Data analysis methods: from business question to decision

You have a table, can write a query and know how to build a chart. Then someone asks why sales fell or what should change in the process. This is where choosing an analytical approach becomes useful. Selecting a tool does not settle how to group customers, which periods to compare or what evidence you are actually looking for.

I will work through eight ways of examining a problem: segmentation, cohorts, RFM, funnels, unit economics and costs, distributions, root cause investigation and systems thinking. A fictional shop connects the examples. The aim is to choose a practical next step, rather than apply every method to every report. We will finish by separating findings from hypotheses and planning how a proposed change could be evaluated.

Start with the question and decision

“Analyse sales” is too broad a task. I would ask which decision is approaching: changing advertising spend, improving payments, offering something to returning customers or increasing warehouse staffing. Each requires different evidence. More orders might be a useful outcome for one decision and insufficient information for another, particularly if the cost of serving those orders is unknown.

Record the observation unit, period and outcome definition. A session, person, order and order line are different levels of detail. One person can have several sessions, while one order can contain many products. Joining tables without checking those relationships can multiply sales. The technical foundations are covered in SQL for analysts.

I would also establish which information could change the recommendation. If the decision depends on whether a problem affects one device type, start with that breakdown. Dozens of additional cuts can wait. Give the analysis a stopping condition: enough evidence for the stated decision, or a clear account of what is still missing and why it matters.

Check data availability separately. The system owner should confirm when event collection began and whether the latest period is complete. Trace a small sample from source records to the reported result. That can expose an incorrect filter before you spend time explaining a difference created by the query itself rather than the underlying business activity.

Example: conversion falls from 8% to 5%

Our fictional shop compares April and May 2026. Each month contains 1000 correctly recorded sessions. Conversion means the proportion of sessions containing at least one purchase. Several purchases in one session still produce one purchasing session. We assume unchanged measurement and no technical traffic; a real assignment would require those assumptions to be checked against evidence.

At the start of each session, we assign one of two groups. “New” means no earlier purchase in the customer’s history, while “returning” means at least one earlier purchase. These are our definitions, independent of any tool’s labels. A person can move between groups over time, but an individual session belongs to exactly one group.

There are 80 purchasing sessions in April and 50 in May. Conversion falls by three percentage points, equivalent to a relative decline of 37.5%. That difference does not yet establish whether payments need redesigning. We need to examine the composition of traffic and the rates within groups before assigning a cause to the overall movement.

Segmentation: check whom you are comparing

New-customer conversion is 20/500 = 4% in April and 35/875 = 4% in May. For returning customers, the figures are 60/500 = 12% and 15/125 = 12%. Each group’s rate is unchanged between months. What changes is the mix: new customers account for 50% of sessions initially and then 87.5%.

Applying April’s weights to May’s rates gives 0.5 × 4% + 0.5 × 12% = 8%. With this grouping, the entire overall decline can be explained arithmetically by the changed session mix. That is a description of aggregates. It does not prove why traffic changed or that every individual person behaved identically in both periods.

My next question would concern the origin of the extra new sessions and the reduction in returning sessions: campaigns, seasonality, product availability or customer identification. I would not immediately move the entire budget towards the group with higher conversion. Acquisition cost, order value and the group’s role in future growth may differ. Segmentation makes the comparison clearer but does not select an action by itself.

Record group sizes and assignment rules. An empty group has an undefined conversion ratio, rather than zero conversion. Missing customer history also does not necessarily mean that no earlier purchase occurred. If measurement cannot connect devices, disclose that limitation instead of treating the labels as a flawless account of individual people.

Cohorts: compare similar time since the start

A cohort groups people around a shared starting point, such as the month of their first purchase. You then observe behaviour during subsequent periods measured from that start. Google’s cohort documentation also separates inclusion and return criteria. This gives “do customers return?” a specific denominator and clock. A repeat purchase in the first month after joining is different from any repeat purchase during a full available year.

