Skip to main content
NexaLink

Benchmark

Receipt scanning accuracy benchmark 2026: what an on-device OCR pipeline gets right and wrong on 20 receipt types

Most receipt apps advertise an accuracy percentage with no test set behind it. This page publishes the test set, the exact checks, the pass count and every known miss for NexaLink Scanner's on-device pipeline, measured on 14 and 15 November 2026.

By Sachin Anand

Founder, NexaLink

Published November 15, 2026

Every figure links to its source in the list at the end, with the date it was checked (2026-11-15). Where a number is NexaLink's own measurement, the method is described on the page. We make NexaLink Scanner, and say so where it appears.

Short answer

On a fixture set of 20 receipt types (17 generated images plus 3 originals), NexaLink Scanner's parser matched 146 of 169 field expectations (86.4%) on saved Apple Vision OCR output; 11 of the 20 receipts matched every field. Totals were correct on 18 of 20 receipts and no wrong total was produced: the two misses (a Japanese yen receipt and a refund) returned no total, which the app then asks you to type and confirm. In a device run on 14 November 2026, all 33 hard checks on merchant, total and currency across 15 receipt kinds passed.

Key takeaways

  • Field-level accuracy on the fixture set: merchant 18/20, date 16/20, subtotal 14/18, tax 15/20, total 18/20, currency 18/20, card last four 19/20, category 10/11, confidence decision 18/20 (NexaLink measurement, 15 Nov 2026).
  • Every one of the 23 misses is a documented, reproducible gap with a stated cause (table below), not a statistical residue.
  • The hardest receipts are not faded or skewed paper: they are a Japanese receipt (OCR runs with Latin-script languages), an Australian invoice with a bare "$", a German VAT table and a refund with negative amounts.
  • The fixtures are generated images with real OCR output, so the benchmark measures parsing on clean-to-moderately-degraded receipts; it does not measure camera quality, lighting or handwriting beyond one case.
  • No competitor was measured. The page exists so that the method can be repeated against any app.

What was measured, and how

Two measurements are reported. The first is a field-level comparison run on 15 November 2026: the production parsing pipeline of NexaLink Scanner (document flattening is skipped; text recognition is the saved output of Apple's Vision framework, VNRecognizeTextRequest at the accurate level with language correction on, recorded on macOS 26.5[1]) is run over 20 receipt fixtures and each extracted field is compared with the value a person reads off the receipt. The comparison is the repository's Jest test suite[2], which prints an expected / actual table for every field.

The second is a device run on 14 November 2026: a Maestro end-to-end flow[3] on an iOS Simulator, with a standalone staging build, imports 15 single-image receipts one by one through the real app (photo import → flatten → on-device OCR → parser → Confirm screen) and asserts what the Confirm screen shows for merchant, total and currency.

The fixture set

17 receipts were generated as images with macOS system fonts from drawn data (merchant, date, line items, tax, total, card line), so the correct answer is known exactly and no expectation was ever edited to make a test pass. Three older fixtures are included: a clear USD coffee receipt, a hardware receipt with no currency symbol, and a photo with no text at all. Two receipts span two pages (a hotel folio and a 42-item supermarket receipt photographed in two overlapping sections). Kinds covered: US grocery thermal roll, restaurant check with a handwritten tip, fuel pump with three-decimal figures, UK till receipt with a VAT analysis block, German bakery receipt with comma decimals and two VAT rates, Indian GST tax invoice, Canadian HST receipt, Australian tax invoice with GST included, Japanese café receipt in yen, online order screenshot, small parking slip on a dark background, faded low-contrast thermal print, a photo rotated about 8° with perspective, a card-terminal slip with only an amount, and a refund credit note.

Fields and pass criteria

Nine fields are checked where the receipt carries them: merchant, date (ISO), subtotal, tax, total, currency, card last four digits, expense category, and the confidence decision — whether the app saves the total straight away ("high") or blocks saving until the user confirms it ("confirm"). Some receipts have no subtotal or category expectation, which is why the denominators below differ. A field passes only on an exact match (merchant comparison ignores case and punctuation).

