Results published · 3 fixtures approvedProtocol v1.0.0

Controlled test · InvariTech owns one tool

Best Invoice OCR Software8 tools. Same 3 documents.

We tested 8 tools using the same digital invoice, handwritten bill, and handwritten non-invoice note. We checked for usable structured results, not simply returned text.

Upload invoices free — get line items and totals as CSV See how the test was run

Verdict

Results published

InvariTech was strongest for structured invoice-to-CSV extraction.

InvariTech produced all expected line rows across the two financial documents and correctly rejected the handwritten non-invoice note. In this small owned-document test, it was the strongest option for structured invoice-to-CSV extraction.

Results published

31 Aug 2026

01 / Test corpus

Three documents. Three different checks.

Three owned documents cannot prove universal accuracy. They can show whether a tool handles a clean digital invoice, a handwritten bill, and a handwritten document that should not become an invoice.Document details are blurred to protect private information.

Digital invoice

A native or high-quality digital PDF. It has a clear header and multiple line items.

Checks whether standard invoice fields and line items are returned in a usable structure.

Handwritten bill

A real handwritten bill or receipt. Labels, numbers, spacing, and table structure vary across the page.

Checks whether handwritten fields and line items are extracted without invented rows.

Small obscured sample of the non-invoice note

Handwritten non-invoice note

A real handwritten note that is not an invoice, bill, or receipt.

Checks whether the tool rejects a document that is not an invoice instead of inventing financial data.

02 / Comparison results

8 tools. Field and line-item scores.

Each percentage is calculated against approved ground truth. Exact counts are shown where the retained structured output supported scoring.

Private values and full exports are not published. “Not scored” means the retained result could not support a defensible calculation.

Scroll sideways to see every column

Tool	Account	Speed	Digital invoice · 1 line	Handwritten bill · 14 lines	Non-invoice note	Export
InvariTech Owned by publisher	No	2m 01s for all 3 files	Core fields 8/8 · 100% Line rows 1/1 matched · 100% 1 returned Line cells 5/5 · 100% Total Correct	Core fields 4/5 · 80% Line rows 14/14 matched · 100% 14 returned Line cells 65/69 · 94.2% Total Correct Invented fields 1	Correctly rejected	Structured CSV
Thunderbit	No	~30s	Ocr text Text only; structured metrics not scored	Ocr text Text only; structured metrics not scored	Outcome unverified	OCR text, not structured CSV
InvoiceOCR.app	Yes	~2m	Failed Retained output did not support numeric scoring	Core fields 2/5 · 40% Line rows 14/14 matched · 100% 14 returned Line cells 50/69 · 72.5% Total Correct Invented fields 2	Correctly rejected	CSV
Parseur	Yes	~2m	Core fields 7/8 · 87.5% Line rows 1/1 matched · 100% 1 returned Line cells 4/5 · 80% Total Correct	Core fields 2/5 · 40% Line rows 13/14 matched · 92.9% 14 returned · 1 extra Line cells 37/69 · 53.6% Total Correct	Processing failed	CSV
Docparser	No	~30s per file, one at a time	Core fields 5/8 · 62.5% Line rows 0/1 matched · 0% 1 returned · 1 extra Line cells 0/5 · 0% Total Correct	Failed Retained output did not support numeric scoring	Processing failed	Structured demo result
Nanonets	Company email	~1-2m	Core fields 7/8 · 87.5% Line rows 1/1 matched · 100% 1 returned Line cells 4/5 · 80% Total Correct	Core fields 0/5 · 0% Line rows 2/14 matched · 14.3% 9 returned · 7 extra Line cells 4/69 · 5.8% Total Missing	False positive 0 invented fields · 32 invented rows	Structured result
PDF.co	Yes	~1–2m	Core fields 8/8 · 100% Line rows 1/1 matched · 100% 1 returned Line cells 5/5 · 100% Total Correct	Core fields 2/5 · 40% Line rows 14/14 matched · 100% 14 returned Line cells 68/69 · 98.5% Total Correct Invented fields 1 Required manual PDF conversion	Processing failed	JSON
DocuClipper	Yes	~40s	Core fields 7/8 · 87.5% Line rows 1/1 matched · 100% 1 returned Line cells 4/5 · 80% Total Correct	Core fields 1/5 · 20% Line rows 9/14 matched · 64.3% 14 returned · 5 extra Line cells 25/69 · 36.2% Total Correct Invented fields 2	False positive 0 invented fields · 1 invented rows	CSV; malformed multiline export

