Slipsheet · Content Back to ledger
Home / What Does an OCR Receipt Scanner Actually Read?
What Does an OCR Receipt Scanner Actually Read?

What Does an OCR Receipt Scanner Actually Read?

An OCR receipt scanner reads the text on a receipt photo or PDF and pulls out what is printed there: merchant, date, total, and tax. That is the whole job. It cannot decide the category, tell you which client the lunch belongs to, or know whether the number it found is the right one. Your books do not need a page of text. They need rows. So we measured how much of a nine-column expense row OCR can fill at all, and how much of that it gets right once the photo gets worse.

Text is not a row

Anyone who has keyed receipts into a spreadsheet knows the reading is the easy part. The yardstick here is the row we actually export: the nine columns in a Slipsheet CSV.

The nine columns of one expense row, and where each value comes from

20 receipts make 180 cells. Only 80 of them (merchant, date, total, tax) are printed on the paper. Another 40 (currency and file_name) are mechanical. The last 60 (category, client_tag, notes) are judgment calls, and no OCR engine can read them off a receipt because they are not on it. So even on a perfect scan, OCR is working on 80 of 180 cells, and it still has to get those 80 right.

The test: 20 receipts, two engines, six photo conditions

We used 20 synthetic receipts (SaaS invoices, rideshare, cafes and restaurants, hardware, office supplies) whose true totals add up to $2,034.43. Each one ran through two open-source OCR engines, Tesseract and RapidOCR, with default models and no tuning, plus one simple parser that picked merchant, date, total, and tax out of the text. A field counts as right when the merchant matches after normalizing, the date matches exactly, and the total and tax are within $0.01.

This is a test of generic open-source OCR, not of Slipsheet's extractor. None of these numbers describe how Slipsheet reads receipts. We did not run Slipsheet on this set.

The six conditions: a clean file, a 4° tilt, blur, dim light (low contrast), a chat app resend (JPEG quality 25 at half size), and a phone snap combo of tilt, blur, shade, noise, and JPEG at once. "Fields right" is out of 80 (20 receipts x 4 printed fields). "Totals off by" adds up how far the engine's 20 totals missed.

Tesseract: fields right out of 80, by photo condition. Generic OCR, not Slipsheet.

ConditionFields rightTotals off by
Clean file71$0.00
Tilted 4°61$85.44
Blur36$2,001.21
Dim light2$1,980.68
Chat app35$701.20
Phone snap17$1,558.68

RapidOCR: fields right out of 80, by photo condition. Generic OCR, not Slipsheet.

ConditionFields rightTotals off by
Clean file75$0.00
Tilted 4°52$0.00
Blur61$0.00
Dim light57$0.00
Chat app27$3,582.97
Phone snap22$7,604.65
Bar chart of printed fields read right out of 80 by photo condition. Tesseract and RapidOCR: clean file 71 and 75, phone snap combo 17 and 22. Generic open-source OCR on 20 synthetic receipts, not Slipsheet.
Printed fields read right, out of 80, for each photo condition. Built from the test data above. Generic open-source OCR, not Slipsheet.

What the numbers say

The best case still is not 80 of 80. On clean files, Tesseract got 71 and RapidOCR got 75. Tesseract read tax right on only 12 of 20 receipts, and RapidOCR got the merchant right on 15 of 20. That is the ceiling, on images better than anything you will shoot at a counter.

A real phone photo is a different story. In the phone snap combo (tilt, blur, shade, noise, and JPEG at once), Tesseract got 17 of 80 and RapidOCR got 22. Tesseract found no total at all on 17 of 20 receipts. RapidOCR returned a total on 19 of 20, but only 8 were right, and its totals were off by $7,604.65 on receipts that add up to $2,034.43.

Switching engines does not fix it. RapidOCR held up under blur and dim light (61 and 57) where Tesseract fell to 36 and 2. Tesseract did better on the 4° tilt (61 against 52) and the chat app JPEG (35 against 27). Some of that is our simple parser: on the tilt, RapidOCR read every total and 0 of 20 taxes because the lines ran together. A better engine moves the numbers around. It does not get you to a finished row.

