Need VOC Themes in ChatGPT? Paste Reviews and Cite

You dump a page of G2 and Capterra text into ChatGPT and ask for themes. You get a tidy list. Some quotes are paraphrases, some are not in the dump, and a few sound like the product you already know.

ChatGPT will theme a review dump if you put the reviews in the prompt, require verbatim quotes with source IDs, and check every quote against the source before it goes on a slide. That path is valid. The work is the constraint and the check.

If the job is review page URLs or a corpus and you want clustered themes without prompt craft: how to turn G2 and Capterra reviews into VOC themes

How to get cited VOC themes from a review dump in ChatGPT

1. Put the reviews in the prompt with source IDs

Paste the reviews (or put them in a ChatGPT Project / Custom GPT instructions) with stable source IDs: G2-14, Capterra-3, AppStore-8. Keep the original URL next to the ID when you have it.

Do not ask ChatGPT to "remember the market" or to fill gaps from training. If a review is not in this prompt, it does not exist for this run.

One review per block is easier to cite later than a wall of unlabelled paragraphs.

2. Require clustered themes with verbatim quotes

Tell ChatGPT to cluster pain themes from the supplied reviews only. Every theme must include verbatim quotes and the source IDs those quotes came from.

Forbid using general product knowledge. Forbid inventing reviews. Forbid paraphrasing a quote into something punchier. If the model cannot find a verbatim line, it should drop the quote, not fix the wording.

A useful instruction is boring on purpose: cluster from this dump, cite IDs, do not add reviews you were not given.

3. Ask for a fixed output schema

Ask for the same fields on every theme so you can scan the result:

• Theme name
• Who it hurts (the reviewer role or situation named in the quotes, not a persona you invented)
• 2 to 4 cited quotes, each with its source ID
• What to verify (the claim you would check against the original page before you reuse it)

You can add a severity guess if you want one. Treat it as a label on this dump, not a market score.

4. Chunk long corpora, then merge by ID

Long dumps get truncated. Split the corpus into chunks that fit, run the same instruction on each chunk, and keep the source IDs.

Then ask ChatGPT to merge themes across those chunk outputs by ID. Do not paste the whole dump again in one shot and hope the middle reviews survived. Merging from themed lists with IDs is more reliable than one giant paste.

If two chunk themes share quotes, combine them. If they only share a vibe, keep them split until you read the quotes.

5. Verify every quote against the source

Before a slide or a PRD, open the original review and find the line. ChatGPT will invent or paraphrase quotes unless you constrain it and then check.

If you cannot find the sentence, drop the quote. If the theme only stood up because of that quote, drop or rename the theme. Do not publish a cleaned-up line that the reviewer did not write.

6. Prefer pasted text over browsing G2 URLs

ChatGPT browsing and fetch are unreliable for G2 and Capterra pages: login walls, JavaScript, and blocked scrapes. Prefer pasted text, or a tool that is built to ingest the corpus you already copied.

If you paste a URL and the model "summarizes the page," assume it guessed until you have the text in front of you.

When you would rather skip the prompt craft

When the job is review page URLs or a review corpus, and you want clustered pain themes with citations without writing the instruction above, Review VOC Pack does that ingest. Paste the corpus (the reliable path for G2, Capterra, and Trustpilot; those aggregator pages are not fetched). You can also paste review page URLs we can fetch when robots.txt allows it.

Guest runs are free (5/day). Signed-in orgs spend 2 credits per request. You still verify quotes.

ChatGPT stays a valid path if you paste the text, lock citations to source IDs, and check every quote. Use whichever loop you will actually finish.

What this is not

This method is a first pass on reviews you already collected. It is not:

• A ticket, NPS, or survey CSV theming run (Customer Feedback Analyzer is not a live AgentLeverage tool)
• A replacement for interviews, sales calls, or win/loss notes
• A scrape of G2 or Capterra
• A ship decision

You still decide what to do with a theme after the quotes check out.

Frequently asked questions

Can ChatGPT theme G2 reviews if I only give it the URL?

Unreliable. Prefer pasted text with source IDs. Login walls and JavaScript get in the way of browsing those pages.

Why do I need source IDs?

So every quote maps back to one review you can open. Without IDs, "several reviewers said…" is hard to check and easy to invent.

Does ChatGPT invent review quotes?

It can invent or paraphrase them. Constrain the prompt, then verify every quote against the source before you publish.

When should I use Review VOC Pack instead?

When you already have review page URLs or a corpus and you want clustered pain themes with cited quotes without prompt craft. Review VOC Pack guest runs are free (5/day). You still verify quotes.

Does Review VOC Pack replace ChatGPT?

No. ChatGPT is a valid path if you paste text and constrain citations. VOC Pack is the ingest path when you do not want to write that instruction.

Cluster review themes →

---

Have a G2 or Capterra dump and want clustered themes with citations? Try Review VOC Pack. Guest runs are free (5/day). You still verify quotes. ChatGPT works if you paste the text and check every line.