Does editGPT Preserve Citations and Track Changes? A DOCX Test You Can Run

Does editGPT preserve citations? Its public Word workflow says you can import a Word document, edit it with AI, and export a DOCX with Track Changes. That means the output can show a trail of edits. It does not, by itself, prove that citations, numbers, tables, formatting, or revisions already present in the file were preserved. Those need a before-and-after DOCX test.

That distinction matters most for academic writing. Track Changes can tell you that a sentence was replaced. It cannot tell you that a citation still supports the same claim, that 0.48 kept its role as a coefficient rather than becoming a sample proportion, or that an old supervisor deletion remained separate from a new AI edit.

The same page names bold and italics as examples of formatting support. It does not turn the phrase that follows, and more, into a specification for every Word object. A returned .docx and an intact academic document are two different acceptance gates.

I would test them separately.

---

Start with a disposable copy of a real document that you are authorized to upload. Do not use a clean paragraph. Give the file the same kinds of fragile objects that your actual submission contains.

Put these six things in the sample:

1. One real citation attached to one specific claim, plus its reference-list entry.
2. Two numbers with different jobs, such as a sample size and a measured result with a sign, decimal, unit, and date.
3. A small table with a known row and column count and two cells you can identify by coordinates.
4. One pending Word insertion and one pending deletion created before the import.
5. One bold span and one italic span, each applied to a distinctive phrase.
6. A plain sentence outside the intended edit area, so you can tell whether the request changed more than expected.

Keep the original untouched. Record the filename, file size, Word version, and whether the document opens without a repair warning. Open the Reviewing pane and take a screenshot of the two existing revisions. This is your baseline, not busywork. Without it, a returned file can look fine while quietly changing the history underneath.

The citation needs more than a string check. Record the final six words of the claim it supports and the citation immediately following that claim. After the round trip, verify the author, year, locator, punctuation, and reference entry. Then check the attachment: does the citation still follow the same claim?

A citation can survive character for character and still fail if it moves to the next sentence.

Do the same for numbers. Write down each value, its unit, and its sentence role. 128 participants and 12.8 percent share digits but make different claims. A fluent rewrite that swaps their roles is not preserved just because both numbers remain somewhere in the file.

For the table, record its dimensions and two sentinel cells, for example row 2, column 1 and row 3, column 3. After export, count the rows and columns again. Check the two cells by position, not by searching for their text elsewhere in the document. If the table contains a subtotal, recalculate it from the returned cells.

---

Use one bounded edit request. Do not change the mode, prompt, and file at the same time. If the interface offers settings, record them before you start. Export the result as DOCX, then inspect the file in Word rather than relying on a browser preview.

Here is the ledger I would use:

Check	Baseline to record	Pass condition in the returned DOCX
Citation attachment	Exact citation, supported claim ending, and reference entry	Citation is unchanged and still belongs to the same claim
Numeric roles	Value, sign, decimal, unit, date, and sentence role	Every token and its meaning remain attached
Table geometry	Row count, column count, two coordinate-addressed cells, and any subtotal	Shape and cell positions match; subtotal still recalculates
Bold and italics	Exact styled phrases	Each style remains on the same words
Existing Track Changes	One old insertion and one old deletion, with anchor text and author label	Both remain pending and distinguishable from new edits
Returned file state	DOCX suffix, clean open, editable state, and original file size	File opens without repair and can still be edited, saved, closed, and reopened
Scope sentinel	Exact sentence outside the requested edit	Sentence remains unchanged

The revision-history row deserves its own pass. In Word's Reviewing pane, separate edits that existed before upload from edits introduced during the editGPT round trip. Confirm that the old insertion and deletion are still pending, still attached to the same text, and still usable. Then accept or reject one new edit, save the document under a new name, close Word, and reopen it.

That last reopen catches a different failure class. A file may display correctly once but lose editability or ask Word to repair its contents after saving. The deliverable is not finished until it survives that cycle.

Do not collapse the ledger into one score. Six clean rows do not cancel a moved citation. The failed object is the result. It tells you whether this workflow fits the document you actually need to submit.

---

This test also has limits. One passing DOCX supports one file shape, one request, one set of options, and one product state. It does not prove that every reference manager, field, footnote, equation, comment, or complex table will behave the same way. Add those objects when they matter to your own document.

Use non-sensitive material unless your rules allow the upload. editGPT's Privacy Policy says submitted text and files are used to provide the service. It separately says documents, files, and submitted text are not used for training. Its deletion statement is narrower: when a user deletes a document or file, that document or file is permanently removed from the servers. Those are the vendor's policy statements, not an independent security or deletion audit. Your course, client, publisher, or institution may impose a stricter rule.

I did not upload a file to editGPT, run its editor, or test an export for this article. There is no preservation result or detector result to report. The public claim is specific enough to design a fair test. Only the returned file can answer whether your citation, table, revisions, and editable document state survived it.

Track Changes helps you review what changed. The ledger tells you whether the document still means and behaves the way it did before.

Originally published at HumanPen: https://humanpen.net/blog/editgpt-ai-rewriter-review?utm_source=devto&utm_medium=article&utm_campaign=d166