How Accurate Is Receipt OCR? What the Benchmarks Actually Say

Receipt OCR accuracy depends entirely on which accuracy you mean. Leading systems reach roughly 95 to 98 percent field-level F1 on clean, well-scanned benchmark receipts, and considerably less on faded thermal paper, phone photos, and line items. A vendor advertising 99 percent character accuracy can still get a wrong digit into a third of your totals, because character accuracy and field accuracy are different measurements and only one of them is what you care about.

What is the difference between character-level and field-level accuracy?

There are three altitudes at which accuracy gets measured, and vendors quote the flattering one.

Character-level accuracy is one minus the character error rate. It measures raw transcription: how many individual characters came out right. This is the highest number available and therefore the number in the marketing.

Field-level accuracy measures whether a field, the total, the date, the merchant, was extracted correctly and labeled correctly. A field is either entirely right or it is wrong. One bad digit fails the whole field.

Document-level accuracy measures the fraction of documents where every field is right. That is what a bookkeeper actually experiences, and it is the harshest number of the three.

The academic field abandoned raw character accuracy years ago and now scores extraction as field-level F1, precisely because character accuracy hides the errors that matter.

What is Character Error Rate (CER)?

Character Error Rate is the edit distance between the machine output and the ground truth, normalized by the length of the ground truth. Formally, CER equals substitutions plus deletions plus insertions, divided by the number of characters in the reference text. Word Error Rate is the same formula computed on words instead of characters.

Because one wrong character invalidates an entire word, WER typically runs three to four times higher than CER. And because insertions can outnumber the reference characters, CER can exceed 100 percent, an edge case that surprises people the first time they see it.

Why 99 percent accuracy still gets a third of receipts wrong

This is arithmetic, not a study, and you can check it yourself. Assume a system with 99 percent per-character accuracy, and assume errors are roughly independent across characters.

• A total like 1234.56 has seven characters. The probability all seven are right is 0.99 to the seventh power, which is 0.932. So about 7 percent of totals contain at least one wrong character.
• The numbers that matter on a receipt, line item prices, subtotal, tax, total, and the date, easily run to 40 characters. The probability all 40 are right is 0.99 to the fortieth power, which is 0.669. So roughly 33 percent of receipts contain at least one error in a number that matters.
• A whole receipt of around 250 characters comes out perfect with probability 0.99 to the two hundred fiftieth power, or 0.081. Only about 8 percent of documents are flawless.

A truthful claim of 99 percent character accuracy is therefore entirely compatible with getting a third of your key numbers wrong. Nobody is lying. The metric altitude is doing all the work. When you see an accuracy percentage, the only useful follow-up question is: accuracy of what, measured how?

How accurate is OCR on receipts, according to published benchmarks?

Three benchmark datasets dominate this literature, and all three score field-level F1 rather than character accuracy.

SROIE, the ICDAR 2019 Scanned Receipt OCR and Information Extraction challenge, uses 626 training and 347 test receipt images. Its extraction task pulls four fields: company, date, address, and total. The original LayoutLM reported an F1 of 0.9524 on it, and leaderboard leaders reach roughly 98 percent. Four well-structured header fields is an easy task.

CORD, the Consolidated Receipt Dataset, is harder: 30 fine-grained fields under a menu, subtotal, and total hierarchy. FUNSD covers noisy scanned forms rather than receipts.

From the LayoutLMv3 paper (Huang and colleagues, ACM Multimedia 2022), the reported F1 scores tell a clear story about how much layout understanding matters:

Donut, an OCR-free document transformer (Kim and colleagues, ECCV 2022), reports a field-level F1 near 0.84 on CORD, lower because it solves a harder problem without a separate OCR stage.

Keep the caveat attached to every number above: these are clean, well-scanned academic images. They are not thermal receipts photographed in a parked car.

How accurate is Tesseract OCR?

The best evidence is peer reviewed. Hegghammer's 2022 benchmarking experiment in the Journal of Computational Social Science tested 322 English and 100 Arabic scans, each replicated 43 times with different artificial noise, producing over 18,000 documents. The finding: Google Document AI delivered the best results, the cloud processors substantially outperformed Tesseract, and the gap widened as documents got noisier.

That direction is the durable takeaway. Precise Tesseract accuracy percentages circulate widely on vendor blogs and aggregator sites, but they trace back to no primary study, so treat any specific figure you see with suspicion. What you can rely on: Tesseract is respectable on clean, high-resolution printed text and it degrades faster than modern systems on the documents people actually photograph.

Which OCR is most accurate on invoices?

The most useful cross-vendor comparison available is an industry benchmark from Businessware Technologies, published in January 2025, on scanned invoices spanning several layouts. It is not peer reviewed, and the sample size and scoring methodology were not disclosed, so read it as a signal rather than a verdict. It is valuable because it separates header fields from line items.

Look at the second column. Every system except Azure and Textract falls off a cliff on line items. That collapse is the single most important thing to know before you pick an extraction tool.

Why are line items so much harder than the total?

Header fields appear once, in roughly stable positions, so pulling them reduces to key-value extraction. Line items require table structure recognition. The model must detect the table, segment rows and columns, handle descriptions that wrap onto a second line, stitch tables that break across pages, and group cells into the correct item, all before it assigns a single value. Column order, column count, and wrapping vary by vendor.

