How to export and analyze AI conversations in Excel

When a spreadsheet is the right format

Reading a conversation from top to bottom is useful when you care about its narrative. Analysis asks different questions: How many turns did it take to resolve the task? Which prompts produced corrections? Where did the assistant make unsupported claims? Which themes recur across several sessions? A row-based export makes those questions easier to investigate.

Excel is especially helpful for qualitative coding, prompt evaluation, editorial review, support analysis, and building a catalog of valuable conversations. It is not always the best reading copy. Keep Markdown alongside the workbook when the original response includes meaningful headings, lists, tables, links, or code.

Three export formats, three jobs

Prepare the dataset

1. Export the conversation. Create a public share link in ChatGPT, Claude, or Gemini, paste it into the ThreadSift exporter, and choose Excel.
2. Keep the source file unchanged. Duplicate the worksheet before cleaning or coding so you retain an untouched reference.
3. Add a conversation ID. When combining files, give every chat a stable identifier so filters and summaries do not mix sessions accidentally.
4. Normalize your labels. Use consistent values for speakers, topics, status, and reviewer. A dropdown list reduces spelling variants.
5. Separate observations from source text. Do not overwrite message content. Put codes, ratings, and reviewer notes in new columns.

ThreadSift structures the export around message order, speaker, and content. Your analytical columns should reflect the question you are trying to answer, not every attribute you might possibly collect.

Useful columns for analysis

• Conversation ID: connects every row to its source session.
• Message number: preserves sequence and supports turn-by-turn comparison.
• Speaker: separates user prompts from assistant responses.
• Topic code: groups messages into a small, documented set of themes.
• Outcome: records whether the turn advanced, corrected, blocked, or completed the task.
• Verification status: distinguishes unchecked claims from claims reviewed against a reliable source.
• Reviewer note: captures a concise observation without changing the transcript.

If you are studying many conversations, also record the provider, model name when known, collection date, use case, and inclusion criteria. Avoid collecting personal or sensitive fields simply because the spreadsheet makes it easy.

Analyses you can perform

Filter by speaker to review prompts independently from responses. This reveals whether instructions were consistent, whether crucial context appeared late, and how often users needed to restate a request.

Code themes and outcomes with a controlled vocabulary. A pivot table can then count topics, correction types, or resolution status across conversations. Keep a short codebook defining each label so two reviewers apply it consistently.

Measure message length carefully. Character or word counts can indicate verbosity, but length is not quality. Interpret it beside task type and outcome. A concise answer may be excellent for a factual request and inadequate for a complex analysis.

Compare turns within a session by preserving message order. You can inspect what changed after a clarification, where an error first appeared, or how a prompt constraint influenced the next response.

Sample for manual review. Spreadsheet summaries can point to unusual or high-impact cases, but they cannot determine truth on their own. Read the surrounding conversation and verify important claims against authoritative sources.

Avoid common spreadsheet mistakes

Do not split a single message into several rows unless your method requires passage-level coding and you retain a link to the original row. Do not sort only one column; sort the complete table so content stays attached to its identifiers. Be cautious with formulas that interpret cells beginning with special characters, and treat imported links or text as untrusted data.

Document exclusions and transformations. If you remove greetings, redact identifiers, or combine short turns, record the rule and apply it consistently. For team review, protect source columns and put annotations in designated fields. Agreement checks on a small sample can reveal whether category definitions are too vague.

Privacy and platform limits

ThreadSift requires a public share page. Review the full chat before sharing it, particularly if it contains personal data, internal information, unpublished research, or credentials. Private and account-gated chats are unsupported. ThreadSift does not use an application database to retain submitted links or exported conversation history, and no sign-up is required.

ChatGPT, Claude, and Gemini can change how their public pages work, so extraction depends on the provider page remaining available and readable. Once exported, manage the workbook under your own access, retention, and backup policy. See the AI chat export guide for a quick overview of the process.

Frequently asked questions