Extract a Table From a Screenshot — What Actually Works
Short answer: text recognition gives you text, not cells. You get one line per row, with the column gaps collapsed or inconsistent, so the real work is a second step — splitting those lines into columns in a spreadsheet. The good news is that second step takes about thirty seconds once the rows come out cleanly.
This page covers why tables are the hardest thing to pull off a screenshot, the order of fixes that works, and when cropping per column beats trying to read the whole table at once.
Why a table comes out as a wall of text
A text recogniser's job is to find characters and put them in reading order. It does that well. It has no model of what a column is, because a column is not a thing in the image — it is an absence, a vertical strip of white space that a human reads as structure.
So the reader walks left to right, top to bottom, and emits one line per visual row. What you get looks like this:
Widget A 12.50 3 37.50
Widget B 4.00 10 40.00
Widget C 128.00 1 128.00
That is genuinely the table, just flattened. The numbers are correct, the order is correct, and the structure is gone. Everything that follows is about getting the structure back cheaply.
The one thing that decides whether this works
Whether the gaps between columns survive as whitespace — and whether they are consistent.
Two things control that:
- How wide the gaps were in the original. A table with 8-pixel gaps is ambiguous to the reader; a table with 40-pixel gaps is not.
- Whether the font was monospaced. Digits in a proportional font are already different widths, so space runs wobble even when the reader does everything right.
You can change both before you capture. Zoom the page or document to 150–200% before screenshotting, and the gaps grow with everything else. If you control the source — a spreadsheet, a report you are generating — set the font to a monospaced one (Consolas, Menlo, Courier) and the columns line up so cleanly that the split is trivial.
Screen grid lines are usually a mistake. The reader sees vertical rules as characters — |, l, 1, I — and sprinkles them through the output. Before screenshotting, turn the grid off in Excel or Sheets, or use a view without borders. The empty space is a better separator than a line.
The order of fixes that actually works
1. Check whether you need recognition at all
If the table lives in a PDF, try to select a word in it first. If the text highlights, the PDF contains real text and you should copy it — extraction will be exact and take two seconds. Recognition is only for scans and images. The image format guide explains how to tell the two apart.
If the table is on a web page, select it with the mouse and copy. Same reasoning: never run recognition on text you could select.
2. Crop per column, not per table
This is the highest-value trick on the page, and almost nobody does it. Instead of screenshotting the whole table once, take one screenshot per column and run each separately.
You end up with four lists of a dozen values each. Stack them side by side in a spreadsheet, and you have your table with perfect structure and no splitting work at all. For a 4-column table of 20 rows, that is four extra screenshots — roughly forty seconds — and it removes every ambiguity about where one column ends and the next begins.
Do this when the table is wide, the numbers matter, or the first attempt came back as one undifferentiated line.
3. Otherwise, extract the whole table and split it
If the columns did survive as whitespace, the rest is a spreadsheet operation:
- Google Sheets: paste into a column, select it, then Data → Split text to columns, and choose Space as the separator.
- Excel: Data → Text to Columns → Delimited → Space, and tick Treat consecutive delimiters as one — without that tick, a three-space gap produces two empty columns.
Before you paste, format the destination columns as Text. This is the step people skip, and it is the one that quietly destroys data: an invoice number 0012345 becomes 12345 the moment it lands in a General-format cell, and 1,350.00 may become a date in some locales. Formatting first costs one click; recovering lost leading zeros costs a manual pass over every row.
4. Read the digits suspiciously
Numbers are where recognition errors hide, because several digit shapes are the same shape as letters. The pairs that matter:
| Confused with | Typical cause |
|---|---|
0 / O |
Round glyph, thin font |
1 / l / I |
All three are a vertical stroke |
5 / S |
Curvature difference of a few pixels |
8 / B |
Closed loops, similar proportions |
6 / G |
Low resolution |
2 / Z |
Diagonal stroke ambiguity |
In a table this is dangerous in a specific way: a single wrong digit does not look wrong. 1,350.00 and 1,35O.00 — the second is obviously broken. But 1,850.00 and 1,860.00 are indistinguishable without checking the source. For any number you are going to act on, total up the column and compare against the printed total. A mismatch tells you a digit is wrong even when you cannot see which one.
What genuinely does not work
Some tables will not come out, and knowing which saves an hour of retrying.
- Merged cells. A header spanning three columns has no whitespace boundary, so the reader has no way to represent it. You will get the header text once, positioned arbitrarily.
- Rotated headers. Vertical or diagonal column labels are read as scattered characters or skipped entirely.
- Faint ruling lines on a scan. Grey lines read as faint marks, which the reader either ignores or promotes into strokes that merge with digits.
- Tables with no whitespace at all — a boxed grid where every cell is outlined and nothing is padded. There are no gaps to find. Crop per cell is the only route, and at that point manual entry is often faster.
- Multi-page scans with skewed pages. Each page's skew is different, so the row heights do not match and rows cannot be joined.
For these, the honest answer is: retype it, or find the original file. Recognition is not the bottleneck — the source is.
A note on column order and multicolumn layouts
There is one more trap worth naming, because it produces output that looks plausible and is wrong. In a wide table, the reader sometimes runs across the full width of a row and then wraps, so a row becomes two output lines. If you then split on spaces, you get half a row in each line and everything after it is shifted by one — every value lands in the wrong column.
The symptom is easy to spot: the row count in your output is roughly double what it should be. The fix is the same as step 2: crop per column.
What to do next
If the earlier advice about format and contrast is new, why OCR comes out messy covers the six causes and how to diagnose which one you have from the output itself. If you are moving the numbers into a spreadsheet, converting an image to text in Excel walks through the same splitting step with the leading-zero trap in more detail.
And if you are on a Mac, taking the screenshot properly — tight crop, grid lines off, zoomed to 150% — removes more of this problem than any setting on this side of the process.