Results: 146 of 169 fields, 11 of 20 receipts fully correct

Field-level results on the 20-receipt fixture set (saved Apple Vision OCR, macOS 26.5)
FieldMatchedCheckedShareWhere it missed
Merchant182090%Japanese receipt (Latin-only OCR); online-order screenshot (wordmark merged with a nav label)
Date162080%Hotel folio (arrival date chosen over printed date); Australian invoice (read month-first after a USD guess); Japanese receipt (年月日 format); refund (original receipt date outranked the refund date)
Subtotal141878%German VAT table; online order ("Item(s) Subtotal"); rotated photo (row shift); refund (negative amount)
Tax152075%German VAT table; Australian "GST included in total"; Japanese receipt; rotated photo; refund
Total182090%Japanese receipt (no ¥ survives OCR); refund (negative totals ignored). No incorrect total was produced.
Currency182090%Australian invoice (bare "$" read as USD); Japanese receipt (default currency used)
Card last four192095%Hardware receipt (Vision splits "VISA ****" and ": 4417")
Category101191%Parking slip ("CITYPARK" matches no travel keyword)
Confidence decision182090%German bakery and rotated café receipts: printed totals sent to confirm (over-cautious, not wrong)
All fields14616986.4%23 known gaps, each with a stated cause
Denominators differ because not every receipt carries every field (two have no subtotal, nine have no category expectation). Source: NexaLink fixture report, 15 November 2026; method above.
Per-receipt results (fields matched / fields checked)
Receipt typeResultKnown gaps
US grocery, thermal roll, 18 items8 / 8—
Restaurant check, handwritten tip and total9 / 9— (handwritten 117.17 read; total correctly routed to confirm)
Fuel pump, gallons × price per gallon9 / 9— (3-decimal figures not mistaken for the total)
Hotel folio, two pages, totals on page 28 / 9Date: arrival date taken instead of printed date
Long supermarket receipt, 42 items, two photos8 / 8—
UK till receipt, GBP, VAT analysis block9 / 9—
German bakery, EUR, comma decimals, two VAT rates6 / 9Subtotal and tax sit under table headers; total sent to confirm
Indian GST tax invoice, INR, CGST + SGST8 / 8—
Canadian receipt, HST 13%, "TOTAL CAD"8 / 8—
Australian tax invoice, bare "$", GST included5 / 8Currency USD not AUD; date read month-first; GST line dropped
Japanese café receipt, JPY3 / 8Merchant, date, tax, total, currency: Japanese text not read (Latin-script OCR)
Online order confirmation, phone screenshot6 / 8Merchant merged with nav label; "Item(s) Subtotal" skipped
Small parking slip, dark background, no "total" word8 / 9Category not inferred
Faded, low-contrast thermal print8 / 8— (total read with high confidence)
Photo rotated ~8° with perspective6 / 9Row shift pairs labels with the amount above; total sent to confirm
Card-terminal slip, amount and last four only8 / 8—
Refund / credit note, negative amounts4 / 8Negative total, subtotal and tax ignored; wrong date chosen
Clear USD coffee receipt (original fixture)9 / 9—
Hardware receipt, no currency symbol (original)8 / 9Card last four split by OCR; total correctly sent to confirm
Photo with no text (original)8 / 8— (every field blank, confirm requested)
The rotated-photo gaps were measured without the app's flatten step; on a device the dewarp runs first and the 14 November run read that receipt's merchant, total and currency correctly.

Device run, 14 November 2026: 33 hard checks, all passed

The end-to-end flow pushes each of 15 single-image receipts into the simulator's photo library, imports it through the app and reads the Confirm screen. For every receipt, the fields the parser is expected to read are hard assertions (the run fails if one is wrong); fields already listed as known gaps are checked as warnings so the report shows them without hiding a new failure. On 14 November 2026 at 11:16, on a standalone staging simulator build, all 15 receipt kinds passed every hard check — 33 field assertions on merchant, total and currency — and 6 of the 12 warning checks fired. The flow's final housekeeping step (returning to Home) failed and was replaced by an app relaunch; that step tests navigation, not reading, and the flow has not yet been re-run since the change.

  • Hard checks per receipt: 3 (merchant, total, currency) on fuel, restaurant, German, parking and rotated receipts; 2 on the US grocery, UK, Indian, Canadian, online-order, faded, card-terminal and Australian receipts, and on the refund; 0 on the Japanese receipt, where nothing usable is read today.
  • A receipt whose total is not read (Japanese, refund) must leave Save disabled; both did.
  • The handwritten restaurant total and the hardware receipt without a currency symbol must stop at the confirm step; both did.

