2026 08:01:37 +0000

In the previous piece I went through 24 million ratings from our app Rate Me (people rate each other's photos 1-10) and showed four things: the dip at nine, men underrating men, countries rating differently, raters getting tired within a session. The most common follow-up question was about age. So here is age - plus one myth I believed myself.

Same disclaimer as last time: this is data about how raters behave, not about who is "more attractive". Birth dates are self-reported, so every cut below uses only pairs where both ages are known and fall in 16-80. A rating of 5 is the slider's default position; it makes up a quarter of all ratings and pulls every group's average down equally, so it does not affect comparisons between groups.

1. Age of the person rated: peak at 18-24, then only downhill

In all four "who rates whom" pairs the maximum sits at 18-24 and the average then declines monotonically:

Men rating women: 6.75 (18-24) → 6.49 (25-34) → 6.17 (35-44) → 6.04 (45-54) → 5.52 (55+). The first cell alone is 6.5M ratings, the biggest slice of the dataset.
Women rating men: 6.73 → 6.59 → 6.05 → 6.08 → 4.88 for 55+.
Women rating women: 6.51 → 6.31 → 6.02 → 5.92 → 5.22.
Men rating men: 5.71 → 5.51 → 5.28 → 4.80 → 4.35, the lowest cell in the entire dataset.

Two observations. First, the gap between "18-24" and "55+" is 1.2 to 1.9 points, larger than the gender effect from the previous article (1.14). The age of the person rated is the strongest factor after who is doing the rating. Second, men are not just harsh on men, they get harsher with every age band.

Caveat: the 55+ cells have an order of magnitude less data (17K to 66K ratings vs. millions), but that is still tens of thousands, so the direction is solid.

2. Age gap: men and women behave differently

This is the interesting part. Instead of absolute age, take the difference "rater minus rated" and see how the rating moves.

Men rating women - monotonic: the older the man relative to the woman, the more generous. A man 10+ years younger gives 5.91, same age 6.54, 10+ years older 6.81. No same-age peak, the curve just climbs.

Women rating men - a completely different shape. Peak at same age (6.71), decline both ways. And asymmetric: a woman 3-10 years older than the man gives 6.61, 10+ older 6.43; a woman 10+ years younger than the man gives 5.64. Young women are the toughest judges of men noticeably older than them.

Same-gender pairs repeat the "same-age peak" shape: women rating women 6.08 → 6.41 → 6.56 → 6.57 → 6.47; men rating men 4.98 → 5.40 → 5.52 → 5.46 → 5.98. The harshest cell in the whole dataset: a young man rating a man 10+ years older - 4.98.

I will not try to explain this with psychology, the data is not enough for that. But the curve shapes are stable and hold across the full 12-year span.

3. There is no reciprocity

A hypothesis I took for granted: generous raters get rated more generously in return. The app has no reciprocity mechanic, but I assumed it would show up through behaviour somehow.

I took 5,207 users who have both ≥100 ratings given and ≥100 received and correlated their average given with their average received. Result: -0.046. Essentially zero. The bucket breakdown is flat: users who give 3-4 on average receive 6.54; users who give 9+ receive 6.59. Everything in between sits in a 6.40-6.62 corridor.

So rating "karma" does not exist: how much you give others has no relation to how much you get. Which makes sense once you think about it: the people rating you are not the people you rated, they are random users from the feed.

What did not make the cut

Three cuts that looked promising and got dropped:

hour of day and day of week - spreads of 0.15 and 0.06 points, and that is in UTC without timezone correction;
"photo number": the fourth and later photos score 0.3 lower - but slots 4-6 only unlock at high levels, so that is a different audience, not an ordering effect. Classic selection, not a finding.

How it was computed

Postgres on production, full dataset, no sampling. Age is age(rating.created, user.date_of_birth) at the moment of the rating, for both sides. There are no per-country cuts here, so the rater-count threshold from the previous article was not needed. All queries are plain GROUP BY with CASE buckets; the slowest one (reciprocity, two aggregates plus a join) took about four minutes.

