How to collect, analyze, report and act on customer insigths

Every product (marketing) decision is only as good as the customer truth behind it. We believe getting that truth right shouldn't be daunting... and it shouldn't take months. That's why we built a customer research process that's practical, simple, and scalable enough to actually run in your role.

Welcome to part one of our customer research methodology: a tested process built by product marketers, for the product and product marketing managers doing the work. The full method moves through four stages: collect, analyze, report, and act. Bubble gives you both halves of each: a method to follow and a product that does the analysis, so the conversations you're already having become fundamentals you can trust and act on.

This article is a sneak peek of the Collect chapter. Read it, see how it fits the way you work, and decide for yourself whether the full method is worth going deeper on (free pdf below).

Introduction

If you've got "product" in your job title, like product marketer, product manager, or something close, this course is for you.

Here's what it isn't: a course for professional researchers, and it won't try to turn you into one. No research degree required, and you will not be asked to read a 500-page academic textbook on qualitative data analysis. The goal is the opposite — to give you just enough foundation to do genuinely good research without training for years to do it.

Why does that matter for your role? Because you already do research, whether or not anyone calls it that. Listening to sales calls, running customer interviews, building the roadmap or messaging from "what customers told us": that's all qualitative data analysis.

Right now you might do all of that by hand; scrubbing through call recordings, copying quotes into a doc, building your messaging from memory and gut. Or you might already lean on AI to speed parts of it up. Either way, AI is about to become a much bigger part of how this work gets done. So the real question isn't whether you'll use it, but whether you'll use it well.

With AI, you can already be fast. But you don't want to be fast and wrong. That's why this course will help you to be fast (with the help of AI) and accurate (with the human in the loop). Here, accurate means built on what customers actually said: your fundamentals and decisions come from real customer conversations, and a validation log flags every assumption so nothing runs on a hunch. Fast and accurate, not one at the cost of the other.

And it matters more, not less, now that AI tools do the heavy lifting. AI can help you surface themes, and tie insights back to a quote in the transcript. But to trust what it hands you — to know when it's right, when to push back, and how to defend a finding to a skeptical stakeholder — you need a feel for what it's doing underneath.

That's what this methodology is for. It will teach you the minimum theoretical background you need to pass judgement on what is happening in your research, give you a clear process from start to end on how to run it, and give you tips and tricks along the way on how AI can help you be fast, but stay accurate.

We hope you enjoy it.

Anne, Erinc & Steven
Co-founders of Bubble

3. Collect Your Data

3.1 Collect Your Data

Your plan is set, so now you gather the raw material your analysis will run on. But before you book a single new interview, you take stock of what you already know and pull the customer conversations you already have: recorded sales, success and support calls in tools like Gong, or from past research. Only then do you run new interviews, and only to fill what's still missing.

That order matters because interviews are the expensive part of your research, because it takes time to find people to interview, schedule the meeting and prepare the interview properly. And let’s be honest, when you haven’t done it a lot before, it can also be pretty scary.

So this chapter shows you how to map the gap between what you know and what you need, pull existing transcripts well enough that your AI tool can actually use them, build an interview guide that doesn't bias people, recruit and run the conversations, and use AI at this stage without letting it distort what you hear.

3.2 Step 1: Start With What You Already Have

Good research starts by collecting what you already know — to see what you can build on and what genuinely needs new work. Before collecting anything new, map three things: your current situation, your desired situation, and the gap between them. The gap — the evidence you don't yet have — is what you actually go and collect. Everything you already hold, you don't.

Maybe there's an old ICP profile or a persona deck; maybe it's outdated; maybe there's nothing. That inventory is the point. Then you apply three questions when analyzing the inventory:

1. What do we already have that could support this research?
2. What do we think we know but haven't validated against real customer data?
3. What is unclear, opinion-based, or missing

Worked through for a persona-update workflow, the picture looks like this:

The output of this step is a short research action plan: the prioritized list of gaps to fill. That list, not a vague "we need better personas," is what drives the rest of the Collect phase.

3.3 Step 2: Collect Existing Customer Conversations

Most teams already sit on months of customer conversations: recorded sales and success calls in Gong, Fathom, Meet, Chorus, or Fireflies. Start here. It's faster than new interviews, and it often closes part of the gap outright — sometimes a gap you'd otherwise have spent two weeks interviewing to fill.

3.3.1 What to Pull

Don't export the entire call library. That buries the signal and slows you down. Instead, work step by step to select the right transcripts:

1. First you filter on the segment and persona you defined in your planning template.
2. You triage existing conversations, using an AI tool like Claude to identify good transcripts
3. You define the gap by already analyzing these transcripts with Bubble
4. You define the gap by identifying elements with low confidence levels in the report.

3.3.2 How to Pull It

If there’s no integration between your AI tool and these tools, you're exporting by hand — and how you export decides whether the data is usable:

• Full transcripts, not clips. Export the whole conversation, not a highlight, snippet, or the tool's AI summary. Bubble codes what was actually said; a summary has already thrown away the verbatim wording your analysis and messaging depend on.
• One file per conversation, with speaker labels and the date preserved, so every quote can be traced back to a source.
• A format most AI tools can easily read (plain text, .docx, or .vtt). Name files so you can tell segment, source, and date apart at a glance.

