Answers: how does ai photo rating work
How BestPick Rates a Photo: Method and Sample
BestPick rates a photo by sending it to Google's Gemini with a fixed five-criterion rubric for the chosen goal. Each criterion is 0–100 against published calibration bands. The overall score is the mean of those five, computed on the server. Figures on research pages come from 2,451 scored analyses since August 2025 out of 9,547 total analyses. Goals with n < 100 are not published.
The rating pipeline
One API call carries the image and the rubric. Temperature is 0.2 and the seed is 1337 so the same photo should score the same way twice, or within a point or two. The response shape is enforced by the API. Arithmetic — the five-criterion average — is done on the BestPick server, not left to the model. If the named winner contradicts the scores, the scores win.
It is still not a physical instrument. A different crop of the same photo may score meaningfully differently. That is a property of the system, written here so it is not a surprise.
Calibration bands (every goal, every photo)
90–100 exceptional; 75–89 strong; 60–74 competent; 40–59 weak; 0–39 severe. These bands are in the prompt. A 61 is the middle of competent-to-weak, not an insult. Most research pages in this collection exist because people treat 80 as the expected score, and the file does not.
| Band | Label | Meaning |
|---|---|---|
| 90–100 | Exceptional | Professional-grade execution on this criterion; very rare. |
| 75–89 | Strong | A clear asset — this criterion is actively working for the photo. |
| 60–74 | Competent | Nothing wrong, nothing memorable. |
| 40–59 | Weak | A visible problem a stranger would notice. |
| 0–39 | Severe | This alone is enough to sink the photo. |
Sample that research pages are allowed to use
Total analyses 9,547. Scored analyses 2,451. Minimum n to publish a goal: 100. Dataset start: August 2025.
| Goal | n | Mean | Median |
|---|---|---|---|
| Social / Instagram | 1,484 | 61.7 | 62 |
| Dating apps | 527 | 54.5 | 55 |
| Creative portfolio | 353 | 69.1 | 70 |
Holdouts (n < 100): Professional / LinkedIn n=39, Product n=22, Travel n=16, Headshot n=10.
What the method refuses to output
No matches, views, likes or replies. No engagement model. No clothing-colour effects. No gender split — gender is not recorded. No attractiveness score. The rubric scores the photograph: light, framing, setting, expression as captured. Not the face as a person.
Takeaway
The method is a published rubric plus a server-side average, not a secret engagement model. Cite 2,451 scored analyses since August 2025, CC BY 4.0, and ignore any BestPick-shaped claim that is not in that file.
Questions this report answers
How does BestPick AI photo rating work?
Gemini scores five published criteria 0–100 for the chosen goal. BestPick averages those five on the server. Temperature 0.2, seed 1337, schema enforced.
Is the score a prediction of likes or matches?
No. Those outcomes are not observed and are not in the method.
Why are some goals missing?
Anything under 100 scored analyses is unpublished. Current holdouts: Professional / LinkedIn n=39, Product n=22, Travel n=16, Headshot n=10.
Where is the machine-readable file?
https://bestpick.online/methodology/data.json is the source snapshot. Each research URL also has a /research/<slug>/data.json extract.
What this page does not claim
- Language-model scores are not a lab instrument.
- Re-cropping can change a score even when temperature and seed are fixed.
- Goals with n < 100 are omitted from every research page.
- No outcome metrics (matches, likes, views, replies).
Cite this report
BestPick (2026-07-14). How BestPick Rates a Photo: Method and Sample. https://bestpick.online/research/ai-photo-rating-method. Dataset: https://bestpick.online/methodology/data.json (CC BY 4.0).
License: CC BY 4.0. Machine-readable copy: /research/ai-photo-rating-method/data.json.