When to stop user interviews: a practical saturation rule

You have talked to seven trial users about second-week drop-off. By the fifth conversation, a pattern is clear: a silent CSV-import failure makes people think the product cannot handle their data. Six and seven add variations, but no new reason to leave.

Can you stop?

For a focused interview study using inductive thematic analysis, a defensible answer is: stop when a pre-set batch of interviews adds little or no new information relevant to the decision, after checking that your sample, probing, and analysis are strong enough to make that signal credible.

That is more useful than a universal target such as “always interview five” or “always interview 20.” It turns “this feels repetitive” into a rule another person can inspect.

This article explains the rule, when it applies, and the three ways it can tell you to stop too early. If you need a number for planning before recruitment begins, start with how many users you should interview.

What saturation means—and what it does not

Researchers often call the stopping point saturation. Nielsen Norman Group defines it as the point at which themes are developed enough that more interviews are unlikely to add insights that would change them.

Saturation is not certainty, statistical representativeness, or proof that the next person will say nothing new. It is a transparent judgement about whether more data is likely to alter the answer to a particular research question.

It is also not appropriate for every qualitative approach. Braun and Clarke caution that themes are created through interpretation rather than simply discovered in transcripts, so a claim of “no new themes” partly reflects the researcher’s analytic choices. Saturation makes most sense here as a practical stopping aid for a narrow, applied study—not as a universal quality stamp.

Define the scope before counting:

For recent self-serve trials at companies with 10–50 employees, what prevents a successful first data import during the first two weeks?

“What do users think of onboarding?” is too broad. A stopping rule cannot settle a question that mixes audiences, stages, and decisions.

The base-run-threshold rule

Guest, Namey, and Chen proposed a simple way to assess thematic saturation during data collection. Their method has three settings:

• Base size: the first set of interviews used to establish the denominator of distinct themes.
• Run length: the number of consecutive interviews checked for themes not found in the base.
• New-information threshold: how much new information you will accept before stopping.

The authors tested base sizes of four, five, and six; run lengths of two and three; and thresholds of no new information or at most 5%. They found the base size had little effect in their three datasets and suggested four interviews as an efficient default, followed by a run of two.

Their calculation is:

New-information rate = themes new to the base in the current run ÷ distinct themes in the base

The paper’s hypothetical example begins with 37 themes in four interviews. Interviews five and six add seven themes, or 19%. The next overlapping run, interviews six and seven, contributes four themes, or 11%. Interviews seven and eight contribute one, or about 3%. With a threshold of at most 5%, that last run triggers the stop. The overlapping run is important: the method moves forward one interview at a time rather than checking only separate pairs.

The threshold is a chosen decision rule, not a law of nature. The paper explicitly says there is no guarantee that saturation has been reached when a threshold is met. A longer run or a zero-new-information threshold is more conservative.

A blank tally sheet

Use one row per interview and update it while fieldwork is running:

Write a short code definition for every theme. “CSV problem” is too loose; “import returns no visible error after unsupported date format” is specific enough for another reviewer to apply consistently.

Nielsen Norman Group recommends starting with a small sample and analysing interviews as you go. Without ongoing analysis, you cannot see the stopping signal until after you have already completed every booked conversation.

Do not generalise the six-to-twelve result

In the Guest study’s two relatively homogeneous datasets, six or seven interviews captured roughly 80% of all themes found in the complete datasets; at the 95th percentile, eleven or twelve were needed to meet the 5% threshold. The more varied third dataset needed more interviews and still captured a smaller share of all themes at that threshold.

Those results are useful calibration, not a SaaS benchmark. The method was validated on three previously coded interview datasets for narrow applied questions. It was designed for open-ended interviews with inductive probing and inductive thematic analysis. It has not been validated as a universal stopping rule for product discovery, usability tests, diary studies, or mixed-method research.

Francis and colleagues used a related approach in theory-based interview studies: specify an initial analysis sample and then a number of additional interviews that must add no new ideas. Their worked recommendation used an initial ten plus a stopping criterion of three. The lesson is not that 13 is always enough. It is that the rule should be declared before the current results tempt you to move the goalposts.

“Heard it all” is different from “understood it”

A theme can stop being new while its meaning is still changing.

Hennink, Kaiser, and Marconi analysed 25 in-depth interviews and reported that the range of issues stabilised at nine interviews, while richer understanding required 16–24. They called these code saturation and meaning saturation.