Data is from Rate Me (rateme.lv) - happy to answer questions in the comments.

We analysed 24 million ratings of human faces. Four things surprised us

Антон Темербеков — Thu, 13 Aug 2026 07:43:31 +0000

We run an app where people upload a photo and get rated from 1 to 10 by strangers around the world. It has been running since 2014. In that time the database has collected 24,097,993 submitted ratings from 109,821 people across 232 countries — plus 8.4 million skips, where someone looked at a photo and moved on without scoring it.

This is not a scientific sample, and I'll get to the limitations in a moment. But data like this is hard to come by: academic work on attractiveness perception typically runs on hundreds of participants and dozens of photographs. Here we have twelve years of continuous observation of how people judge other people's faces. I went into our own PostgreSQL and found four things I did not expect.

First, the limitations

This is data from one app, not from humanity. The people here actively chose to be rated — a very particular sample. Country comes from the city in the user's profile, not from geolocation. Gender is self-declared. We don't know who is actually in the photo.

So everything below is about the behaviour of the raters, not about who is better looking. The question "which country has the most attractive people" is meaningless in this data, and I'm not asking it.

Finding 1: nobody gives a nine

The first thing that jumps out of the distribution is the hole at nine. A ten is given 3.4× more often than a nine: 14.6% versus 4.3%. Between "eight" and "ten" people barely stop.

The single most common answer is five, at 24.45%. That's the slider's default position, and a quarter of raters simply never move it. Harsh scores are rare: ratings 1 through 4 together account for 12.4%.

Score	Ratings	Share
1	476,591	1.98%
2	602,882	2.50%
3	898,204	3.73%
4	1,009,530	4.19%
5	5,892,680	24.45%
6	3,548,185	14.72%
7	4,379,993	18.18%
8	2,719,997	11.29%
9	1,042,271	4.33%
10	3,527,681	14.64%

A ten doesn't seem to work as a point on the scale — it works as a separate gesture. It means "wow", not "nine plus one". A nine demands that you weigh something; a ten doesn't. Anyone who has collected NPS scores or five-star reviews knows this shape: people cluster at the ends and on round numbers instead of spreading across the range.

Finding 2: men punish men

This is the sharpest gap in the entire dataset. Women rate men and women almost identically — 6.45 and 6.32. Men rate women 6.54, and other men 5.40.

That's a gap of 1.14 points between how a man rates a woman and how the same man rates another man. The equivalent gap for women is 0.13 points — essentially zero.

Pair	Ratings	Average
Man → man	780,766	5.40
Woman → woman	1,410,741	6.32
Woman → man	2,617,973	6.45
Man → woman	19,288,521	6.54

The effect holds across all twelve years and isn't explained by sample size. I can't tell from the data what drives it: competition, disinterest, or simply that a man looking at a man is answering a different question — not "is he good looking" but "is he better than me". But the size of the gap is hard to write off as noise.

Finding 3: a 2.4-point spread between countries

The same photo will get a noticeably different score depending on who is looking at it.

There's a methodological detail worth dwelling on. My first pass put East Timor, Anguilla and Senegal at the top of the generosity ranking — and that turned out to be a pure artefact. All 5,942 of East Timor's ratings came from a single user; Anguilla and Senegal had three raters each. Once I required at least 100 distinct raters per country, the picture became meaningful:

Country	Raters	Average
Norway	101	7.87
Czechia	114	7.21
Mexico	139	6.96
Canada	726	6.89
Netherlands	213	6.72
Belgium	145	5.88
Romania	121	5.87
Türkiye	224	5.84
Hungary	126	5.78
Ukraine	108	5.48

And a separate detail that breaks intuition: people are harsher on their own. When rater and rated are from the same country the average is 6.34; when they're from different countries it's 6.48. The difference is small but stable across four million ratings, and it points the opposite way from what you'd guess. Shared nationality works against you here.

Finding 4: the first ten ratings of the day are the kindest

I ordered each person's ratings within a day: first of the day, second, tenth, hundredth. People start generous and harden fairly quickly.

