Pivot or persevere? Decide what experiment results justify

Decide whether to pivot or persevere by examining what the experiment actually tested, how trustworthy the observations are and which assumption they affect. A missed target can justify a better test, a limited refinement, a structural change, a pause or a stop. It does not identify the right response on its own.

Persevering means continuing a proposition whose important assumptions still have credible support, while addressing bounded weaknesses. A pivot changes a fundamental part of the proposition or strategy so a different hypothesis can be tested. Both require an account of the evidence and the next commitment.

This guide follows a fictional bicycle repair shop exploring an add-on service based on automated collection reminders. The shop hopes a dependable message will help customers collect promptly and free storage space. Its first trial produces some encouraging reactions but misses its collection threshold. The shop and every quantity or outcome below are hypothetical.

Define what would count as a pivot in this proposition

The Lean Startup methodology connects build-measure-learn with a structural course correction to test a new fundamental hypothesis. The distinction matters: correcting a broken link is usually a refinement. Changing the target customer, central promise, delivery channel or revenue logic can be a pivot if it changes what the proposition depends on.

For the shop, adjusting the wording of a reminder preserves the concept. Moving from “a message makes collection workable” to “an agreed evening collection appointment makes collection workable” changes the add-on’s promise and delivery model. It would be a bounded pivot of that concept, rather than a reinvention of the repair business.

Write what changes and what remains. Our Business Model Canvas guide helps trace whether a proposed pivot also changes resources, partners, costs or customer relationships. A new collection window involves staffing and building access; it is more than a new message template.

Do not use “pivot” as a flattering word for any alteration. Nor should a team avoid a necessary change because the term sounds dramatic. Name the specific fundamental belief that the new direction would replace.

Reconstruct the experiment before debating the verdict

Bring the original hypothesis, protocol, threshold, cohort, observations and version history to the review. If the team never agreed a decision rule, acknowledge that. It can still make a judgment, but it should not invent a threshold after seeing the result and describe it as pre-agreed.

The fictional shop’s initial rule was that at least 16 of 24 eligible customers would collect by the end of the next business day after being told their repair was ready. Ordinary closing hours stayed unchanged. The test also recorded message delivery, extra staff contact and storage use.

Use the small-experiment design playbook to make such conditions visible before the next round. A live MVP rather than a prototype is needed to observe actual collection choices. A screen demonstration can reveal understanding without establishing whether somebody can leave work in time to arrive.

Check the practical exposure. Were all 24 repairs genuinely ready? Did the message specify the complete repair and collection hours? Were refunds, payments or identity checks handled consistently? A trial with unreliable readiness information may test a different proposition from the one the team intended.

Separate the result from the explanation

Suppose the invented records show that 22 messages were delivered, two failed to deliver, and 10 of the 24 customers collected by the deadline. All 24 collection outcomes are recorded. The continuation threshold of 16 has not been met.

These observations do not mean that reminders caused the 10 collections or that the other 14 customers do not value the repair service. There is no control group. The two delivery failures expose one implementation problem; the remaining outcomes may involve timing, transport, payment or another dependency.

Strategyzer’s Learning Card distinguishes the tested hypothesis, observations, deductions and intended action. Use that separation literally. “Fourteen did not collect by the deadline” is an observation. “They do not care about collection” is an explanation needing support.

Interview a relevant range of customers about what happened. In this fictional follow-up, several late collectors describe working until after the shop closes. Treat this as a candidate explanation, not a population estimate. Selecting late collectors is useful for understanding that outcome, but cannot establish how common the difficulty is among all customers.

Do not discard positive evidence because the threshold was missed. Some customers may understand the message clearly or value the reassurance. Equally, polite praise should not overrule the actual behaviour the test was meant to observe. Preserve both without merging them into a single score.

Decide whether the evidence is weak, contrary or conditional

Weak evidence leaves uncertainty largely unresolved: an incomplete record, an unsuitable cohort, little opportunity to act or an intervention that was never delivered properly. Improve the design or narrow the conclusion. Continuing unchanged because the evidence is weak is not the same as having evidence in favour.

Contrary evidence directly challenges an assumption under conditions where the test was capable of exposing it. Customers who understand the offer, can use it and repeatedly choose an alternative provide a different signal from customers who never received it. A single disappointing count may contain both kinds of case.

Conditional evidence suggests the proposition works only in a narrower situation. The reminder may help people who already have flexible schedules while leaving commuters unable to collect. That could support a narrower target segment, a different delivery arrangement or a decision that the commercial opportunity is too small.

GOV.UK’s alpha guidance connects testing risky assumptions with deciding whether to progress. Its public-service context differs from a bicycle shop, but the useful discipline is the same: the next stage requires an account of what has been learned and what remains uncertain.

Return to assumption mapping after classifying the findings. A newly exposed constraint can change which belief deserves attention. The map is a working account of uncertainty, rather than a record that every earlier assumption has been permanently validated.

