How to Analyze Customer Interview Notes and Research in Notion

Notion is good at holding research and bad at counting it, so fix the structure before you add AI. Create one database, such as Customer Conversations, with a row for every interview, sales call recap, and advisory board session.

Backfill only the last 90 days; older notes describe a product you've since changed. As a rule of thumb, moving 30 pages takes about an hour.

Set this as the database's default template:

# [Account], [YYYY-MM-DD], [Type]

## Context
Role, team size, plan, how long they've been a customer

## What They Said (Verbatim)
Participant: "We export the report every Friday and rebuild it in Sheets."
Interviewer: "What do you change?"
Participant: "We add the renewal date. Without it, finance ignores the numbers."

## What They Do Today
- The workaround, and how long it takes

## Our Interpretation
- What we think this means (hypotheses, not evidence)

• Paste quotes word for word from the transcript. Mark anything written from memory as a paraphrase.
• Write one Speaker: text line per turn. Tools such as Deck use the labels to tell the participant from the interviewer.
• Give workarounds their own section. A workaround proves the problem already costs the customer time; a feature request is only their guess at a fix.

Recording with Notion AI Meeting Notes (Business and Enterprise)? Use the desktop app. In the browser, Notion transcribes only your microphone, so you lose the customer's side of a web call (Notion Help). Get every participant's consent first.

Tag from a fixed list of 8 to 15 problems, named after the customer's pain. "Can't see which seats are active" is a problem; "Admin dashboard" is a solution someone already picked. To build the list, see how to build a customer feedback taxonomy.

Notion AI can take the first pass (Notion Help):

• Basic Autofill fills tags and summaries from the page's own content. It's included on Business and Enterprise, uses no credits, and runs manually or when a page is created or edited.
• Custom Agent Autofill can search the workspace and run on a schedule, but each task uses Notion credits.

Give Basic Autofill instructions like these:

Read only "What They Said" and "What They Do Today".
Pick up to 3 problems from the list below that the participant describes in their own words.
Ignore "Our Interpretation". If no problem fits, leave this property empty.

- Can't see which seats are active: admins can't tell who uses the product
- Report needs rebuilding: users export data and fix it in another tool

For the first month, spot-check one in five tagged pages. If you'd change more than one tag in five, tighten the definitions before trusting the counts.

One interview is an anecdote, and Basic Autofill only ever sees one. To read across thirty, use a grouped view or an AI assistant; the third route, a feedback tool, is under Tools That Help.

In Notion. Filter a board view to the last 90 days and group it by Problems; a page tagged with three problems appears in all three columns. Change the number beside each column heading to Count unique values of Account, so each column shows accounts, not pages (Notion Help).

With Claude or ChatGPT. Notion's MCP server lets assistants such as Claude, ChatGPT, and Gemini search your workspace and work with databases (Notion Developers). Notion Agent does similar work inside Notion (Notion Help). Use a prompt that forces evidence:

In the Customer Conversations database, read every page with Type = Interview and a Date in the last 90 days.

1. First, list every page you read, by title. Say so if you could not open one.
2. Group the problems participants describe into 4 to 8 themes, named after the customer's problem.
3. For each theme, list the accounts and one verbatim quote per account from "What They Said", with the page link.
4. Never use "Our Interpretation" as evidence. Do not count, rank, or invent quotes.

The first instruction matters most: if the list is shorter than your database, the assistant skipped pages. Work in batches of 10 to 15 interviews, then merge themes. For other setups, see how to connect customer feedback to Claude or ChatGPT.

A workable bar: a problem is a finding once three unrelated accounts in your target market describe it in their own words. Below that, it's a lead worth more calls. It is the same bar the weekly synthesis session uses. Run two checks first:

• Where the evidence sits. A theme that rests mostly on "Our Interpretation" is your hypothesis, not the customer's problem.
• Three quotes in context. Open three quotes in their source pages. AI summaries flatten "annoying" and "blocking" into the same word.

Then mark the pages Synthesized and link your decision to them. Don't let a "Q3 research findings" page replace the interviews: teams cite it for a year, and nobody can check it.