RESOURCES/Methodology
Where the hypotheses come from
Markin generates its own revenue hypotheses instead of waiting to be given them. It reads behavioural, transactional, product and operational signal continuously, looks for where value is leaking or unclaimed, writes a hypothesis with an expected direction and a sized value, and puts it in a queue ordered by money rather than by opinion.
The bottleneck nobody budgets for
Every growth team has a backlog of ideas and a much shorter list of things it will actually test. The filter is not quality. It is who had time to size the idea, define the segment and write the brief. Hypotheses that would take a day of investigation to formulate never get formulated, so the business optimises the things that were easy to describe.
- Ideas that require cross-domain evidence rarely survive the backlog.
- Anything outside the campaign surface needs a different team, which means a different quarter.
- Technical causes of revenue loss are usually found by accident, months late.
Four domains, one queue
The point of authoring hypotheses rather than ranking given ones is that the search space stops being the campaign tool. All four domains compete in the same queue on the same currency: expected incremental revenue.
01
Detect
Continuous scanning for deviations that matter commercially: a cohort behaving differently from its own history, a funnel step degrading, a segment whose margin profile has drifted, an anomaly nobody reported.
02
Investigate
Before a hypothesis is written, the candidate cause is checked against alternatives. A conversion drop that is really a traffic-mix change should not become a pricing experiment.
03
Write the hypothesis
Stated with a direction, a mechanism and an eligible population. 'Doing X for population Y will increase Z, because W.' A hypothesis without a mechanism cannot be learned from, only won or lost.
04
Size it
Expected value is estimated on the eligible population, net of margin and contact cost, so the queue is ordered by money. Sizing is a forecast and is never reported as a result.
05
Kill it cheaply
Most hypotheses are wrong. The design target is that a wrong hypothesis costs a small holdout and a short window, so that expensive-to-formulate, high-information hypotheses become affordable.
What Markin is allowed to hypothesise about
| Domain | Typical hypothesis | How it is tested |
|---|---|---|
| Marketing | This cohort is worth a win-back offer at this depth, and this other cohort is worth leaving alone. | Randomised holdout on the eligible population, read on incremental margin. |
| Product | Customers who never reach this feature in week one churn at a materially higher rate; surfacing it earlier will move retention. | In-product treatment arm against a control, read across a full cycle. |
| Commercial | Annual framing beats monthly for this cohort once discount depth is priced against margin. | Price and packaging variant with a margin floor as a guardrail. |
| Technical health | Checkout error rate rose on one device and one region a fortnight ago and is suppressing conversion. | Anomaly confirmed against baseline, routed to the owning team, effect of the fix read against the pre-fix trend and a matched control where one exists. |
Sizing is a forecast, not a claim
Expected-value estimates used to rank the queue are explicitly not published as results. BCG's finding that 20% to 40% of measured next-best-action uplift fails an incrementality test is the reason we treat the two as different classes of number.
Independent researchBCG, incrementality in personalisation programmes (2026)
Score your own hypothesis pipeline
Take last quarter's tested hypotheses and classify them.
- What share were about a message, an offer or an audience?
- How many originated outside the growth or CRM team?
- How many were about pricing or packaging?
- How many started from an anomaly nobody was looking for?
- For each one, is the mechanism written down, or only the variant?
- How many were killed within four weeks? If the answer is none, the pipeline is only testing safe bets.
Where automated hypothesis generation adds little
- Businesses whose revenue problem is already well understood and simply unbuilt. Execution capacity is the constraint, not hypotheses.
- Environments where outcome data arrives too late to close the loop within a planning cycle.
- Highly regulated decisions where every treatment needs individual legal review, which reintroduces the human bottleneck by design.
- Very small bases, where nearly every hypothesis is underpowered and judgement beats testing.
Markin is an autonomous growth-science team for large B2C businesses. It investigates why revenue per customer is stuck, forms its own hypotheses across marketing, product, pricing and technical health, chooses the next best action for each customer, launches it through the systems the business already runs, and proves every one against a randomised holdout.
Decisioning tools choose between the actions your team already built. Markin decides what to build.
Questions people ask
- How does Markin decide which hypothesis to test first?
- By expected incremental revenue on the eligible population, net of margin and contact cost, with statistical power as a hard filter. A large opportunity that cannot be measured within a reasonable window ranks below a smaller one that can, because an unmeasurable result cannot be scaled responsibly.
- Does Markin only test marketing hypotheses?
- No. Marketing is one of four domains. Product, pricing and packaging, and technical health are in the same queue and compete on the same currency. In practice technical health is where the fastest, least contested wins usually sit, because nobody was looking for them.
- What stops it from generating noise?
- Two filters. A candidate must survive an investigation step that checks it against alternative explanations, and it must be sized above a threshold on an eligible population large enough to reach power. Anything that fails either filter never becomes a hypothesis.
- Who approves a hypothesis before it runs?
- Guardrails are set by humans in advance: margin floors, contact economics, brand and regulatory constraints, eligible populations. Within those, hypotheses run automatically. Anything that touches a guardrail is escalated rather than executed, which is the difference between autonomy and a system nobody can control.