Syntha Lab RU

Customers and loyalty

Who brings the revenue, whom to win back, whether the loyalty program pays and what extra points of returns are worth.

both Sales and channels

Open interactively →

About the topic

Revenue almost always rests on a small share of customers, and a brand needs to know exactly which ones. The rest buy rarely, in small amounts, or have left and not returned. The topic is about how to divide customers by behavior, decide whom to retain, whom to win back and on whom not to spend, and how to check whether a loyalty program really creates purchases or just hands out discounts to people who would have bought anyway.

The decisions here are these. Which segments get which communication and budget. What an extra percentage point of repeat purchases is worth and how much can be spent on retention. Whether the loyalty program pays back. How much revenue depends on a few large customers (for wholesale this is a matter of survival).

The reasoning rests on three dimensions of behavior. Recency: when the customer last bought. Frequency: how often. Monetary value: how much in the period. These three signs (RFM) give segments: champions (recent, frequent, large), sleepers (long ago, once frequent), newcomers and so on. Each segment gets an action and a cost: champions get early access, not a discount, sleepers are offered a reason to return, non-target customers are left alone. It helps to word the decision concretely: not “improve loyalty” but “win back those silent for three months, spending no more than RUB 800 per customer”.

The main catch of a loyalty program is self-selection. Members often spend more even without it, because the most active people join. Example: members spend RUB 24,000 a year, others RUB 12,000, and the RUB 12,000 difference looks like the program’s merit. But before joining, members spent RUB 21,000, so the real uplift is RUB 3,000, which at a 50% margin is RUB 1,500. A 5% discount on RUB 24,000 costs RUB 1,200: the program barely pays back, and the cost of running it still has to be subtracted. So to evaluate a program you compare behavior before and after and against a control group, not members against non-members.

Start with a customer table that has the date of the last purchase, the number of orders and the amount. Traps: segments without actions, comparing members with the rest without a control group, measuring return rate by the number of orders instead of cohorts, and giving a discount to those who would have bought without it.

Key ideas

RFM: recency, frequency, monetary value

Three behavioral signs divide the base into understandable groups. A customer who bought a month ago, three times in a year, for RUB 30,000 is different from one who bought once a year and a half ago for RUB 3,000. A segment is useful only together with an action and the cost of that action.

Self-selection

A control group removes self-selection. Out of 2,000 customers, 1,000 were chosen at random and invited to the program. A year later the invited spent RUB 18,000 and the rest RUB 16,500: an effect of RUB 1,500, at a 50% margin that is RUB 750. A 3% discount costs RUB 540, so the net gain is RUB 210 per customer. Random choice makes the groups comparable.

The price of a point of repeat rate

A point of repeat rate can be priced in roubles. If the base is 20,000 buyers, the average order is RUB 6,000 and the margin is 50%, every returning customer brings RUB 3,000 of profit from one extra order. One percentage point (200 customers) gives RUB 600,000: this is the ceiling on retention spending, above it winning back loses money.

Customer concentration

The share of revenue from the five largest customers shows how fragile the business is. If a customer provides 12% of revenue (RUB 12 million out of 100) at a 35% gross margin, its departure removes RUB 4.2 million of gross profit while fixed costs remain. The higher the share, the more important the contract, diversification and payment terms.

Discount to those who would have bought anyway

Any discount splits into two flows: those whose decision it changed and those who would have bought at full price. For the second group the discount costs money with no return. So the aim of loyalty is to reward behavior that would not otherwise have happened, not to pay for what already exists.

A segment is read as a separate survey

With 400 answers the margin of the overall estimate is about ±5.3 pp (95%, quota effect 1.2). A segment of a quarter of the sample is 100 answers, and its margin is about ±10.6 pp: twice as wide. Segments cannot be compared by a difference of a few points; either the sample is enlarged (for three segments with a margin of ±5 pp 1,332 answers are needed) or only large differences are read.

Formula: Segment margin = z × √(effect × p × (1 − p) ÷ n of the segment)

What to do: Decide in advance which segments you will read and count the sample for each; otherwise read only large differences.

Think first: 400 answers are collected, the margin of the overall estimate is ±5.3 pp What is roughly the margin of a segment of 100 answers?

