PDF to Excel converter
Convert PDF tables to Excel in your browser. Rows and columns are rebuilt from where the text sits on the page, one sheet per page, with nothing uploaded.
Drop files here
or
Accepts .pdf · Up to 20 files at a time · up to 80 MB each "Upload" and "download" here mean your browser reading a file from this device and writing one back to it. Your file is never sent to a server — open your browser's network panel and watch: no request carries your file.
What you get in the spreadsheet
- One sheet per PDF page, named Page 1, Page 2 and so on. A page break in a report is usually a table break, and two sheets are easier to join than one sheet with header rows buried in it is to split.
- Numbers you can add up. A cell holding a plain number is stored as a number, so SUM works on the column straight away, and the PDF's decimal places are kept: 48.50 still reads 48.50.
- Text that must stay text. Leading zeros, thousands separators, dates and phone numbers are kept exactly as printed rather than turned into numbers that would lose them.
- Pages that are not tables come back as one column of lines, rather than prose cut into invented columns.
Fonts, colours and cell borders are not carried over; the file holds the values and nothing else, so it opens the same in Excel, Google Sheets, Numbers and LibreOffice.
Which PDFs convert well
| Kind of PDF | What to expect |
|---|---|
| Saved from Word, Excel, a browser or accounting software | Rows and columns rebuilt; 176 of 176 rows exact on our own test files |
| Reports with ruled tables (government, annual reports) | Mostly right; 93.4% on the ICDAR 2013 benchmark, with the misses mainly in multi-row headers |
| Bank statements | Use Bank statement to Excel, which adds a date, description, amount and balance layout and checks every row against the running balance |
| A wide table printed sideways on the page | Rows and columns come out the wrong way round; check by hand |
| Scanned or photographed pages | Not converted, and it says so. OCR PDF makes the text readable first, but check every cell |
Other ways to do it
| Method | Where it works | Columns kept | Upload |
|---|---|---|---|
| Copy and paste from the PDF | Everywhere | Often not: a row lands in one cell, or everything in one column | No |
| Excel: Data → Get Data → From File → From PDF | Excel for Microsoft 365 on Windows only | Yes, through Power Query | No |
| Adobe Acrobat: Export PDF → Spreadsheet | Acrobat Pro (paid); the online version uploads the file | Yes | Online version only |
| This page | Any modern browser, Mac and phone included | Yes | No |
If you have Excel for Microsoft 365 on Windows, its own import is worth trying first. This page is the option on a Mac, on a work laptop where you cannot install anything, or when the file should not leave your computer.
What it does not do
- No merged cells. A cell spanning several columns lands in the first of them.
- No "same as above". A cell that is blank in the PDF is blank here.
- No formulas. Totals arrive as the numbers printed, not as SUM formulas.
- No scanned tables, for the reason above.
How accurate it is, measured
| Test set | Result | What is in it |
|---|---|---|
| ICDAR 2013 table competition | F1 93.4% (precision 94.3%, recall 92.5%) | 158 tables in 67 PDFs from EU and US government publications |
| Our own test PDFs | 176 of 176 rows exact | 14 PDFs saved from Chrome, Word and Excel: a bank statement, an invoice, wrapped cells, a merged header, two tables on one page, Chinese text, a wide landscape page |
The benchmark does not mark a table simply right or wrong. For every cell it takes the nearest neighbour to the right and the one below, and counts how many of those pairs a converter reproduces — so one column out of place costs a few pairs rather than the whole table. Precision is the share of the pairs it produced that were right; recall is the share of the answer key's pairs it found.
Where the misses are
The score is not spread evenly. 103 of the 158 tables scored 90% or better and 27 were perfect; 33 scored below 80%. Most of the worst ones have one of two clear causes, and the rest mostly lose their points in the header:
- Pages turned on their side. A wide table printed sideways on a portrait page stores its text running vertically, so rows and columns are read the wrong way round. Tables like this score close to zero.
- Cells holding a paragraph. Forms and guidance documents where one cell carries several lines of prose. A wrapped line is joined back only when the layout proves it is a wrap; otherwise it stays a row of its own, which is safer than gluing two real rows together but shows up as missed pairs.
