JustOCR guide
Annual report OCR: test the numbers, headers and source pages
Test annual report OCR on year columns, Group and Company figures, footnotes, signs and source pages before trusting extracted financial data in your workflow.
Evaluate annual report OCR by the cost of producing figures you can check against the source. A file can look clean and still assign a revenue figure to the wrong year. Financial teams need a test that checks labels and source evidence as well as digits.
JustOCR offers page-based pricing, Markdown, JSON and source previews for a reading-and-review workflow.1 Its semantic Markdown/HTML tables have whole-table source boxes, without explicit per-cell objects, geometry or confidence. For automatic cell-level highlighting, evaluate Textract or Azure's cell schema too.740
Use these checks to measure your own sample. They aren't reported benchmark results.
Build the sample from documents you'll buy
Choose a representative annual report or audited financial statements (AFS) document that you're permitted to process. Include a scanned report if the real workload contains scans; don't test only clean digital PDFs and extrapolate to every source.
A practical starting pack is 30 pages, matching JustOCR's published job minimum.1 Select pages for difficulty, rather than taking the first 30: a two-column management discussion, a statement of financial position, cash flows, an accounting-policy page and several notes. Add one continuation page whose table header appears only on the preceding page.
Write an answer sheet from the originals before running OCR. For each selected figure, record the value as displayed, its label, year, currency, scale, Group/Company category, note reference and PDF page. A reviewer who didn't prepare the sheet should check the outputs against it.
Keep your selection fixed across providers. Record the model or processor version, output options and any preprocessing. Otherwise a second run with different pages or settings can't explain a changed result.
Reading order can change the meaning
On a two-column page, test whether text follows one column before moving to the next. Watch for a sentence that ends in the other column, a sidebar inserted into an accounting policy, or repeated headers entering the body.
Search for a phrase near the bottom of the left column and inspect the next paragraph. Then repeat on a page containing a table and footnotes. A readable first paragraph proves little about the rest of that layout.
Keep the printed page number and the PDF page position separately. Front matter can make them differ. If a research note cites page 104, its link should reopen the intended source page rather than whichever batch index happens to match.
Financial tables need association checks
Use a statement with multiple year columns and Group/Company headers. Confirm that each amount stays under the correct combination. Then inspect wrapped labels: a two-line expense name must remain one row, without shifting the numbers beneath it.
Merged headers deserve a separate test. Azure exposes row and column spans; Textract represents merged cells explicitly.740 Google's Form Parser documentation limits its tables to conventional cells without row or column spanning.19 Those are capability differences, not an accuracy ranking for your report.
For JustOCR, inspect the rendered grid and downloaded JSON together. Check that a multirow header still identifies the year and reporting entity for each amount. Record any repair needed before importing the figures.
Signs, units and footnotes can defeat clean digits
Create checks for parentheses around losses, a minus sign, a dash in a comparative year and a decimal in earnings per share. Treat a dash as an unresolved source symbol until your accounting rule defines it; don't silently turn every dash into zero.
Record whether a statement reports units, thousands or millions. Preserve the currency and scale with the imported values. A correct sequence of digits with the wrong multiplier can corrupt the analysis without looking like an OCR mistake.
Pick a superscript footnote and follow it to the corresponding disclosure. Check whether the note number remains separate from the amount. For a continuation table, confirm that the earlier header still governs the later rows. Keep that link in your review notes even if the service returns each page separately.
Make the output useful beyond the preview
Save both recognized text and source images. JustOCR's published output set includes Markdown, JSON and page previews, while its workspace documentation describes per-page artifact links and a whole-job ZIP.12 Check whether the files retain the original document/page mapping after download.
For a spreadsheet workflow, test the actual import or paste operation. Don't assume a rendered table implies native XLSX export or correct expansion of merged cells. Recheck the year columns after pasting.
For search or retrieval, ask a question whose answer spans a policy paragraph and a note table. Require the answer to cite the source page. Keep OCR review and financial interpretation as separate steps; a recognizer doesn't establish an accounting conclusion.
Score corrections before choosing a supplier
Count missing text separately from wrong digits and wrong column associations. Record how many selected values need intervention and how long the reviewer spends correcting them. Use your own agreed thresholds; a single error in a valuation-critical input can matter more than dozens of harmless header mistakes.
Only compare cost after completing that review. Page fees, setup work, export cleanup and recurring corrections all belong in the buying decision.
Try JustOCR with a public annual report or a redacted document you're permitted to process. Review current pricing and follow the workspace's submission instructions. Run the sample through the full download-and-review workflow before moving a larger report batch.