The first ten ratings of a day average 6.98. By the fiftieth-to-hundredth the average has fallen to 6.27 — a drop of 0.71. After that the curve turns back up: people who reach three hundred ratings in a day score higher again, but that's a different, much smaller audience that's into the process itself.

Position in day	Ratings	Average
1–10	283,457	6.98
11–25	178,691	6.66
26–50	130,570	6.31
51–100	144,877	6.27
101–250	188,050	6.30
250+	265,438	6.51

There's a practical lesson here for anyone who collects ratings of anything — CVs, contest entries, products: presentation order affects the result more than you'd think. The difference between reaching a reviewer fifth and fiftieth is roughly seven tenths of a point out of ten. If you don't randomise order, you aren't measuring quality — you're measuring position in the queue.

What to take from this

None of these numbers say anything about how anyone actually looks. All four findings are about the rater: their fatigue, their gender, their culture, their relationship with the scale. The person looking at your photo brings far more to it than you'd assume from the number that comes back.

Age plays a part too: the harshest raters are 18–24 (6.14), rising to 6.64 among 35–44 year olds. The effect is real but weaker than the others, so I left it out of its own section.

All aggregates were computed directly in PostgreSQL on production; no personal data was used or published. The app is called Rate Me. But the most useful thing to take away from these 24 million ratings is a healthy scepticism toward any single number somebody has assigned to you or your work.

How we distribute an iOS app outside the App Store: a practical AltStore PAL walkthrough

Антон Темербеков — Thu, 06 Aug 2026 19:52:07 +0000

App Store review rejected our app under Guideline 1.2 — the core mechanic (users rate each other's photos) is banned as a category, and no metadata change fixes that. So we became one of the first apps distributed exclusively through an alternative marketplace under the EU's DMA. Documentation for this path is thin and sometimes contradictory, so here is a step-by-step of what actually had to be done.

Step 1. Alternative Terms Addendum

You sign an addendum to the developer agreement in App Store Connect. You don't need to be an EU resident. Distribution is free up to 1M first annual installs (companies under €10M revenue get the Core Technology Fee exemption).

Step 2. Registering with the marketplace

AltStore has a REST API to register your developer ID — important: it's the UUID from your ASC profile, not the Team ID (registering with the Team ID silently fails — we stepped on that). The token goes into ASC: Users and Access → Integrations → Marketplaces.

Step 3. Notarization instead of review

You build and upload with the usual xcodebuild + upload. Then, instead of "Submit for Review", you send it for notarization: an automated security/functionality check, without content guidelines. Ours passed on the first try in ~30 hours. A mechanic banned on the App Store passes here.

Step 4. Hosting the ADP yourself

After notarization you download the Alternative Distribution Package: manifest.json, a signature file (no extension!), and several .ipa variants. You host it as-is; hierarchy and hashes must not change. Two gotchas:

the extensionless signature file — our SPA fallback on ASP.NET served index.html instead; fixed with ServeUnknownFileTypes on the static handler;
the .ipa must be served as application/octet-stream.

Step 5. Source JSON and federation

Next to it you put a source.json: marketplaceID (= the app's Apple ID), the manifest downloadURL, size, icon, screenshots, versions. Users add the source via an altstore://source?url=… link. Enable federation (a fediUsername field) and the app becomes searchable right inside the marketplace, no manual source add.

Updates

Each new version is the same loop: build → notarize → new ADP on the server → new entry in versions[] of source.json. AltStore picks up auto-updates itself.

Limits and fallbacks

Marketplaces work in the EU, Japan and Brazil, iOS 17.4+. For the rest of the world we run a PWA (standalone manifest + a minimal service worker — Safari installs it to the home screen) and a sideload IPA via AltStore Classic with our own source. The biggest non-technical problem is conversion: you have to explain to users what a marketplace even is, so we funnel all traffic to a single install page.

Takeaways

Cost beyond the developer account: zero. Actual time spent: about a week, most of it on the undocumented details above. If the App Store is closed to your app (or you just don't want to hand over 15–30%), this path works today.

The app from this case study is Rate Me (rateme.lv/ios) — happy to answer questions in the comments.