A table screenshot goes through OCR, every word comes back right, and a paste into a spreadsheet would still land in the wrong columns. On plain grids a coordinate rebuild fixed this: ignore the Tabs in the export and rebuild rows and columns from the text-box coordinates. On a table with a merged header it scored worse than doing nothing. Splitting on Tab left 41 of 45 cells misplaced, the rebuild left 43, and what placed all 45 was cutting the image along its ruling lines.

The sample is one self-made table with fictional numbers, clean and not tilted: 9 rows by 6 columns, 45 non-empty cells, a title spanning all 6 columns, two group headers spanning 2 columns each, and two cells merged vertically. Recognition ran through ImgIng https://imging.ai/ at its default tier, in a desktop browser on an Apple-silicon Mac, one run on 2026-11-20. The TXT and JSON were exported, and every text box in the JSON carries four corner coordinates. "Paste" below means a script that splits each exported line on Tab and fills a grid. No spreadsheet app was involved, so a real paste is not tested.

How the same 45 text boxes were placed What came out Cells misplaced Status
Split each line on Tab 9 lines of 1, 4, 1, 1, 2, 2, 2, 2, 1 cells 41 of 45 measured
Rows by box centre, columns by projection, 80% vote 5 columns, 1 extra row 43 of 45 measured
Same script, vote relaxed to 70% 7 columns 45 of 45 measured
Cells cut along ruling lines 45 cells detected 0 of 45 measured

Tab paste misplaced 41 of 45 cells

The text was fine: all 45 strings came back intact and the export had the right 9 lines. Split on Tab, those lines gave 1, 4, 1, 1, 2, 2, 2, 2 and 1 cells for a table 6 wide. The row with the group headers came out as 4 cells, with nothing holding the place of the columns they span, and the four sub-headers arrived glued into one string. Only 4 cells landed in the right place. Across the seven sample tables in this batch, a Tab showed up where the gap between two boxes was wide, and a narrow gap got a space or nothing. Other OCR tools were not tried.

Rebuilding from coordinates got 43 wrong

The rebuild: boxes whose vertical centres sit within half a box height of each other form a row. Every box is then projected onto the x axis, and a column gap is a stretch that is empty in at least 80% of the rows and at least 0.2 box heights wide. The same script had taken a ruled 9 by 7 table from 60 of 63 misplaced to 0, and a borderless one from 60 to 0. The 0.2 was chosen after seeing results, because 0.4 merged two columns on the borderless table and got 0 of 63 right.

On the merged-header table it found 5 columns instead of 6, produced one row too many, and misplaced 43 cells, two more than the Tab split. The extra row comes from the two vertically merged cells, whose centres sit between two rows. The title spans the whole table, so the projection finds no gap in that row, which fits the missing column. That is an inference, not checked box by box.

A looser vote gave 7 columns and 0 right

Relaxing the share of rows that must be empty from 80% to 70% gave 7 columns, one more than the table has, and none of the 45 cells in the right place.

Ruling lines placed all 45 cells

This table has printed borders, and borders say where the cells are. The second script ignores text spacing: it binarises the image, keeps the horizontal and vertical lines, treats each white region enclosed by lines as a cell, and puts every text box into the region containing its centre. It detected 45 cells and misplaced none.

The catch is in the name: no lines, no cells. A merged header on a borderless table is not among the samples, so it is not tested. A cheap alarm before trusting any OCR table is counting the Tabs on each exported line: if the lines disagree, the columns are already wrong. Borders mean cutting along the lines. A plain grid with no spanning cells went through the coordinate rebuild at 0 misplaced here. For a spanning header without borders, rebuilding only the body rows first is an idea that was not run.