You’re in the roadmap review. Someone asks the only question that matters: did we validate this?
“Yeah,” someone says. “We did five customer interviews.”
Heads nod. The build gets greenlit.
Nobody in the room can tell you what interview six would have shown them.
Ask five different product people how many user interviews are enough and you’ll get five different numbers back, each delivered like it’s gospel. Five. Seven. Ten to fifteen. Fifty, if you ask the person who thinks qualitative research needs a real sample size.
Here’s the part that should bother you more than it does:
They can’t all be right, and none of them can tell you why their number is the right one for your product.
That’s because the number was never the point.
Saturation is a signal you detect in what people are telling you. It is not a headcount you pick before you’ve talked to anyone.
In 2026, an AI agent can run fifty of those conversations while you sleep. The real question has quietly flipped. It used to be how few can I get away with. Now it’s when do I stop learning.
How many user interviews are actually enough?
For a single, well-defined customer segment, most product teams reach saturation, the point where new conversations stop surfacing new patterns, somewhere between 8 and 15 conversations. A 2022 systematic review of qualitative studies found data saturation typically lands between 9 and 17 interviews. The number moves. The signal doesn’t.
That range isn’t a guess. A 2022 review in Social Science & Medicine pulled together 23 empirical studies that tracked when new themes actually stopped showing up in the data, rather than just recommending a number from experience.
Across studies with a reasonably homogeneous group of participants, code saturation landed in that same 9-to-17 range. Broader questions and multiple segments push it higher.
Here’s the deal: that’s a landing zone, not a rule. Your actual number depends on two things, and neither of them is how many interviews someone else ran.
Why the “test with five” rule doesn’t apply here
You’ve heard it: test with five people and you’ll catch most of the problems. It’s true.
It’s also being misapplied to your situation, and that’s the whole reason the search results for this question are such a mess.
The number-five rule comes from usability testing: watching someone try to use an interface and counting how many bugs and confusing moments you catch.
Nielsen Norman Group’s research on interview sample size is blunt about the difference: finding usability problems in a screen is a narrow, bounded task. Understanding a person’s experience, motivation, and workflow is not.
Five interviews are often not enough, NN/g’s own research concludes, because an interview study has more variability to capture than a usability test does. You’re not counting broken buttons. You’re mapping how someone actually thinks about a problem, and that takes more than a handful of conversations to see clearly.
The “test with five” rule was built for usability testing: catching interface bugs. Not for interviews: understanding why someone behaves a certain way. Different job, different math.
The saturation framework: how you know you’re actually done
The best-known study on this, run by Guest, Bunce, and Johnson in 2006, analyzed 60 interviews on a narrow, well-defined health topic. Code saturation hit by interview 12. Most of what they’d ever find had already shown up by interview six.
That’s the finding everyone quotes. What most people cut is the context: a narrow question, a homogeneous group of participants, and researchers who already understood the domain going in. Change any of those and the number moves.
Two things decide your number, not the calendar and not a rule of thumb:
- How narrow is your question? “Why do trial signups churn before day 14 specifically because of onboarding friction” saturates fast. “How do small businesses think about software in general” does not.
- How similar are the people you’re talking to? Interview five VPs of Sales at 50-person SaaS companies and you’ll see patterns fast. Interview five VPs of Sales at companies ranging from 10 to 10,000 people and you won’t.
| Situation | Rough range |
|---|---|
| One narrow question, one tight segment | 5 to 8 |
| Standard product discovery, one segment | 8 to 15 |
| Broad, exploratory question | 15 to 25 |
| Multiple distinct segments or personas | multiply the per-segment number by the segment count |
After every conversation, ask one question: did I learn anything I didn’t already know? When three in a row come back “no,” you’ve hit saturation for that segment. Stop counting people. Start counting new information.
Who you talk to matters more than how many
A weak signal from ten conversations is worse than a strong one from six, and the fastest way to get a weak signal is talking to the wrong people at the right cadence.
Two failure modes show up constantly. Teams interview whoever’s easiest to reach instead of their actual ICP, which pads the count without ever hitting real saturation. Or they ask people to predict the future (“would you use a feature like this?”) instead of describing their past, which produces confident-sounding answers that mean nothing.
If you haven’t nailed down exactly who you should be talking to yet, our guide to defining your ICP without sales data walks through how to find them before you burn interviews on the wrong crowd.
Once you’ve got the right people on the call, asking about their past instead of their opinion of your idea is what turns a pleasant chat into a data point you can actually use.
- Repetition, not novelty. The eighth person describes the same workaround the third person did.
- You can predict the answer. You start finishing sentences in your head.
- The follow-ups dry up. You run out of genuine curiosity about what comes next.
The real shift in 2026: the ceiling was never the sample size
Here’s what the saturation research doesn’t tell you: the reason most teams stop at five or six interviews was never a methodology decision. It was a calendar problem.
Product researcher Teresa Torres has recommended for years that product teams interview at least one customer every week, as a standing habit.
She’s also candid about why most teams don’t: the habit matters more than the headcount, but most teams can’t sustain even one interview a week once recruiting and scheduling friction sets in.
That friction was the actual constraint the whole time. Not the theory. The calendar.
When the scheduling ceiling disappears, the question changes.
An AI agent that runs a real conversation, not a fixed script, doesn’t care whether it’s talking to 8 people or 80. It doesn’t need a calendar invite, and it doesn’t get tired by interview eleven.
Our own breakdown of when AI-moderated interviews actually work goes deeper on spotting a real one from a survey wearing an interview’s clothes. But the short version: once the marginal conversation is nearly free, “how few can I get away with” stops being the right question.
It becomes “when do I stop learning something new,” which was always the right question. It just wasn’t affordable to ask it.
This is the bet behind hollie. She has real conversations with your customers, not fixed-script ones, and brings back what she heard, ranked by how often it came up.
If you want to see what running past your usual five looks like, holito’s free plan is a 14-day trial with full access, no credit card required.
Frequently asked questions
Is 5 user interviews ever enough?
Yes, for a genuinely narrow question with a tightly defined, homogeneous group. Testing whether one specific onboarding step confuses a specific persona can saturate in 5 to 6 conversations. Broader discovery, “why do people churn,” “what should we build next”, almost never does.
What if I have more than one customer segment?
Treat each segment as its own saturation count. If you’re studying three distinct personas, plan for roughly 8 to 15 conversations per segment, not 8 to 15 total. A sample that’s diverse across unrelated segments will never feel saturated, because you’re really running three separate studies at once.
How many interviews do I need to validate a brand new idea versus a feature for an existing product?
A brand-new idea usually needs a wider net first: 15 to 20 conversations across a broader set of prospects, because you’re still discovering who has the problem at all.
A feature for an existing product is narrower. You already know the audience, so 8 to 12 conversations with active customers usually gets you a clear signal.
Does an AI-moderated conversation count the same as a live interview toward saturation?
The saturation research was built on human-run interviews, but the underlying logic (new conversations stop revealing new information) doesn’t care who’s asking the questions. What matters is whether the follow-ups actually dig into what each person said, not a fixed script. A real, adaptive conversation counts. A survey with a chat interface bolted on doesn’t.
Stop picking a number before you’ve heard a single answer.
Run the conversations. Watch for the repeats. Stop when the room stops surprising you.
The Bottom Line
There’s no universal number of user interviews that’s “enough.” There’s a signal: new conversations stop surfacing new information. For most single-segment product research, that lands between 8 and 15 conversations, wider for broad questions, narrower for tight ones. Pick your number based on what you’re hearing, not what you read on a forum thread.
