Proof PerimeterRequest a demo
Evaluation & Quality

OCR Accuracy Rate

The number everyone quotes and nobody defines — what OCR accuracy claims actually mean, and how to read them.

OCR accuracy rate is the headline number of text recognition — "99.5% accurate" — and the glossary entry it most needs is a decoder ring, because the number is meaningless without its definitions: accuracy of what (characters, words, or fields — each a different metric with different arithmetic: 99% character accuracy can be ~95% word accuracy and far lower field accuracy, since errors compound upward), measured on what (clean printed scans or the fax-and-handwriting tail — the corpus choice moves results by tens of points), under what matching rules (case, punctuation, whitespace, normalization — undisclosed leniency inflates freely), and against what ground truth (whose transcription, resolved how).

Reading vendor and benchmark claims therefore means asking the qualifying questions: the metric's level and formula (character error rate's complement is the honest default; "accuracy" without a formula is marketing), the test population's composition (document types, capture channels, languages, quality mix — and whether it resembles your stream), the normalization and scoring rules, and the segmentation of results (an aggregate hides the segments; per-type numbers inform decisions). The strongest response to any claimed rate is the one this glossary's benchmarking entries systematize: measure on your own documents, with your own sealed ground truth, under stated rules — the only accuracy number that predicts your production experience.

Used internally, accuracy rate becomes a managed metric rather than a claim: tracked per document type, channel, field, and language; paired with confidence calibration (does the score predict the accuracy?); trended against drift; and connected to the business numbers it drives — because the operational meaning of "99% vs 98%" is not one point but a doubling of the error volume that review queues, exception handling, and downstream corrections must absorb. The rate is a summary; the distribution underneath it is the truth.

Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.

Book a demo