- Headers over several rows. A heading that spans several columns, or a header split across two lines, lands in a different column from the one the answer key puts it in, while the figures underneath come out right. These tables mostly score between 50% and 80%.
How it was measured
The corrected release of the ICDAR 2013 dataset, all 67 documents of its competition set. Each PDF was converted whole, exactly as this page converts it, without being told where on the page the table is, and cells were compared with spaces and letter case ignored. Measured on 19 October 2026 with the current version of the converter. Scanned PDFs are not part of either test: they have no text to measure, and this converter says so rather than guessing (see below).
PDF to Excel FAQ
Why not just copy the table out of the PDF, or import it in Excel?
Copying usually pastes each row into a single cell, or every value into one column, because the PDF holds text placed on a page rather than cells; splitting it back with Text to Columns goes wrong as soon as a description contains a space. Excel for Microsoft 365 on Windows can import a PDF itself, under Data → Get Data → From File → From PDF, but Excel for Mac and older versions cannot. This page works in any browser, on a Mac too, and rebuilds the columns from where the text sits.
How can it find tables when a PDF has no tables?
It cannot read them, so it measures them. A PDF stores glyph runs with a position each — the ruled lines you see are separate drawing operations, and nothing in the file says a cell exists. Rows come from shared baselines, which is how the page was typeset. Columns come from the horizontal span each cell occupies: cells whose spans overlap are the same column, and a span that recurs on several rows is a column rather than an indent.
Can I add up the numbers afterwards?
Yes — a cell that holds a plain number is stored as a number, not as text, so SUM works on the column straight away. The source's decimal places are kept as a display format, so a price written 48.50 still reads 48.50 rather than 48.5. What deliberately stays text is anything where the formatting is information: a leading zero, a thousands separator, a date, a phone number. Coercing those would quietly destroy them.
Why is each page its own sheet?
Because a page break in a report is usually a table break, and the header row repeats on the next page. Two sheets are easy to join if they turn out to be one table; a single sheet with header rows buried in the middle of it is not easy to take apart.
What happens to a page that is not a table?
It comes back as one column of lines. That is deliberate: inventing columns from a paragraph that happens to be indented would produce a grid that looks like a table, and someone would trust it. A page only becomes a grid when at least two column positions recur across several rows.
Which tables does it get wrong?
Three kinds, all for the same reason — the geometry is ambiguous. A cell whose text wraps onto a second line is joined back when the layout proves it is a wrap — one line apart, same font, the line above full to the column edge — and otherwise left as a second row rather than risk gluing two real rows together. A merged cell lands in the first column it spans. And a cell that is empty in the PDF is empty here, so a table that relies on blank cells to mean "same as above" arrives with the blanks.
Does it work on a scanned table?
No, and it says so rather than returning something plausible. Recognition gives one line of text per detection box, and those boxes merge neighbouring cells often enough that the grid would be wrong in ways that are hard to spot. Running OCR PDF first makes the text extractable, but a scanned table is the case to check by hand whatever tool you use.
Related guides
Related converters
Bank statement to Excel
One row per transaction, with every running balance checkedPDF to PNG
Every page becomes an image; several pages come back as a ZIPPDF to JPG
Every page becomes an image; several pages come back as a ZIPPDF to Word
Rebuilds headings, paragraphs, ruled tables and pictures as an editable Word fileAI text recognition, on your devicePDF to PowerPoint
One slide per page, each a picture of it — the text is not editablePDF to TXT
Pulls out the text layer, or reads a scan on your deviceAI text recognition, on your devicePDF to Markdown
Rebuilds headings, paragraphs, bullets and ruled tables as Markdown textAI text recognition, on your deviceOFX to Excel
One row per transaction, with real dates and amounts you can add up
PDF tools
- Shrink the file size by re-encoding the images inside the PDF — by level, or to a size you set, like 300 KB.
Compress PDF
- Break a document into several PDFs — one file per page, one every few pages, or by the ranges you type.
Split PDF
- Pull the pages you need into a new PDF and leave the rest behind.
Extract PDF pages
- Turn sideways or upside-down pages the right way up and save the result.
Rotate PDF