Compare five possible responses

The table below is an original decision aid. A real review may combine responses, such as fixing a delivery fault while investigating a different collection arrangement.

Response	When it may fit	What the decision must say
Persevere	Important assumptions have useful support and remaining risks are bounded	Why the next commitment is justified and which uncertainty remains
Refine	A specific implementation defect plausibly obscures the existing proposition	What changes, what stays and how the next test separates versions
Pivot	A central assumption is undermined and another credible proposition addresses the need	Which fundamental belief changes and what evidence motivates it
Pause	Harm, missing records, capacity or another condition prevents responsible interpretation or delivery	What must be resolved and who will decide whether to restart
Stop	Need, feasibility or value lacks sufficient support for further justified effort	Why additional learning is not worth the commitment now

Stopping need not mean that the team learned nothing. It may preserve resources for a more promising question. Persevering need not mean expanding: it can authorise one carefully bounded next investigation.

If the alternatives involve materially different commitments, a weighted decision matrix can expose trade-offs. Use anchored evidence for cost and effort. A ranking based on imagined benefits would simply make the uncertainty harder to see.

Make a small pivot that can be tested

The fictional shop chooses to investigate an agreed evening collection window. Its reason is specific: timely notification may be understandable while collection remains incompatible with some customers’ working day. The new hypothesis concerns access, rather than a more persuasive reminder.

It proposes inviting a small cohort of customers who have reported the timing problem to one scheduled evening session. Before the session, the team defines completed collection, appointment information, payment handling, staff time and the fallback if a customer cannot attend.

This is an investigation of an alternative, not evidence that the alternative will work. A concept test can first examine whether customers understand the appointment offer and what it requires. A later live session observes behaviour and operating effort.

Do not compare the new selected cohort’s percentage directly with the earlier 24-customer trial as if only the collection window changed. Selection, customer circumstances and scheduling have changed too. A causal comparison would need a different design. The next small test can still reveal whether the arrangement is practically deliverable.

Also inspect whether the new promise serves the business. Extra opening hours might free storage while using more staff time than the benefit warrants. A customer-facing improvement and a sustainable operating model are related questions, but they require distinct observations.

Protect the review from selective storytelling

Ask a person who did not lead the experiment to challenge the interpretation. What observation would weaken the preferred explanation? Which customers or operating periods are missing? Are the same standards applied to evidence for and against continuing?

Keep prior commitments out of the evidence column. The time already spent explains the organisation’s position; it does not establish that the next expenditure is useful. Use staged innovation funding to describe the cost and learning purpose of the next commitment.

Set a review date and a condition for stopping another inconclusive round. Otherwise a team can repeatedly call small changes “learning” while never changing the decision. A useful continuation makes the future evidence requirement clearer.

Sometimes the test challenges the goal itself. The shop may discover that next-day collection is an internally convenient target that does not reflect customer circumstances. Single-loop and double-loop learning distinguishes improving execution within a rule from reconsidering the governing rule. Questioning the deadline still requires the owner to assess storage, service and financial consequences.

Write a decision record that survives the meeting

Keep the original result and the revised proposition together. State the decision, owner, evidence, competing explanations, remaining risks, next test and authorised resources. Describe which observations would change the next decision in either direction.

For the example, the record might read: “The reminder trial missed its timely-collection threshold. Message comprehension remains promising, while delivery defects and reported opening-hour constraints require separate attention. We will correct delivery monitoring and investigate one evening appointment session before deciding whether to change the add-on proposition.”

Connect the commitment to the innovation strategy. A feasible alternative may still be outside the business’s intended scope or capacity. That is a strategic decision requiring an explanation, rather than a conclusion extracted from the pilot.

When the investigation changes staff routines, the Lean Change approach can help connect feedback, participation and further small steps. It does not remove the need to agree decision rights and operating responsibilities.

What research can and cannot tell you about pivoting

Camuffo and colleagues’ randomised study examined training in hypothesis-based entrepreneurial decision making with 116 Italian startups. Its publisher abstract reports differences in performance and pivoting between groups. That provides a reason to take structured testing seriously; it does not supply a universal pivot threshold for this repair-shop example.

The full study methods and estimates are not assessed here. Avoid claiming that every startup should pivot more frequently or that following this decision record will improve revenue. A pivot is useful only if the changed proposition and subsequent evidence justify it.

The Magenta Book’s Test and Learn guidance distinguishes early testing and refinement from the stronger evaluation needed when considering wider scale. Its emphasis on context is valuable: a promising local outcome may support adaptation rather than wholesale expansion.

The shop’s decision remains bounded by a small cohort, incomplete delivery and no counterfactual comparison. The practical achievement is a defensible change in what the team believes and what it will investigate next. Neither a disappointing result nor an encouraging one removes the need for that judgment.