The average family doesn't exist

An average is a correct summary of numbers and can still describe almost no real customer. Before you plan around a "typical customer", count how many customers actually look like it.

Leo's dashboard says Klasna's typical learner is about 15 and takes about six lessons a month. Both numbers are right: in the sample export of 160 accounts the mean age is 15.39 and the mean pace is 6.27 lessons a month, where a learner's pace = lessons_total / active_months.

Now filter the rows: learner_age = 15 and a pace of 5 to 7. The answer is 0. Eleven accounts are 15, and none of them takes 5–7 lessons a month. Widen it to ages 14–16 and you find 3 of 160 (1.88%).

The average sits in an empty spot between groups that pull in opposite directions, some taking many lessons for a few months, others few lessons for most of the year. A plan aimed at the average fits none of them.

When an average hides groups, split the file by columns that change behavior and compute the same measure inside each part.

In Klasna's export, filter goal = national_exam and first_lesson 2026-01 or later: 40 accounts that started exam prep only after New Year. Compute each account's pace (lessons_total / active_months), then average those 40 paces: 9.71 lessons a month, against 6.27 for the whole file.

That's a different kind of customer: families who arrive late, in a rush, and study two to three times a week. Their needs (a fast start, a plan that fits the weeks left, frequent slots) differ from a family that started in September.

Note the method: average the per-account paces, the same way the dashboard does, so the two numbers are comparable. Dividing the group's total lessons by its total months would give a slightly different figure.

Two groups can share a goal and still be different segments, because when and how they buy differs.

Klasna's exam families split by start month. Those with first_lesson 2025-12 or earlier, the steady exam families, are 24 accounts. Their mean pace (the average of each account's lessons_total / active_months) is 7.61 lessons a month, over about 9.67 active months, and 12 of 24 (50%) were still studying in September 2026.

Compare the spring crammers: 40 accounts that started in 2026-01 or later, 9.71 lessons a month, about 4.78 active months, 30% back in September. Same exam, two rhythms: one plans ahead and studies most of the year; the other arrives late and studies hard for a few months.

For marketing this means different timing (autumn vs winter), different offers (a year-long plan vs an intensive sprint) and possibly different messages, even though a dashboard would file both under "exam prep".

When a sample can be checked against a figure the business already trusts, check it. Agreement makes both more credible; disagreement is a question to answer before the plan relies on either.

Of the 40 spring exam accounts, look at last_lesson: 15 stopped in May and 13 in June, right around the exam (28 in all, 70%), and 12 show 2026-09, back for the new school year. 12 ÷ 40 = 30%, exactly the share Hanna's finance figures report for spring families returning in September.

The group's shape is now clear: they arrive in a rush (January–March), study hard (about 9.71 lessons a month) for a few months and mostly leave after the exam. That shape matters for marketing: they need to be reached in winter, and a September message to returning spring families is a separate, smaller opportunity.

A market segment is a group of customers who need, buy and behave alike, so that one marketing choice (an offer, a message, a channel, a timing) fits all of them. The test is practical: would the same decision work for everyone in the group?

Klasna's spring exam crammers pass it: they start exam prep in January–March, take about 9.71 lessons a month, and most stop after the exam. One choice, say an intensive winter package promoted in January, fits the whole group.

Groups that fail the test:

• An age band ("14 to 16") mixes crammers, steady exam families and others with different rhythms.
• A meaningless label such as an account id.
• A channel's audience ("website visitors") mixes every need.

The dashboard's average blended several real segments and hid every one. Finding the segments, not the average, is what row 2 of the plan is for.

First-party data is a company's own record of what people actually did: lessons taken, packages bought, sign-up answers. It's the best starting point for segments because it shows behavior, not opinions or guesses.

Its blind spot is built in: it only shows people who already came. Klasna's export of 160 accounts says nothing about families who looked at Klasna and chose free videos, a school teacher or waiting instead, nor about groups Klasna never reached at all. Segments found in it are segments of Klasna's current customers, not of the whole market.

Two habits keep that in view. Label findings for what they are ("among our customers, crammers are 25%"), not as market facts. And fill the gap with other sources: interviews with families who didn't sign up, the "what else did you consider?" answers, public data about the wider market. A group that's small in your data may be small because you never spoke to it.

Split customers by columns that change what they need and how they buy, never by labels without meaning or by sensitive traits.

Good candidates in Klasna's export:

• goal: an exam, school grades, a career, relocation and keeping up abroad call for different offers.
• first_lesson: starting in September or in February is a different rhythm and urgency.
• payer: a parent deciding for a child buys differently from an adult paying for themselves.
• country: families in the EU need other time slots and keep up with a different program.

Poor candidates: customer_id carries no meaning; it only orders the rows. And sensitive traits such as gender, ethnicity or religion: they aren't in the file on purpose. Klasna doesn't segment or target on them; they invite stereotypes, raise legal and ethical risks and don't explain behavior that goal, timing and payer explain far better.

A plan built around an average customer risks an offer and messages that fit no real group. The arithmetic isn't the problem; the summary is.

In Klasna's case, "a 15-year-old who takes six lessons a month" is almost absent from the data. Spring crammers want more lessons, fast, and stop after the exam; steady exam families study for most of the year at about 7.61; younger learners want about one lesson a week for most of the year; adults and EU diaspora families have their own patterns. A six-lessons-a-month package aimed at parents of 15-year-olds misses most of them, and the adults entirely.

Here plain rules on a few columns found the groups. When the groups aren't this obvious, clustering methods can suggest them; the Statistics for Product Managers and Marketers course covers that.

How it fits together: an average can describe nobody; splitting by goal and start month reveals groups with different pace and return rates; a segment is such a group; first-party data finds them, sensitive traits never. The next lesson describes each group so you can find it, and names the need that wins it.