Which first hundred days metrics actually hold up on a small sample?

Early metrics have one property that makes them dangerous: they are all available immediately and almost none of them mean anything yet. The instinct is to report whatever the dashboard shows. The discipline is to know which figures have had time to become real.

This page is about counting in a window too short for most of the ratios people want. No thresholds appear, because a threshold from someone else's business applied to a fortnight of your own is worse than no number at all.

What to take away

Cohort, before anything else

A cohort is the group who arrived in the same week. Every early figure has to be tied to one, and the reason is arithmetic rather than fashion.

If ten people arrive in week one and forty in week five, a week six whole-population retention figure is dominated by people who had one week to leave. It looks excellent.

It keeps looking excellent while arrivals grow. When they flatten, it collapses, and everyone hunts for a cause in the product.

Mark the arrival week on day one. It cannot be added later, and without it none of the rest of this page is available to you.

The return curve, read as counts

Read the shape, not the level. A curve that falls by the same amount each week has not found a floor. A curve that falls then flattens has one; the flat part is the group who genuinely want the thing.

At these sizes, do not compare your curve with anyone else's. Their offer, price, and audience differ, and those three produced their curve.

Two cohorts are worth more than one long cohort. If cohort one flattens at five out of ten and cohort two flattens at two out of ten, something changed between them: the audience, the message, or the product. That comparison is the single most useful early reading available, and it exists only if you marked the weeks.

Question density

The support and question log is a measuring instrument and almost nobody treats it as one.

Count questions per active customer per week, and group the questions by what they are about. Then read two things. Whether the total per customer is falling, which is what comprehension improving looks like. And which group is largest, which is where the product is unclear rather than where the loudest customer is.

A rising question count with flat arrivals means the product got harder, usually because something shipped. A question group that never shrinks after you fixed it means you fixed the wrong thing.

The refusal and cancellation record

The count that matters here is not how many left. It is what share of them you asked, and you control that entirely.

Record every cancellation with a date, cohort, and reason in the person's own words, not a dropdown. Dropdown labels collapse three different findings into one.

"Too expensive" from someone who never used it differs from the same phrase from a daily user. Only free text separates them. That same split appears in the pricing questions.

Canceling has to be easy for this record to exist at all. Where billing recurs, the ability to stop without obstacles is also the subject of the negative option rule, and a person who gives up and disputes the charge tells you nothing.

Four figures that mislead early

The one ratio worth building early

Not lifetime value. Recovery time: how many payments it takes to get back what it cost to acquire one customer, using costs you have actually paid and including your own hours at what you would have to pay somebody.

It is buildable inside a hundred days because both terms are observable. It is comparable against itself over time.

This figure decides how much cash growth needs, a different question from whether the business works. One-time and recurring costs behave differently in that calculation. The startup cost guidance states the difference clearly.

The wider identity this ratio sits inside is set out in the market metrics.

Writing the weekly figure so it stays honest

One line per cohort per week, in the same file, appended rather than overwritten.

Every figure carries three things: the count, the denominator, and the cohort. Any figure that arrives without all three does not go in the file. A dashboard can display whatever it likes; the file is the record, and the record is what you will read in month four when someone asks whether things improved.

Read four weeks at a time. Single-week movements at these sizes are noise, and the only signal available is a direction that holds. What to do once a direction holds is the subject of the first hundred days.

Common questions

How many customers before retention means anything?

There is no count that switches it on. What changes is what the number can carry: with ten you can see a shape and name a hypothesis, with a hundred you can compare two cohorts with some confidence. Say which of those you are doing.

Should I track engagement inside the product?

Only if a specific action predicts staying, and you can only learn that after enough people have stayed or not. Early on, tracking everything produces a dashboard nobody reads and no decisions.

What about comparing against industry benchmarks?

Do not. A published figure comes from a different offer at a different price to a different audience, and the comparison will be internally consistent and irrelevant. Derive what you need from your own arithmetic instead.

Is a fortnight ever enough to conclude something?

Yes, for arrival and comprehension questions. Where people came from and where they stopped in the purchase path are both readable quickly. Anything with the word "retention" in it is not.

What if the numbers are too small to write down?

Write them anyway, as fractions. Three out of four is a legitimate entry. It is the conversion to seventy five per cent that causes the damage.