Dates and tax go early. Under blur, Tesseract read 3 of 20 dates and 1 of 20 taxes. Through the chat app, RapidOCR read none of either. A wrong date puts the expense in the wrong month, which is the mistake you find at quarter end.

The 60 judgment cells stay blank in every condition. Category, client_tag, and notes are yours to decide whatever the photo looks like.

If you are weighing engines rather than rows, our Tesseract OCR alternative comparison covers that side.

A blank total you notice. A wrong one you don't.

A missing total jumps out of the sheet. A plausible wrong one sits there until the month will not reconcile. In a second run on the same 20 receipts (RapidOCR only, November 13, 2026), a receipt that was too small in the frame, 182 px wide, came back with 7 confidently wrong totals. Notion's $96.00 became $596.00. Home Depot's $156.88 became $5,143.37. Cutting the photo off through the TOTAL line gave 14 wrong totals out of 20, and 11 of them were the subtotal: a real number, printed on the receipt, just not the one you paid.

The full table of capture conditions, and how to shoot around them, is in our mobile receipt capture guide.

The safety step that works is boring: review the total and date before export. Look at the total against the paper and the date against the month you think it belongs to. It is the check that catches the $596.00 Notion bill before it lands in your books.

Where Slipsheet fits

Slipsheet is the step after the text. You upload receipt photos or PDFs, and each receipt becomes one row with the date, merchant, total, and tax read off the receipt, and the category, client_tag, and notes columns in the same file. One receipt in, one CSV row out. You check each row against the receipt before you export.

Every plan exports the same nine columns as one CSV named slipsheet-YYYYMMDD.csv: date, merchant, total_amount, currency, tax_amount, category, client_tag, notes, file_name. Google Sheets sync is on Pro. QuickBooks and Xero work only through importing that CSV yourself. Slipsheet does not connect to them directly.

Pricing is a flat subscription. There is a 14-day free trial with no card. Starter is $5 per month for up to 250 receipts. Pro is $12 per month for up to 1,500 receipts, plus Google Sheets sync. Yearly billing is $50 for Starter and $120 for Pro.

The test above did not measure Slipsheet. Whatever reads your receipts, the habit is the same: check the total and the date, then export the row.

Frequently asked questions

What does an OCR receipt scanner actually extract?

The text printed on the receipt, which a parser then turns into fields: usually merchant, date, total, and tax. In a nine-column expense row, that covers 4 of the 9 columns. Category, client, and notes are not printed on a receipt, so no OCR engine can read them. In our test of 20 synthetic receipts with two open-source engines, the best clean-file result was 75 of 80 printed fields right.

Why did my receipt scanner read the wrong total?

Usually because the photo made the total hard to find or hard to read. In our test of 20 synthetic receipts with an open-source OCR engine, a receipt too small in the frame turned $96.00 into $596.00, and a photo cut off through the TOTAL line returned the subtotal on 11 of 20 receipts. Fill the frame, keep the TOTAL line in, and check the total before you export. These are results for generic OCR, not for Slipsheet.

Do I still need to check scanned receipts?

Yes. Review the total and date on every row before you export. In our test, a phone snap photo dropped two open-source OCR engines to 17 and 22 of 80 printed fields right, and one engine's totals were off by $7,604.65 on receipts that add up to $2,034.43. A blank total is easy to spot. A plausible wrong one is not.

About the test

20 synthetic receipts, six photo conditions, two open-source OCR engines with default models, one simple parser. The field results come from our runs compiled on November 10, 2026, and the capture results from our run on November 13, 2026. These numbers describe generic OCR on this set only. They are not a Slipsheet accuracy number.

Ready to stop typing receipts?

Open your ledger.

14-day trial. No card. Cancel anytime. Your receipts write themselves while you do literally anything else.