The 23 known gaps, by cause

Causes of every miss on the fixture set (15 November 2026)
CauseFields affectedStatus
Non-Latin script: OCR runs with Vision's default Latin languages, so Japanese labels (合計, 内消費税, 年月日) and the ¥ sign are not read5 (Japanese receipt)Language configuration work; not fixed
Negative amounts: refunds are ignored by the subtotal, tax and total finders, and the original receipt's date outranks the refund date4 (refund)Parser rule; not fixed
Table layouts: amounts sit under a header row (German MwSt table) rather than beside a label3 (German bakery)Parser rule; not fixed
Currency hints: a bare "$" is treated as USD; ABN, "TAX INVOICE", GST and an Australian postcode are not used as an AUD hint, which also flips the date order3 (Australian invoice)Parser rule; not fixed
Row grouping on a rotated photo pairs each label with the amount above it (measured without the flatten step)3 (rotated café)On-device flatten likely hides it; not verified in Jest
Label guards that are too strict: "Item(s) Subtotal:", "GST INCLUDED IN TOTAL", a colon in the card line, "parking" as part of a word4 (online order, Australian, hardware, parking)Parser rule; not fixed
Date preference on multi-date documents (hotel folio: arrival vs printed)1 (hotel folio)Parser rule; not fixed
A gap is removed from this list only when the fixture test turns green for that field; the test is written so that a fixed gap fails until its entry is deleted.

What this benchmark does not tell you

  • It is one app. No competitor was run through the same fixtures. Vendors' advertised accuracy figures could not be compared because none we found publishes a test set.
  • The images are generated, not photographed, apart from simulated skew, fade and a dark background. Real-world blur, creases, shadows and handwriting beyond one tip line are not represented.
  • OCR output was recorded on macOS, not on a phone, for the field-level table. Apple Vision on a device can read a page slightly differently, and the app flattens the image first; the device run exists to catch that difference and on 14 November found none on the hard checks.
  • Percentages are over fields, not receipts. 86.4% of fields matched; 55% of receipts (11 of 20) were perfect. Quote whichever you mean.
  • Small sample. 20 receipts is enough to list failure modes precisely; it is not enough to estimate accuracy to a decimal place in the wild. We publish one decimal only because the arithmetic is exact for this set.

What the app does with uncertainty matters as much as the raw hit rate: a low-confidence total cannot be saved until you confirm it, and when nothing is read, the receipt image is still kept and the total is typed. Read how the pipeline fits into a tax workflow in how to scan receipts for taxes, compare apps in the receipt scanner apps guide, or see plans on the NexaLink Scanner page.

Frequently asked questions

What accuracy does NexaLink Scanner claim?

We do not publish a single accuracy percentage as a claim. On the 20-receipt fixture set described here, 146 of 169 field expectations (86.4%) matched on 15 November 2026, and 18 of 20 totals were correct with no incorrect total produced. On other receipts your results will differ.

Is the OCR done on the phone?

Yes. Text recognition uses Apple's Vision framework on iOS and runs on the device; the fixture OCR for the field-level table was recorded with the same API on macOS. Receipts are not uploaded for reading.

Why is the Japanese receipt so poor?

The app currently runs Vision with its default Latin-script languages, so Japanese labels and the yen sign are not recognised; without a currency marker, integers are not treated as money. It is listed as a known gap, not hidden.

Do faded thermal receipts scan?

The faded pharmacy fixture, which washes out towards the bottom, matched on all eight fields including a high-confidence total. Real receipts can be fainter than that; see the thermal fading page for how fast they degrade and what to do.

Can I reproduce this?

The method is written out above: generated fixtures with known values, saved OCR output, a field-by-field comparison and a device flow that reads the Confirm screen. The fixture images and expectations live in NexaLink's private repository; the structure is simple to recreate for any app.

Will the numbers change?

Yes, in both directions as the parser changes and as Apple's Vision model changes with OS updates. The page states the date of each measurement; we will update it when the fixture set or results change.

Sources

  1. Apple Developer — VNRecognizeTextRequest (Vision framework text recognition) — checked on 2026-11-15
  2. Jest — JavaScript testing framework — checked on 2026-11-15
  3. Maestro — mobile UI testing framework documentation — checked on 2026-11-15

NexaLink Scanner

Every receipt captured, checked and ready for tax time.

50 receipts free, read on your phone, with CSV export on every plan. Pro $39.99/yr · Premium $59.99/yr.