Support conversations are the largest body of unsolicited user research most companies own, and the one they use least. It is continuous, unprompted, high-volume, and written by people at the exact moment something went wrong — which is the moment research is most honest.
The obstacle has always been volume: nobody can read forty thousand conversations. Now that a machine has already read and categorised them, the obstacle is gone, and what remains is the habit of looking.
Four signals worth extracting
1. Questions that should not need asking
Cluster your top intents and ask, for each: why does the interface not make this obvious? "Where do I find my API key?" asked two hundred times a month is not a documentation gap. It is a navigation bug with a documentation workaround.
The strongest version of this signal is a question asked inside the screen that answers it. That is a design brief.
2. The vocabulary gap
Compare the words customers use with the words in your interface. If everyone writes "workspace" and your product says "organisation", you are paying a small tax on every interaction — in support, in onboarding, in search, in every doc you write.
3. Features requested in the phrasing of the request
"Can I export this to CSV?" is a feature request wearing the clothes of a question. Bots answer these politely and truthfully — "that's not currently available" — and the demand disappears into the archive. Tag them. Count them. Sort by count. That list is better prioritisation input than most roadmap workshops, because it is weighted by people who wanted the thing badly enough to ask.
4. Where conversations end badly
Find the topics with the highest escalation rate, the longest conversations, or the most rephrasings. Long conversations are not engagement; they are difficulty. That ranking is a map of your product's roughest edges, ordered by how many people hit them.
A monthly routine that takes an hour
- 1Export last month's conversations grouped by intent.
- 2Take the top 20 intents by volume. For each, write one sentence: what would make this question stop being asked?
- 3Sort those sentences into docs fix, UI fix, feature gap.
- 4Send the docs fixes to whoever owns the help centre — they are usually done that week.
- 5Bring the UI fixes and feature gaps to planning with the volume attached. "Four hundred people asked this last month" ends a prioritisation debate faster than any argument.
A support queue that stays the same size while your user base grows is a product getting clearer. That is a better health metric than most of the ones on the dashboard.
Handle it responsibly
Conversation logs contain personal data and often account details customers pasted in without thinking. Before this becomes a routine anyone can run: redact identifiers on ingest, set a retention window and enforce it, keep analysis on aggregates rather than individual transcripts, and make sure your privacy policy actually describes what you are doing. The insight is worth having. It is not worth having carelessly.
Then the loop closes on itself. The bot answers questions, the questions reveal the product's rough edges, the edges get sanded down, and the questions stop being asked. The best possible outcome for a support assistant is a support queue that quietly shrinks.
- Research
- Product
- Analytics