Answer: About ±10.6 pp: four times fewer answers, the margin twice as large.

The lapsed threshold is taken from data

The threshold is a quantile of the gaps between repeat purchases: at a share of 80% in the example 248 days, “lost” after 496. Fixed 90 and 180 days are not tied to your buying rhythm.

Formula: Threshold = the P quantile of gaps with linear interpolation at position (n − 1) × P

What to do: Count by category and season and launch the win-back for sleeping customers.

Think first: There are only five gaps. Can the threshold be trusted?

Answer: No: a threshold from such a sample is shaky, at least a few dozen are needed.

NPS without an interval misleads

NPS has a spread that depends on the number of responses: 32 ± 7.7 at 400 responses. A rise of 6 points at z = 1.04 is within the noise.

Formula: z = (NPS₂ − NPS₁) ÷ √(σ₁² + σ₂²)

What to do: Report NPS with the interval and the number of responses.

How to work through it

  1. Prepare a customer table: identifier, date of the last purchase, number of orders, amount for the period.
  2. Divide customers into segments by recency, frequency and monetary value and count each one’s share of revenue.
  3. For each segment assign one action and its cost: retain, win back, leave alone.
  4. Estimate a point of repeat rate in roubles and set a ceiling for retention spending.
  5. Test the loyalty program: behavior before and after and a control group, not members against the rest.
  6. Calculate concentration: the share of the largest customers and the effect of one of them leaving.
  7. For points estimate the liability and expiry so the picture of costs is not lost.
  8. Repeat after a quarter on cohorts and check that segments really moved.

Where to start

Start with “RFM segments from your customer table”, then “Loyalty program economics” and “Repeat purchases”. Calculations of the topic in order: “RFM segments from your customer table”, “Repeat purchases: what extra points of return are worth”, “Loyalty program economics”, “Customer concentration: share of the largest”, “Lapsed-customer threshold from repurchase gaps”, “NPS confidence interval and significance”. See also (their primary topic is another one): “Customer base: active, repeat, new and returning”, “Cohort retention: curve, churn and cohort LTV”, “Customer segment migration: where the value flows”, “Win-back of lapsed customers: campaign payback and break-even response”, “CRM coverage: how many sales and customers you can see and reach”, “Bonus points liability and breakage”, “Survey sample size and margin of error: how many answers are needed and what can be read by segment”.

Calculations in this topic

  • RFM segments from your customer table

    Splits customers by recency, number of purchases and amount into segments, shows the share of customers and revenue in each and whom to win back first. Everything runs in the browser; the table goes nowhere.

    Who holds the revenue, whom to win back first and which segments are not worth spend.

  • Repeat purchases: what extra points of return are worth

    Shows how much extra gross profit a year comes from raising the share of returning customers by a set number of percentage points, what one point is worth and how much can be spent on retention.

    How much profit extra points of customer returns give and how much can be spent on it.

  • Loyalty program economics

    Compares the profit from members’ extra purchases with the cost of rewards and running. Allows for part of the member gap not being created by the program, shows the reward rate at which the program breaks even and the points that remain a liability.

    What the loyalty program gives beyond an ordinary discount and at what reward rate it loses money.

  • Customer concentration: share of the largest

    From a list of customers and their revenue computes the share of the largest, the top N and all the rest, the concentration index and the effective number of customers, and what happens to revenue and contribution if part of the top-N revenue leaves.

    How far revenue depends on a few large customers and what happens if part of their revenue leaves.

  • Lapsed-customer threshold from repurchase gaps

    Derives the threshold from your data and not from a round number: takes the gaps between repeat purchases and finds the period within which a set share of repeats falls (80% by default). A customer with no purchase for longer than the threshold is “sleeping”, longer than the threshold times a multiplier (2 by default) is “lost”. If you enter customers by the recency of the last purchase, shows the shares of active, sleeping and lost customers and the number of customers for a win-back campaign.

    After how many days without a purchase a customer can be considered sleeping and lost by your data, and how many customers are in each status.

  • NPS confidence interval and significance

    NPS is the difference between the shares of promoters and detractors, and it has a spread that depends on the sample size. The calculation gives NPS with a confidence interval and compares two periods: whether the change is significant or just survey noise. It also shows the minimum difference that can be noticed at all at your sample sizes.