3.3.3 What to Pay Attention To

Transcript quality caps insight quality. Whatever you load is exactly what Bubble reasons over: garbage in, garbage out. A run built on mislabeled, partial, or off-segment transcripts produces confident (typical LLMs!), but wrong insights. Check each one against this before you load it:

• Speaker attribution is correct — who said what. If the rep and the customer are merged into one voice, the analysis can't tell a customer need from a sales pitch.
• Names, product terms, and numbers are right. Auto-transcription mangles exactly the terms your findings hinge on.
• It's the whole conversation — not a trimmed highlight reel that drops the context around a quote.
• It's the right customer. Off-segment conversations don't just add noise; they pull themes toward the wrong people.
• You have enough per group. Count usable existing transcripts against the per-segment targets from 2.5.1 — that tells you how much Step 3 still has to cover.

3.4 Step 3 — Run Additional Interviews to Fill the Gap

Existing conversations rarely cover the whole gap and they weren't recorded with your research question in mind. Whatever's still unanswered after Step 2 is your interview scope, and nothing more. Interview for the gap, not for everything: it's the difference between three sharp conversations and fifteen unfocused ones.

3.4.1 Prepare Your Research Guide

Don’t build the guide straight from your research question, but use the framework and our sample question sheet (see resources in this Chapter). As Chapter 2 laid out, your research question carries a framework (SPICED, the 5 Rings of Buying Insight, an Opportunity Solution Tree, and so on). This is where that framework earns the first half of its keep: its dimensions become your interview topics, and you write questions under each. Research question → framework → interview questions — the same chain from 2.3.1, now put to work. Building the guide this way is also what keeps your questions focussed and saves you lots of time.

Additional guidance here comes from Rob Fitzpatrick's The Mom Test. His warning: the truth is fragile, and a blunt question shatters it before you can reach it. Most of us bias our interviews without noticing, and a bad conversation hands us a false negative, or worse, a false positive that leads us to over-invest time, money, and team. Fitzpatrick reduces good interviewing to three rules:

Table 1. The Mom Test in practice

3.4.2 Spot Bad Data

Even with good questions, three kinds of misleading answers creep in. Catch them in the moment:

• Compliments. "That's a great idea" costs the participant nothing and tells you nothing. Deflect, and dig for the why behind it.
• Fluff. Generic, future-tense answers — "I usually," "I'd probably," "I would." Anchor them back to a specific, real event.
• Feature requests. Don't take them at face value. Ask what problem the feature would solve and what they do about it today.

3.4.3 Shape of a Good Conversation

Pre-plan the three things you most need to learn from each type of person; that makes good follow-ups far easier. Then structure the conversation loosely:

• Warm up. Ease in with easy, real questions about their role and day.
• Past behavior. Walk through the last time the problem actually happened, step by step.
• Dig into how. Get at how they experience and make sense of the problem — how they feel about it and what they've already tried. In Chapter 1's terms, this is the "how" question (motivation and meaning), not a covariation "why."
• Close with a next step. Ask "Who else should I talk to?" and "Is there anything I should have asked?" — the best follow-ups and warm intros come from these.

Keep it short: 30–45 minutes gets you most of what you need, though people happily talk longer once they're going.

3.4.4 Recruit Your Participants

Recruit against the customer segment you defined in Chapter 2 — the tighter the screen, the more the findings apply to the decision.

• Source from what you already have. CRM, product analytics, support and CS tools, win/loss notes, and any voice-of-customer program surface relevant people fast.
• Screen to the segment. Filter on the behavioral or firmographic criteria that define your group — usage, tenure, plan, role, recent activity — not just "any customer."
• Reach out honestly. Tell people you're doing research and would value their perspective; offer to share what you learn. Warm intros from Sales, CS, or a first participant convert far better than cold asks.
• Offer an incentive when needed. A clear thank-you (gift card, credit, early access) lifts response rates, especially for busy or senior participants.
• Build a research pool over time. Build an email sequence towards contacts in your CRM to ask people to subscribe to future research. Ask participants if they're open to future research after you spoke to them and add the willing ones to a list — over time you grow a research pool, so recruiting interviewees gets easier each round.

3.4.5 Run Your Interviews

Set up. Get explicit consent to record — and again if AI tools will process the audio or transcript — and explain how it'll be used. If two of you interview, one moderates and stays present while the other takes notes; solo is fine, just record so you're not writing and listening at once. Budget 30–45 minutes.

In the conversation. Park your idea; if you slip into pitch mode, name it and steer back to them. Follow their energy — when a topic lights someone up, stay there and ask why. Embrace silence; a pause is an invitation for them to say the thing they were deciding whether to share. Capture their words — verbatim quotes and emotional signals, not just your paraphrase, since exact wording is what your semantic coding and messaging lean on later.

Right after. Debrief within the hour, while it's fresh. Pull the top quotes, surprises, and anything that challenges your assumptions, and log them against the research question. Notes you never revisit are wasted notes.

3.5 How to Use AI at This Stage

AI can take real work off your plate during collection — in both steps — as long as it stays in a support role and a human owns every judgment call. This mirrors the human-in-the-loop principle from Chapter 1: AI surfaces, you decide.

• Triaging existing conversations. Use AI to skim and shortlist which Step 1 transcripts are relevant to a gap — but you still verify and decide what goes in.
• Drafting additional questions for the guide. Draft topics and candidate questions from your research question, then edit hard — strip anything leading or hypothetical before it reaches a participant.
• Recruiting. Draft screener criteria and outreach messages, then personalize them.
• Transcription. AI transcription is a big time-saver and produces the clean, quote-level text your analysis depends on — but verify it, especially names, product terms, and numbers.

Resources

Build your guide with these:

• Good, bad, and ugly interview questions — a reference for clean interview questions
• Sample question sheet — a starting set to adapt to your research question.