Our session table cannot produce a valid customer retention calculation. It lacks unique identifiers, first-purchase dates and complete return histories. Even if every new-customer purchasing session represents a different person, we do not know which of them returned. I would request the necessary fields instead of constructing a cohort table from an unsuitable observation unit.

After obtaining the history, I would define the return event and compare cohorts at the same age. A younger cohort has not yet had the opportunity to experience every subsequent month. Mark those cells as unobserved rather than zero. Also distinguish no purchase from unavailable data: those are different explanations for an empty result and should not be silently combined.

A decline in returns at a particular age identifies something to investigate. It does not establish poor onboarding or product quality by itself. I would check purchase seasonality, availability and acquisition-channel differences. The ecommerce KPI guide develops retention and acquisition-cost definitions further, including the importance of comparable observation windows.

RFM: describe purchasing history at a stated cutoff

RFM combines recency, frequency and monetary value. Microsoft’s documentation describes those dimensions and configurable ratings. In our example, I keep the raw values first. A scoring scale becomes useful only after the observation window, scoring direction and intended decision have been established. A three-digit score alone hides those choices.

Consider a separate small transaction extract from the same fictional shop: four customers and seven May orders. It is not the complete list behind the 50 purchasing sessions. Each extracted order has a unique identifier and comes from a separate session. Values are after discounts, excluding VAT, with no returns in this extract.

The calculation cutoff is the beginning of 1 June 2026, using a window from 1 May inclusive to 1 June exclusive. We use differences between UTC calendar dates. Recency is days since the latest purchase, Frequency counts orders in the window, and Monetary sums their value. A has R=2, F=2, M=200; B has 17, 1, 300; C has 12, 3, 60; D has 1, 1, 80.

B has the largest purchase value, C the highest frequency and D the most recent purchase. None of those findings establishes a profitability ranking. Product costs, later returns and service costs are unknown. I would classify a person without purchases separately instead of inventing a date. Scoring thresholds would be documented and checked for stability before being used to guide decisions.

Funnels: identify the stage and define progression

A funnel arranges events into a sequence. Google describes open and closed funnels and ordered steps in a user-based exploration. Below, I construct a custom session funnel, so its values are not automatically equivalent to a GA4 report. State the unit and rules alongside the chart rather than assuming that the visual shape communicates them.

For the same 1000 May sessions, we define a closed path: entry, product view, basket, payment started and purchase. All steps must occur in that order within one session, although other actions may happen between them. Each session is counted at most once at each stage. In this example, all 50 purchasing sessions meet the complete sequence.

Progression rates between adjacent stages are 60%, approximately 16.67%, 70% and approximately 71.43%. Conversion through the entire path is 5%. The largest loss is between product and basket: 500 sessions. Between payment started and purchase, the loss is 20. These figures locate attrition but do not reveal how much of it could realistically be recovered.

I would not assume that the largest loss offers the best return on work. Viewing a product may be part of comparing offers, and some people will come back later. Attrition at payment could involve an error, delivery charges, a missing preferred method or measurement. Logs, support reports and process observation are needed before a specific diagnosis can be supported.

Unit economics: connect volume with costs

The RFM extract contains 640 PLN across seven orders. That is revenue within the stated scope, with no profit information. B’s 300 PLN order may require more expensive goods or service than the others. Until costs are connected, you cannot call B the most profitable customer or promise the outcome of a campaign aimed at similar customers.

I would choose the decision unit: an order, acquired customer, parcel or working hour. Then I would assign variable costs and show how the remaining contribution covers fixed costs. Average parcel cost is not the cost of every order if orders split into multiple parcels. Likewise, acquisition cost depends on whom the business counts as a newly acquired customer.

Allocating a fixed cost does not mean that it disappears when a product is discontinued. In a scenario, ask what would actually change within the chosen horizon and whether added volume would exceed process capacity. The financial KPI article demonstrates COGS, contribution margin and OPEX calculations. Here we need them to evaluate a decision, rather than simply add another dashboard card.

