Why pasting ChatGPT into Word loses the formatting
ChatGPT writes its answers in Markdown and renders them to HTML in the browser. When you select the rendered answer and paste it into Word, Word receives that HTML, and the result looks almost right while being structurally wrong. Headings arrive as Normal paragraphs with a bigger bold font, not as Heading styles, so the navigation pane is empty and a table of contents cannot be built. Tables come in as web tables stretched to 100% width with no borders. Code blocks become paragraphs in Consolas whose leading spaces collapse, which breaks Python indentation. Nested lists flatten to one level or turn into lines that begin with a literal bullet character. Equations, which ChatGPT renders with KaTeX, paste as a jumble of letters, or as an image you cannot edit. If ChatGPT was in dark mode, you may also get white text on a white page.
The fix is to bypass the rendered HTML and convert the Markdown itself. The Copy icon under each answer puts that Markdown on the clipboard. Paste it here and every construct is mapped to the real Word object: ## to Heading 2, pipe tables to Word tables with a header row, fenced code to shaded monospace paragraphs with preserved spacing, nested bullets to Word's multilevel list numbering.
ChatGPT's Markdown dialect
ChatGPT emits GitHub Flavored Markdown with a few habits worth knowing. Answers often start at ## or ### rather than #, so by default the converter promotes headings until the top level is Heading 1; switch that off if you are assembling several answers into one document. Bold labels such as **Step 1:** are used as pseudo-headings and are kept as bold text. Nested lists are indented by two or three spaces, both of which are accepted. Inline math appears as $…$ and as \(…\), display math as $$…$$ and \[…\]; all four forms are detected. Web-search answers carry citation chips that copy as 【12†source】 tokens, which are removed by default or kept as [12]. Horizontal rules between sections are preserved, and emoji in headings are kept as characters.
Tables and equations survive
Pipe tables become native Word tables with a shaded header row and column widths estimated from the content, so a comparison table with inline code or **bold** in cells arrives editable. For equations, the free converter shows them rendered in the preview and writes the LaTeX source in a monospace run in the file. Pro converts $…$ and \(…\) into Word's native equation objects (OMML), which you can edit with Word's equation tools, and inline equations stay inline with the text baseline instead of breaking the line. See Markdown to Word with equations for the details.
How the extension adds a button to the page
The MDtoDOCX extension runs a small content script only on chatgpt.com, claude.ai, gemini.google.com and chat.deepseek.com. It watches for message containers and adds a "Word" button next to the existing Copy button under each answer. When you click it, the extension reads the same Markdown source that the site's own Copy button uses, rather than scraping the rendered HTML, so nothing is lost. The conversion runs inside the extension, on your machine, and a file named after the first heading is downloaded. There is also a PDF button. The extension does not read your conversations in the background, does not need an account, and requests no permissions beyond those four sites.
Claude to Word
Claude also writes Markdown, with a slightly stricter heading discipline and long answers that use # headings at the top level. Copy an answer with the copy control below the message and paste it here. For content Claude placed in an Artifact, open the Artifact and copy from there; the Markdown comes through intact. Display math uses $$. The extension adds the same Word button on claude.ai.
Gemini to Word
Gemini has a built-in "Export to Docs" action, but it targets Google Docs only and drops code-block formatting and table borders. Its Copy control gives you Markdown, which converts here into a .docx with real Heading styles that Google Docs also opens correctly. Gemini favours bold lead-ins on list items and tables with header rows, both handled. The extension adds the Word button on gemini.google.com.
DeepSeek to Word
DeepSeek answers frequently contain heavy math in \[ … \] blocks and a "thinking" section before the answer. By default only the final answer is converted; enable the reasoning block in the extension's options if you want it included as a shaded section. Math is handled the same way as ChatGPT's, with Pro producing editable Word equations. The extension adds the Word button on chat.deepseek.com.