Start a search · $6 USD

Single-query case study · not a benchmark

What one saved face search returned: 100 candidate pages analysed.

One supplied Jensen Huang photo produced 299 parsed candidates. 221B preserved the first 100 usable candidate pages in saved result order and analysed that saved set below.

Case study by Di Tang · Updated 6 August 2026

This page replays a real 221B public-web search saved 5 Aug 2026. Scores are the exact saved values. Opening the case runs no live search, uploads no photo, and uses 0 credits.

Start one public-web search — $6 USDOpen free saved demo · 0 creditsRead methodology

Saved 221B search snapshot

Jensen Huang

One query portrait · 100 saved candidate pages · 3 featured examples

Saved 5 Aug 2026 · 0 credits

Candidate pages

100

Hostnames

27

Unique thumbnails

79

Credits to open

0

Jensen Huang in the exact photo used for the saved demoQuery photo

Exact saved demo input

Supplied photo · 5 Aug 2026

Independent context checkOfficial source

NVIDIA · Jensen Huang

NVIDIA's official biography provides independent public context for the demo subject.

Open nvidia.com

These match-quality scores are shown exactly as saved on 5 Aug 2026. They rank candidates for manual review; they are not probabilities and do not prove identity. Public pages may have changed since capture.

The demo query photo was supplied for this saved case. Result thumbnails are locally preserved low-resolution copies linked to the captured candidate-page URLs.

Saved-set findings

100 candidate pages do not mean 100 independent images.

The 100 saved rows contain 100 distinct candidate-page URLs across 27 canonical hostnames, but only 79 unique thumbnail files. That leaves 21 rows whose thumbnail exactly repeats one already present in the set.

Parsed candidates
299
Candidate pages saved
100
Canonical hostnames
27
Unique thumbnails
79
Repeat-thumbnail rows
21
Score range · mean 87.65
87–89

Surface distribution

Where the 100 saved URLs lead

Share of saved pages
LinkedIn59 pages · 59%
Instagram9 pages · 9%
Other web32 pages · 32%

Saved score distribution

A narrow 87–89 range

The mean saved match-quality score is 87.65. These values prioritise manual review; they are not identity probabilities.

Score 89
1
Score 88
63
Score 87
36

Method

  1. 1One supplied query image was run once through the production 221B public-web search flow.
  2. 2221B parsed usable URL, thumbnail and score records, sorted them by score, and saved the first 100 candidate pages in that result order.
  3. 3Canonical hostnames remove a leading “www.”. Unique thumbnails and duplicate rows use exact SHA-256 equality, not visual-similarity judgement.

Limitations

  • This is one image of one public figure from one search run. It is not a benchmark and does not estimate typical performance.
  • There are no ground-truth relevance labels, so this case cannot support accuracy, precision or recall claims.
  • The report saves 100 of 299 parsed candidates; it does not analyse the remaining 199 candidate records.
  • Source pages can change, disappear, require login or restrict access after capture.
  • Repeated thumbnails or pages are not independent corroboration, and a match-quality score is not an identity probability.

How to read the candidates

A candidate page is a lead to inspect, not proof of identity.

  1. 01

    Open each linked page

    A saved candidate becomes useful only after you inspect the linked page and its publisher.

  2. 02

    Check page context

    Compare the caption, date, image role, and surrounding text instead of relying on the face score alone.

  3. 03

    Keep it provisional

    The saved score ranks a candidate for review. It is not a probability, identity proof, or claim about account ownership.

Ready when you are

Use the same review workflow on your own photo.

Start with the $6 USD Starter Pack for one public-web search, or switch to a monthly plan once you know the workflow helps. Each search uses 3 credits.

Start one public-web search — $6 USDUpload a photo