    How uncertain NPS is at your sample and whether its change between periods is significant.

  • Customer base: active, repeat, new and returning

    Shows the health of the base over a period: how many customers came back, how many left, the share of new and repeat buyers and how the ticket and revenue per customer differ for new and returning ones. You set what “active” and “repeat” mean, so the figures are comparable only inside your company.

    Is the customer base healthy: how many customers came back, how many left, whom revenue rests on.

  • Cohort retention: curve, churn and cohort LTV

    For cohorts of customers (the month of the first purchase) it builds the retention curve: what share came back in the 1st, 2nd, 3rd and 4th period after the first purchase. It compares cohorts by LTV over a common horizon and shows which cohort brought more valuable customers.

    How fast cohorts lose customers and which cohort brought more valuable customers.

  • Customer segment migration: where the value flows

    For your segments (for example from RFM), ordered from the best to the worst, it counts how many customers moved up, stayed or moved down a step over a period and what it cost in money. Shows where the base loses value and what one step down costs.

    Where the base loses value: from which segments customers move down and what it costs.

  • Win-back of lapsed customers: campaign payback and break-even response

    Calculates whether a campaign to win back lapsed customers pays off: how many respond, how much margin is left after the discount, what part would have bought anyway and at what response the campaign breaks even. The discount and the response are your numbers.

    Will a win-back campaign pay off and at what response does it break even.

  • CRM coverage: how many sales and customers you can see and reach

    Shows what share of receipts and revenue is tied to a customer (card, phone), what part of the base consented to advertising messages and how much of it can actually be reached. Without this, conclusions about customers rest on a smaller part of sales.

    What part of sales and of the base you see and can reach: what customer analytics stands on at all.

  • Bonus points liability and breakage

    From points issued, redeemed and expired calculates the points balance, the obligation to members at face value, the historical breakage of a closed cohort, the reserve allowing for expected breakage, the cost of the program in revenue and the addition to the reserve already created.

    How much the brand owes members in points, what part will expire and how much reserve to hold.

  • Survey sample size and margin of error: how many answers are needed and what can be read by segment

    From the confidence level, the tolerable margin, the expected share, the size of the base, the effect of quotas and the response rate counts how many answers are needed, how many invitations to send and what it costs. For an already collected sample shows the margin of the overall estimate and of each segment: a segment several times smaller gives a wide interval.

    How many answers are needed for the chosen margin, how many invitations to send, what it costs and what margin each segment of an already collected sample has.

Self-check questions

  • What share of customers gives most of the revenue and which of them are going quiet?
  • Is an action assigned to each segment and is its cost known?
  • What does one percentage point of repeat purchases cost and what is the ceiling for retention spending?
  • Is there a control group for evaluating the loyalty program?
  • Is self-selection accounted for: how did members behave before joining?
  • What is the reward rate and at what rate does the program become loss-making?
  • What share of revenue do the five largest customers have?
  • What happens to profit if the largest customer leaves and is there a plan?
  • Is the return rate counted by cohorts, not by the total number of orders?

Knowledge check

Multiple-choice questions. The correct answer and the explanation follow each question.

  1. There are 10,000 active customers, 30% are in the loyalty program. Members buy 3.0 times a year, similar customers outside the program 2.2; 50% of the difference is attributed to the program. The basket is RUB 5,000, margin 50%. Rewards are 4% of the basket, 80% of points are used, servicing costs RUB 300,000. What is the program’s net effect for the year?

    • A. RUB 900,000
    • B. RUB 1,260,000
    • C. RUB 1,560,000
    • D. RUB 4,260,000

    Correct answer: B. RUB 1,260,000. Members 10,000 × 30% = 3,000. The frequency difference is 3.0 − 2.2 = 0.8, 50% attributed to the program: 0.4 purchases, or 1,200 additional orders. Gross profit 1,200 × 5,000 × 0.5 = RUB 3,000,000. Rewards 3,000 × 3.0 × 5,000 × 4% × 80% = 1,440,000, servicing 300,000, total 1,740,000. The effect is RUB 1,260,000. 4,260,000 comes from attributing the whole difference of 0.8 to the program (6,000,000 − 1,740,000), 900,000 from assuming everyone uses points (rewards 1,800,000), 1,560,000 from leaving out servicing. Unspent points remain a liability.