That distinction maps cleanly to product decisions:

• Code saturation: Have we identified the main failure modes in this segment?
• Meaning saturation: Do we understand expectations, context, workarounds, consequences, and differences within each failure mode well enough to design a response?

If the decision is whether a reproducible import defect exists, you may not need saturation at all: reproduce it, assess impact, and fix it. If the decision is how to redesign an import workflow for bookkeepers and operations managers, merely listing “CSV failure” is not enough.

Add a second stop signal to the tally sheet: did this interview change the meaning, boundaries, or consequence of an existing theme? Continue if that column is still producing decision-relevant information, even when the new-theme rate is low.

For analysis techniques that preserve differences and evidence trails, see how to synthesise user interviews into themes and the affinity-mapping guide.

Three ways false saturation happens

1. Your interviews are too shallow

Repetition can mean you have covered the subject. It can also mean you keep asking the same top-layer question.

Weller and colleagues analysed 1,147 interviews across 28 datasets. Ten interviews with exhaustive elicitation captured 95% of the most salient items, while ten interviews limited to three responses per person captured 53%. Their conclusion was direct: saturation can be an illusion caused by insufficient probing.

Before accepting a stop signal, audit the conversations:

• Did you ask for a recent, specific event?
• Did you follow the sequence of what happened?
• Did you ask what the participant tried next?
• Did you probe consequences and workarounds?
• Did participants have room to introduce something outside your guide?

If the answers are thin, recruit depth before volume. The guides on asking why without asking “why” and avoiding leading questions provide copyable probes.

2. Your sample hides or manufactures variation

Six friendly power users may agree quickly because you have saturated one narrow group. That is valid only if the decision is about that group.

At the other extreme, a study that mixes new trial users, enterprise administrators, daily operators, and churned customers may never settle. The problem is not necessarily sample size; it is that several research questions have been placed in one tally.

Malterud and colleagues describe this through information power: sample needs depend on the study aim, sample specificity, relevant theory, dialogue quality, and analysis strategy. A narrow aim, specific sample, and deep conversations can support a smaller study; broad aims, convenience recruitment, and weak dialogue require more data.

Analyse each decision-relevant segment separately. When one segment meets the rule, stop recruiting that segment. Add another only because your decision needs it, not because a generic sample target says so. The guide to recruiting participants explains how to build those cells deliberately.

3. The analyst has stopped noticing

Founders enter interviews with product knowledge, strong hypotheses, and a personal stake in the result. That makes fast interpretation possible—and makes confirmation bias especially plausible.

A low new-theme rate can reflect the codebook rather than the conversations. If every complaint is forced into “CSV issue,” distinctions between file format, permissions, error feedback, and mapping may disappear.

Use three checks before stopping:

1. Keep a column for evidence that contradicts the current story.
2. Let someone who did not run the interviews code a small transcript sample without seeing your theme list.
3. Revisit early transcripts with the latest code definitions and check whether the classification still holds.

Agreement is not proof, but a second reading can reveal themes your current story made invisible.

Set the rule by the decision

Choose the settings before interview one. A useful decision table looks like this:

The more consequential and irreversible the decision, the stronger the evidence should be. More interviews are only one way to strengthen it. Better sampling, deeper probes, independent analysis, product logs, support records, and a reversible experiment may contribute more than another batch of similar conversations.

Stopping this study does not end customer research

Saturation is relative to a question, sample, time, and analytic approach. You may have enough evidence about why recent self-serve trials fail to import data. You do not automatically have enough evidence about enterprise setup, third-month retention, or whether the fix worked.

Close the study explicitly:

• record the stopping rule and when it fired;
• name the population and segments covered;
• list important counterevidence and missing groups;
• separate what the study supports from what remains unknown;
• state the next decision or test.

Then start a new study when the question changes.

How Maren supports the workflow

Maren can run adaptive user interviews, preserve the transcripts, and synthesise patterns across completed conversations. Use those capabilities to keep the guide consistent and probe participant-specific detail.

The stopping decision remains yours. Define the segment and threshold, review the underlying evidence, and check whether additional interviews are still adding new themes or meaning. A synthesis makes patterns easier to inspect; it does not turn a convenience sample into a representative one. See how Maren runs and analyses user interviews.