The academic picture is more nuanced and worth reporting honestly. The DocILE benchmark (Simsa and colleagues, 2023) measures key information extraction and line item recognition separately, and finds them roughly equal: LayoutLMv3 scores 0.698 on key information and 0.721 on line items. That is because DocILE's per-field metric gives credit for partially correct line items, whereas the commercial benchmark above scores whole-table correctness. Both are legitimate. In production, where you need the whole table right to match a purchase order, line items are harder.

Can OCR read handwriting?

Not reliably, and not with ordinary OCR. Handwriting needs handwritten text recognition, a different class of model. The canonical benchmark is the IAM Handwriting Database, and modern systems reach a character error rate below about 4 percent on it. One recent architecture reports a test CER of 4.0 percent alongside a word error rate of 12.9 percent, a vivid demonstration of the CER-to-WER gap described earlier.

Vision language models are markedly worse at this. In a 2023 evaluation of GPT-4V's OCR capabilities (Shi and colleagues), the model scored a 9.84 percent word error rate on IAM English handwriting, well behind dedicated systems, and performed poorly on tables and non-Latin scripts.

For receipts, the practical rule: a handwritten total or a handwritten tip line needs human eyes. Even four percent character error means a wrong digit lands in amounts often enough to matter, and a wrong amount is a hard failure.

Are AI models better than traditional OCR at reading receipts?

Better at understanding, worse at transcribing, and the evidence is thinner than the marketing suggests.

Vision language models are weak at raw text recognition. The GPT-4V evaluation above found 62 to 88 percent word accuracy on scene text and significant underperformance against specialized supervised OCR. Yet on the invoice benchmark, GPT-4o with an OCR front end posted the highest field accuracy of any system tested, at 98 percent, while scoring only 57 percent on line items. Language models are good at knowing which number is the total. They are not good at reading a table.

A 2026 preprint called ReceiptBench, which has not yet been peer reviewed, evaluates 10,656 real receipt images across 19 fields and reports F1 scores near 0.74 for Gemini 3 Pro and 0.71 for GPT-5, with structural parsing the weakest sub-task for every model tested. Treat those figures as provisional.

The honest summary is that independent, apples-to-apples, peer-reviewed comparisons of vision language models against specialized document models, on receipts specifically, largely do not exist yet. Anyone who tells you otherwise is quoting a vendor.

What actually breaks accuracy on real receipts

The failure modes are well understood, and every one of them attacks the recognition layer, where the arithmetic in the second section then amplifies small character errors into large field errors.

• Thermal paper fading. The dominant receipt problem. Thermal print oxidizes with heat, light, and time, leaving low-contrast gray text on gray paper.
• Crumpling and folds. Fold lines create false strokes and split characters across the crease.
• Low light and motion blur from phone photos taken quickly.
• Skew and perspective. Photographing at an angle, rather than flat, measurably degrades recognition.
• Glare. Specular highlights blow characters out to white, which the model reads as nothing at all.
• Non-standard merchant formats. Every point-of-sale system prints differently, which is why template-based tools fail on layouts they have not seen.
• Currency and decimal confusion. A thousands separator read as a decimal point turns 1,234.56 into something else entirely.
• Ambiguous dates. 03/04/2026 is March 4 or April 3 depending on locale, and no amount of pixel quality resolves it.

How do I improve OCR accuracy?

Most of the gain is in capture, not in the model. Photograph the receipt flat and square rather than at an angle. Get even light and avoid glare. Fill the frame with the document. Capture receipts the day you get them, before thermal print fades. Scan at reasonable resolution rather than relying on a compressed phone thumbnail.

Then choose a tool that reads layout rather than matching templates, so a new merchant needs no configuration. And review the output before it posts to anything. Every serious pipeline in this industry sets a confidence threshold and routes low-confidence fields to a person.

The most reliable accuracy improvement, though, is to avoid the photograph entirely. A growing share of business receipts arrive as email confirmations that were never printed at all, and for those you can pull the data straight out of the email rather than screenshotting it and asking a model to read pixels. No OCR step means no OCR error.

What is a good OCR accuracy for a business?

The wrong target is 100 percent automation. The right target is that mistakes announce themselves.

Here is the asymmetry that should drive your decision. A confidently wrong value is far more expensive than a blank one. A silently misread total posts an incorrect payment, corrupts the ledger, and may not be caught until reconciliation or audit. A flagged field costs somebody ten seconds. So the design goal is not maximum character accuracy. It is the highest rate of correct automatic extractions, with the errors surfacing as low-confidence flags rather than as confident mistakes.

That is why human review remains standard even at high accuracy, and why an accuracy percentage without a calibrated confidence score attached is close to meaningless. Ask a vendor for both, and ask which altitude the percentage was measured at.

How to test any receipt OCR tool in fifteen minutes

Skip the comparison tables, including the ones above. Take fifty documents that represent your real mix: the faded thermal slip, the crumpled restaurant check, the multi-page vendor invoice, the merchant nobody has heard of. Do not use the clean ones. They all pass.

Run them through, then check the fields by hand. Count field-level errors, not character errors, and count line items separately from headers. Fifteen minutes of this beats every benchmark on the internet, because it measures the only distribution that matters, which is yours.

If you want to run that test now, upload a batch to receipt OCR software and read the extracted fields, or push vendor bills through invoice OCR software to see how the line items come back. Developers comparing engines can hit the receipt OCR API directly. If you are shopping the market, the Veryfi alternative and Nanonets alternative comparisons cover how the API and platform vendors price the same capability, and what OCR means in accounting explains where extraction sits in a bookkeeping workflow.

Benchmark figures cited here are attributed to their original publications and dates. Vendor self-reported accuracy claims, including any published on this site, should be verified against your own documents.