  2. The points balance at the start is 8.0 million, 5.0 million were accrued, 3.0 million redeemed for payment, 1.5 million expired; a point is worth RUB 1. In a closed cohort 30% of points expire, and a reserve of RUB 5.2 million is already created. How much must be added (confirm the reserve accounting with an accountant)?

    • A. RUB 1.8 million
    • B. RUB 3.3 million
    • C. RUB 5.95 million
    • D. RUB 0.75 million

    Correct answer: D. RUB 0.75 million. Points balance = 8.0 + 5.0 − 3.0 − 1.5 = 8.5 million. Not all points will be used, so the reserve is adjusted for expiry: 8.5 × (1 − 0.3) = RUB 5.95 million. 5.2 is already created, so add 5.95 − 5.2 = RUB 0.75 million. 1.8 comes from forgetting to subtract expired points from the balance (10.0 × 0.7 − 5.2), 3.3 from ignoring expiry (8.5 − 5.2), 5.95 is the reserve itself, not the addition.

  3. The January cohort has 1,000 customers, 400 of whom came back in the first period. The February cohort has 3,000, 600 came back. The March cohort has 1,000: its first period has not yet occurred. What is the weighted retention in the first period?

    • A. 20%
    • B. 30%
    • C. 25%
    • D. 40%

    Correct answer: C. 25%. Only cohorts whose period has already occurred are counted: (400 + 600) / (1,000 + 3,000) = 25%. The mean of the two percentages (40% and 20%) gives 30%, but ignores cohort sizes. 20% comes from also putting the March cohort, for which there is no data yet, in the denominator (1,000 / 5,000), 40% is January alone.

  4. The margin of the overall estimate is ±5.3 pp at 400 answers. What is the margin of a segment of 100 answers?

    • A. the same ±5.3 pp
    • B. about ±2.6 pp
    • C. about ±21.1 pp
    • D. about ±10.6 pp

    Correct answer: D. about ±10.6 pp. The margin is inversely proportional to the root of the number of answers: four times fewer answers, twice the margin. It grows fourfold only at a sixteenfold reduction. Covered in: A segment is read as a separate survey.

  5. A survey of ±5 pp, 95%, base 12,000, quota effect 1.2, 12% respond. How many invitations to send?

    • A. 3,700
    • B. 444
    • C. 1,332
    • D. 2,220

    Correct answer: A. 3,700. 444 answers are needed, invitations 444 ÷ 12% = 3,700. 444 is the number of answers and not invitations. Covered in: A segment is read as a separate survey.

  6. How is the lapsed-customer threshold derived?

    • A. Take 90 days for everyone
    • B. Take the warranty period
    • C. As a quantile of the gaps between repeat purchases
    • D. Take the number of shops

    Correct answer: C. As a quantile of the gaps between repeat purchases. The threshold is tied to the buying rhythm of a specific category. Covered in: The lapsed threshold is taken from data.

  7. When is a change in NPS considered significant?

    • A. When |z| is not below the quantile of the chosen confidence level
    • B. When NPS rose by at least a point
    • C. When there are over a hundred responses
    • D. Never

    Correct answer: A. When |z| is not below the quantile of the chosen confidence level. The difference must exceed the survey noise. Covered in: NPS without an interval misleads.

Common mistakes

  • Mistake: Comparing program members with the rest

    What it costs: Self-selection overstates the effect several times

    How to avoid: Compare before and after and against a control group.

  • Mistake: Building segments without actions

    What it costs: The table is pretty, sales do not change

    How to avoid: Attach an action and a cost to each segment.

  • Mistake: Measuring return rate by the total number of orders

    What it costs: New and old customers mix and the picture blurs

    How to avoid: Count by cohorts of first-purchase month.

  • Mistake: Giving a discount to everyone

    What it costs: Most of it goes to those who would have bought anyway

    How to avoid: Target the segments where a discount changes the decision.

  • Mistake: Not watching concentration

    What it costs: The departure of one customer breaks the plan

    How to avoid: Set an acceptable customer share and payment terms.

