Two ways a PDF editor can implement “edit.” Almost every commercial tool picks the same one — and that pick is the reason your tables drift when you add a row.
Overlay editing. The page is treated as a finished image. Your edit is a sticker placed on top: a new text fragment positioned at coordinates the editor guesses from where you clicked. The page underneath doesn’t move. This is what Adobe Acrobat, pdfFiller, Smallpdf, Sejda, and DocHub do.
Structural editing. The editor first reconstructs the document as a model — paragraphs that have an order, tables with rows and columns, headers that own their pages. Edits modify the model; the layout engine recomputes positions. This is how Google Docs, Word, and the web work. Until recently, no PDF editor worked this way.
The taxonomy explains every weird thing your PDF editor does.
What overlay editing actually does
Acrobat shows you blue outlines around text and calls it “editable.” It looks like a word processor. It isn’t. The blue outlines are clusters of text fragments that share approximate y-coordinates. When you type into one, Acrobat updates the string at that position and stops thinking about it. There’s no model that knows “this is a paragraph” or “this paragraph belongs to section 4.” So nothing else can possibly move when your edit changes the text width.
For typo fixes — same word length in, same word length out — this is invisible. The illusion of editing holds. The first time you replace a short word with a longer one, the illusion breaks: text bleeds into whatever’s next to it, and there’s no surrounding context to push around because the editor never built one.
This isn’t a bug. The reason every PDF editor reaches for overlays is in why PDF editors can't reflow: the format itself stores text as positioned drawing commands, with no semantic layer. Building anything else means reverse-engineering that layer.
A scenario that exposes the limit
The clearest test is a multi-line cell in a table. Take any invoice with a “Description” column. Pick a service named “SEO” — three characters wide. Replace it with “Search Engine Optimization & Content Strategy.” About thirty times wider.
In overlay editors the new string runs past the cell boundary into the “Quantity” column, then through “Quantity” and into “Rate.” The “Total” row below stays exactly where it was, now visually intersected by the bleeding text. Nothing in the editor’s data model says “if a cell grows, the column should grow too” — because nothing in its data model says “this is a cell” or “this is a column.”
In a structural editor, that same edit widens the description column, narrows the others proportionally, may push the row to taller height if the new text wraps, and shifts every row below by exactly the new height. Subtotal, tax, total, footer — all reposition. Page breaks recompute. If the table now spills onto a second page, the header row repeats there.
That last part — the header repeats — sounds like a small detail. It’s the single hardest part of building structural PDF editing.
Why structural editing took two decades
It wasn’t that nobody tried. Building it requires solving three problems at once that everyone else has been solving separately: AI that recognizes structure from positioned text fragments (the input direction), a layout engine implementing modern CSS specs from scratch (the output direction), and pagination integrated with layout instead of running after it (the part that breaks every existing library on multi-page flex containers). Each is hard individually. Doing one without the other two leaves you stuck either way — that’s the technical reason structural PDF editing didn't exist before.
What changes for you
The short version: overlay editing is fine when you’re not actually editing — annotations, signatures, fillable form fields, single-character corrections in same-width text. Anything that touches structure (rows, columns, paragraph length, page breaks) hits the wall the moment you try.
For real document work — invoices that need a new line item, contracts that need a clause inserted, reports with multi-page tables — structural editing isn’t a nicer-to-have. It’s the difference between a 30-second change and a 30-minute manual repositioning job that still looks edited at the end.