CoinRecognitionApp tests coin recognition apps for collectors and casual finders who want an honest 75% over a marketed 99%. We believe an AI that tells you when it's uncertain is more useful than one that pretends every photo is a confirmed ID.
Who We Are
One of us inherited a box of coins and reached for a recognition app—it confidently called a worn Mercury dime a Barber quarter. The app gave a certainty score of 97%. That moment taught us something: an app that will not admit 'I'm not sure' is dangerous. We started testing coin recognition apps in 2023 to understand what was really happening under the hood: how computer vision actually works on coin photography, why lighting destroys accuracy, why a worn obverse fools most algorithms. We tested seventeen apps and found that the honest ones were rare. Most claim near-perfect accuracy. None of them do. Our team is two working collectors with backgrounds in software and photography—not numismatists, but people who use these apps weekly and have learned to read their failure modes. We believe you deserve to know what an app can and cannot do, and at what confidence level you should trust its verdict.
Methodology
We built a test set of 34 coins across three strike types: Business strikes, Proofs, and Specimen strikings. The coins span Lincoln wheat cents, Mercury dimes, Morgan dollars, Roosevelt dimes, and Standing Liberty quarters—denominations where wear, patina, and strike variation create real visual ambiguity. We photograph each coin under five lighting conditions: direct overhead, side raking light, diffuse indoor light, outdoor overcast, and strong shadow. For each app, we feed the same 170 photos (34 coins × 5 lights) and log the result, confidence score, and how the app handles low-confidence scenarios. This takes between 60 and 100 hours per app over four to six weeks. We re-test quarterly and after any major app update that claims improved accuracy. The heavy lifting is the light variation—it's where recognition fails hardest and where apps most often refuse to admit uncertainty.
Our Standards
We score apps on whether they admit uncertainty, not on how many coins they claim to identify. Computer vision on coins is hard. A worn Lincoln cent and a heavily cleaned Morgan dollar can look visually similar under poor lighting. Most apps handle this by returning a single high-confidence ID. We believe that is wrong. An app that says 'this is probably a Lincoln cent, confidence 68%, but it could be a cleaned Morgan if the weight differs' is more useful and more honest than one that says 'Lincoln cent, 94% confidence' and stays silent on the wear pattern that could flip the call. We test for three things: whether the app shows confidence as a range, not a binary yes; whether the app explains what features it used to make the ID; and whether the app surfaces the alternative IDs it considered. We also test strike-type coverage—how the app handles Proofs versus business strikes of the same year, and whether it accounts for that difference in valuation. An honest confidence score is not the same as a high confidence score. An app that returns a 72% match with three alternatives is, in our testing, more reliable than one returning 96% with no caveats.
Disclosure
We do not accept paid placement or sponsorship from app developers; we do not test an app unless we have used it ourselves for at least two weeks of real ID attempts. We do not score an app higher just because it returns a fast result—speed without honesty is noise. We do not claim expertise in ancient coins, world coins, or error varieties beyond our core test set; we test what most US collectors encounter, and we are transparent about that boundary.
Contact
App developers can request review via the contact form on this site. Readers with suggestions for coins to add to our hard gallery, or questions about why a particular app failed on a specific coin, are welcome to write in.