  • Mistake: A segment is read with the margin of the whole sample

    What it costs: A difference is seen where there is none A difference of 6 points lies within the margin: the decision about the group is made by noise.

    How to avoid: Count the sample for each segment. Enlarge the sample for the segments or decide only when the difference is larger than the sum of the margins.

    What it looks like (illustrative figures): The overall margin is ±5.3 pp, for a segment of 100 answers ±10.6 pp; the shares of two segments differ by 6 pp

  • Accepting fixed 90 or 180 days

    What it looks like (illustrative figures): Seasonal apparel has another buying rhythm.

  • Comparing NPS without an interval

    What it looks like (illustrative figures): A difference within the noise is announced as growth.

Russia and the EAEU

Customer data in Russia is regulated by the personal data law (Federal Law 152-FZ): storing and using data requires consent, a clear policy and storage of Russian citizens' data on servers in Russia. Email, SMS and messenger mailings are allowed only with the recipient’s consent (Article 18 of the advertising law 38-FZ), and advertising must be marked in the ERIR system. A loyalty program is set up through a public offer or rules: write down the lifetime of points and the expiry terms in advance.

Points and bonuses create a liability in accounting, and their tax treatment (a discount or a marketing expense, the effect on VAT) depends on the program design: check with your accountant. On marketplaces the customer base belongs to the platform: you receive buyer data in limited form, and loyalty is built within the platform’s rules, so your own site and messengers remain the main place where the data belongs to you. For wholesale customers concentration is estimated by counterparty; the credit limit and payment terms are fixed in the contract, and as the share of the largest buyers grows, watch overdue receivables.

Personal data and mailing requirements differ by country: Belarus, Kazakhstan, Armenia and Kyrgyzstan have their own personal data laws, and some require data localization; a guideline, check with a local lawyer. Payment habits and channels also differ: in Kazakhstan Kaspi and instalments are active, in Belarus and Armenia other payment systems and messengers dominate. Currency and VAT affect the receipt and the program: VAT rates by country are on the page “EAEU and CIS countries” (a guideline as of 14.11.2026, verify). Segments and calculations carry over unchanged, while reward rates and receipt thresholds are set in local currency.

Categories and audiences

Apparel
In apparel purchasing is seasonal and frequency is low (two to four purchases a year), so recency is counted in seasons, not weeks. A customer who has not bought for a month is not yet “sleeping”. Recommendations on size and style work well, while discount-based loyalty quickly teaches people to wait for sales.
Footwear
Footwear is bought less often than apparel and in a known size: a saved customer size makes a repeat purchase easier, while high size returns spoil loyalty. The replacement cycle is longer (a year or more), and segments are built by season and type of pair. Winning back lapsed customers works on wardrobe renewal.
Accessories
Accessories are bought on impulse and as gifts: the order is small, frequency varies, and the buyer is often not the owner. For gift purchases the holiday date matters more than RFM by recipient. Add-ons to a main purchase and gift cards work well, and a loyalty program can be cheap in terms of reward.
Kids
The customer here is the parent and the user is a child who grows: the need renews every season and with every size, so repeat purchases are naturally high. The customer’s lifetime is limited by the child’s age; segments are built by children’s ages, and consent for data belongs to the parent. A size-change reminder is effective.
Men
The men’s audience buys less often and in larger amounts, often out of necessity, and stays with a familiar brand longer, so retention is cheaper and winning back is easier. Many purchases are made by a woman as a gift. Discounts are less effective than service and quick selection, and segments differ more by frequency.
Women
The women’s audience buys more often and responds readily to novelties, collections and community, but also reacts more sharply to discounts: self-selection in a loyalty program is stronger. Rich preference data allows precise segments, but the risk of training customers to wait for sales is higher. Reward novelties and early access, not only discounts.

Questions this topic answers

  • What share of customers gives most of the revenue and which of them are we losing?
  • Does the loyalty program pay for itself once self-selection is removed?
  • What is one percentage point of repeat purchases worth?
  • Which customers bring most of the profit and who is going dormant?

Key measures and terms

Study first

What next

Next step: “CRM and the customer base: retention, segments, win-back”. Related: “Marketing budget and testing impact”, “Wholesale terms and credit”, “Store format, rent and the fate of a store”.

Documents on this topic

Open interactively →