Distributions: examine what the average contains

The seven order values are 50, 150, 300, 20, 20, 20 and 80 PLN. Their mean is approximately 91.43 PLN, the median is 50 PLN and the largest order is 300 PLN. Each statistic is valid and answers a different question. The mean helps reconcile the total; the median identifies the middle observation after sorting.

NIST describes histograms as a way to inspect features including spread, asymmetry and multiple clusters. With our seven observations, I would show every value instead of selecting impressive-looking bins. The sample is too small to claim a stable distribution for the shop’s entire sales population or precise knowledge of its tail behaviour.

Do not remove a large value simply because it differs from the others. Inspect the record for duplication, an incorrect currency, an adjustment or a real large order. If you report results after excluding observations, state the rule and show the original result too. In turnaround-time analysis, a similar question concerns delayed cases that the mean may describe poorly. The statistical mistakes article covers further interpretation problems.

Root cause investigation: explanations need evidence

We know that the session mix changed in this example. We still do not know why. The Five Whys method described by ASQ helps develop the questions, and the number of repetitions need not be exactly five. I would use it to organise a discussion and create hypotheses to check. A convincing chain of answers does not replace evidence.

For “a campaign attracted more new people,” I would inspect spending, traffic tags and change dates. For “we lost identification of returning customers,” I would inspect measurement versions, identifier consistency and event samples. Each hypothesis requires a different trace. Record evidence that could contradict the team’s preferred explanation as well, rather than collecting only observations that support it.

There need not be one universal cause. Campaigns and measurement could have changed simultaneously. Two monthly points are also insufficient to describe seasonality or establish a lasting trend. I would request a longer series, a change calendar and comparable periods. Conclusions should preserve the level of certainty that the available evidence actually supports.

Systems thinking: inspect effects beyond one KPI

Suppose the team is considering free delivery. I would map possible relationships: changes in order count and basket value, shipping cost, packing workload, delivery times and later returns. These are hypotheses to investigate, without guaranteed directions for every effect. Higher conversion alone does not establish whether the proposal improves the outcome of the complete process.

I would choose a primary measure linked to the objective, such as contribution per eligible customer, and monitoring measures for timeliness, returns and service workload. Their definitions and observation windows need to capture delayed effects. Additional orders may appear quickly while returns arrive later. Ending observation before those returns emerge would change the assessment of the proposal.

Operations colleagues can identify constraints missing from the sales table. A shared warehouse may serve several channels, so increasing traffic in one can affect the others. Record those connections and the owners of the relevant data. A relationship diagram maps assumptions; measurement is still needed to establish which relationships and effects actually occurred.

From a hypothesis to evaluation and a short recommendation

If the change is suitable for an experiment, eligible customers could be randomly assigned to variants. NIST explains randomisation as random assignment of a factor level to an experimental unit. In this proposed shop experiment, a person should retain the same variant across visits. Shared warehouse effects and mixed exposure also need consideration when deciding whether the design is appropriate.

Before starting, I would record the population, primary measure, smallest business-relevant change, sample-size plan, observation duration and stopping rules. I am not prescribing a universal number of customers or claiming that a test has been conducted. If random assignment is impossible, describe the limitations of the observational comparison explicitly. Choosing a method does not automatically remove differences between groups.

The final note for our example would be brief. Finding: conversion fell from 8% to 5%, with unchanged rates in two groups and changed group shares. Hypotheses: acquisition changes, changes involving returning customers or identification. Missing information: customer histories, costs and process evidence. Next step: check measurement and traffic sources with the person responsible for those data.

I would attach a table that reproduces the calculations and set a date for the next decision. I would not propose redesigning payments solely because the aggregate rate fell. Business analysis methods help turn a broad problem into questions that can be checked. Their value appears when the reader understands what has been established, what remains unknown and what will be investigated next.