How we test face search engines

Why publish the method first

Most face search comparisons online are written by the company that wins them, including one well-known 16-test page that its own vendor paid for and dated May 2024. We are a vendor too, so the only honest way to compare is to fix the rules before we run anything, publish them, and then publish the numbers with a date, whether or not they flatter us. The comparison pages for PimEyes, FaceCheck.ID, Google Lens and Lenso.ai all point here.

The twelve scenarios

Each scenario is one search per engine with the same photo. Scenarios cover the situations people actually bring to a face search: a stolen dating photo, an old picture, a bad screenshot, a face that should not be searched at all. The score is 0 to 5 per scenario, for a maximum of 60 points per engine.

Who and what is in the photos

• Consenting team members who have agreed in writing to have their faces searched on every engine, with the right to withdraw at any time.
• Public-domain press photos of public figures, only where the photo itself is in the public domain and the person's face already appears widely in the public web.
• Seeded pages. Before the test we publish a small set of pages with each test face on domains we control, so every engine has the same ground truth to find.
• No strangers, no customers, no scraped faces. We never test with a photo of a person who did not agree.

How the search is run

1. 1

   Buy the smallest paid tier on every engine

   The plan or credit pack is named on the results page with its price and purchase date. We do not test on free tiers alone, because free tiers often hide the source links that make a result usable.

2. 2

   Search every photo on every engine on the same day

   Indexes change daily. Running all engines within a 24-hour window keeps the comparison fair. The date is published.

3. 3

   Score against the seeded pages

   A match counts only if the source link resolves to a page we seeded or to a genuine public page of the same person. Look-alikes score nothing; see the doppelgänger effect.

4. 4

   Record the removal test last

   For scenario 11 we submit each engine's own removal request for one consenting team member, note the steps, the verification demanded and the time to confirmation, then search again a week later.

5. 5

   Publish everything

   Scores, prices, dates, the name of the person who ran the round, the photos described in words, and the raw notes. Then re-run every six months and keep the old rounds visible.

What we do not test

• Index size claims. Vendors quote numbers of indexed faces or pages that cannot be checked from outside. We report what each engine finds, not what it says it holds.
• Identification. No engine on our list is scored on naming a person, and FaceSearch does not try to. A face search returns public pages; who the person is comes from the page, not the engine. See find someone by photo.
• Private content. Nothing behind a login is in scope, on any engine. A test that found private content would be a failure of the engine, not a success.
• Speed. Normal searches on all engines finish in seconds; the differences are not meaningful for the way people use them.

How to read the results when they appear

A higher total means an engine found more of the seeded faces in harder conditions and behaved well on the two safety scenarios. It does not mean the engine will find a particular photo of you. Face search depends on whether a page has been crawled, how clear the face is and how much it has changed. Our own results explained page covers what a match does and does not prove, and that applies to every engine on the list.