Core-field accuracy: Correct supplier, invoice number, date, customer, currency, subtotal, tax, and total fields ÷ the fields that apply. Example: 7/8 = 87.5%. Invented fields are shown separately.
Line-row coverage: Source line rows matched to a returned row ÷ expected source rows. Example: 13/14 = 92.9%. Returned, missing, and extra rows remain visible.
Line-cell accuracy: Correct description, quantity, unit, unit-price, and amount cells ÷ populated source cells in matched rows. Example: 65/69 = 94.2%. Blank source cells are excluded.

Benchmark 2.0.0 · exact-normalized scoring · generated 31 Aug 2026

03 / Scoring protocol

How every score was calculated.

The approved source documents are the ground truth. We compare each tool's structured result with the fields and line items present in those documents, then publish the numerator, denominator, and percentage.

Metric 01

Core-field accuracy

Correct applicable fields ÷ expected applicable fields

We compare supplier name, invoice number, invoice date, customer, currency, subtotal, tax, and total. A field enters the denominator only when it applies to that document. A missing or incorrect applicable field scores zero; a field that does not exist on the source is excluded.

Worked example: 7 correct fields out of 8 expected fields = 7 ÷ 8 = 87.5%. If the tool also invents a purchase-order number, the score stays 87.5% and the invented field is reported separately.

Metric 02

Line-row coverage

Matched source rows ÷ expected source rows

We count how many real invoice lines can be paired with a returned line. Each returned row can match no more than one source row. The pairing is chosen by the number of exact line-item cells, so a duplicate or unrelated row cannot cover two source rows.

Worked example: A bill has 14 source rows. If 13 are matched, coverage is 13 ÷ 14 = 92.9%. If the tool returns 14 rows but one is unrelated, we show 14 returned, 13 matched, 1 missing, and 1 extra.

Metric 03

Line-cell accuracy

Correct populated line-item cells ÷ expected populated cells

Within the matched rows, we compare description, quantity, unit, unit price, and amount cell by cell. Only cells populated in the approved source data enter the denominator. A missing or incorrect populated cell scores zero; a genuinely blank source cell is not penalized.

Worked example: Fourteen rows contain 69 populated cells because one source cell is blank. If 65 cells are correct, accuracy is 65 ÷ 69 = 94.2%. Returning all 14 rows does not guarantee a high cell score.

What we checked

Core fields

Supplier name, Invoice number, Invoice date, Customer, Currency, Subtotal, Tax, Total.

Each line row

Description, Quantity, Unit, Unit price, Amount.

OCR text is not structured extraction

Finding the words on a page is useful, but it does not prove that a supplier, total, or line-item value was placed in the correct field, row, and column. OCR-text-only results therefore receive no structured accuracy score.

Comparison rules

Exact after limited normalization: We ignore case, repeated whitespace, equivalent currency and decimal formatting, and unambiguous date formatting. We do not award partial credit for a similar supplier name, description, date, or amount.
Totals: Subtotal, tax, and total are compared as normalized numbers. For example, $1,250.00 and 1250 match; 1249.99 does not.
Invented output: Fields and line rows with no source counterpart are counted separately. They remain visible because an apparently complete export can still contain false financial data.
Non-invoice document: The handwritten note passes only when the tool declines to treat it as an invoice. Returned invoice fields or line items are reported as a false positive.
Not scored: We publish a numeric score only when the retained result supports the calculation. OCR text without dependable fields and rows, an empty export, or an unavailable result is labelled Not scored instead of being estimated.

Private source values, document names, and full exports are not published.

Ownership disclosure

We built one of the tools. That matters.

Disclosure: InvariTech owns one of the 8 tools in this comparison. We used the same three owned documents and assessment rules for every tool.

No vendor pays to appear. We use no affiliate links. The comparison reports observed outcomes without publishing private source data or full exports.

Limitations

3 documents can expose failures. They cannot prove universal accuracy.

Results vary with document quality, language, layout, and handwriting. Always review financial data before approving payments or posting entries.

Use the line-item review checklist

Test the document you care about

Turn invoice line items and totals into CSV.

Upload a PDF, JPG, or PNG. Check the header fields and line items against the source, then download one CSV without creating an account.

Upload invoices free — get line items and totals as CSV See how the test was run