From zero to loyal hero: growth through retention and referral

Retention and referral are where growth stops costing you every month

Most growth plans start with advertising. Advertising has one awkward property: you pay for it again every month, and the day you stop paying, it stops working.

Retention (customers who come back and buy again) and referral (customers who bring someone new) work on people you have already paid to win. In the pirate metrics framework, AARRR (Acquisition, Activation, Retention, Referral, Revenue), they are the two stages after the first purchase, and in my experience the two most neglected.

This article covers both, in the order I work on them. If the funnel itself is new to you, the article on pirate metrics and the one on full funnel frameworks set the scene.

Worked example: a coffee roaster in Bristol

Throughout this article I use an invented business, Harbourside Roasters, an online coffee roaster in Bristol run by Tom at harboursideroasters.example. Nobody real, every number made up.

Tom sells whole beans and ground coffee in 500g bags. His average order is £24, and after the beans, the bag, the postage and the payment fee he keeps £10 of margin per order. A customer who stays orders about six times a year, so an active customer brings him $6 \times £10 = £60$ of margin a year.

He spends £2,000 a month on advertising and wins 80 new customers from it. His cost to win a customer is therefore $£2{,}000 / 80 = £25$.

His advertising already pays. The question is how to get more out of the customers it brings.

Retention starts with a definition, written down before anything is counted

"Our retention is 40%" means nothing until you say three things: who is in the group, what counts as coming back, and within what time.

Analytics tools offer two common answers. N day retention is the share of users who come back on one specific day (or week, or month) after they started. Unbounded retention is the share who come back on that day or at any point later.¹ The old version of this article called the second one rolling retention.

For an online shop, "coming back" should almost always be an order, not a visit. A visit costs the customer nothing and says little. An order is the behaviour you want repeated.

Which definition you pick matters less than keeping it. A number that changes its meaning every quarter shows no trend.

Worked example: Tom's definition

Tom writes his down in one line: "A customer is retained in month $k$ if they placed at least one order in calendar month $k$ after the month of their first order."

Take Sophie: first order in January, again in February, nothing in March, again in April. She is retained in months one and three, not in month two. Under an unbounded definition she would count in all three.

He wants the N period version: a coffee customer who misses a month has probably bought beans elsewhere.

For the yearly figure he needs later: the share of one year's new customers who order at least once the following year.

A cohort table shows where customers stop coming back

A cohort is a group of customers who started in the same period. A cohort table gives each group a row and shows what share was still active in each later period. It answers what an average cannot: when do people leave?

The retention rate for a cohort in period $k$ is a simple ratio.

$$r_k = \frac{A_k}{N_0}$$

Here $r_k$ is the retention rate in period $k$, $A_k$ is the number of customers from the cohort who were active (by your definition) in period $k$, and $N_0$ is the number of customers in the cohort when it started. Foundations: ratios and rates

Always divide by the starting size, $N_0$, never by the number still around in the previous period.

Worked example: Tom's first four cohorts

Tom groups his customers by the month of their first order, January to April, and counts who ordered in each later month.

First order in	Customers	Month 1	Month 2	Month 3
January	80	28 (35.0%)	22 (27.5%)	20 (25.0%)
February	75	27 (36.0%)	21 (28.0%)	19 (25.3%)
March	90	32 (35.6%)	25 (27.8%)
April	85	30 (35.3%)

Each percentage is $A_k / N_0$, for January in month one $28 / 80 = 35.0\%$. For one curve, he adds up the counts rather than averaging the percentages, so a large cohort weighs more:

$$r_1 = \frac{28 + 27 + 32 + 30}{80 + 75 + 90 + 85} = \frac{117}{330} = 35.5\%$$

Here $r_1$ is month one retention across all four cohorts: active customers over starting sizes. Foundations: ratios and rates

In the same way, month two is $68 / 245 = 27.8\%$ (three cohorts) and month three is $39 / 155 = 25.2\%$ (two cohorts).

The horizontal axis is the number of months since a customer's first order. The vertical axis is the share of the cohort that ordered in that month, starting at 100% because everyone ordered in month zero by definition. The big fall is between the first order and month one: almost two in three new customers never come back. After that the curve flattens, so Tom has a core of regulars. His problem is the step from first order to second.

To measure how long customers last before they leave for good, use survival analysis, a topic of its own.

A point of retention is worth real money, and you can work out how much

The value of a new customer is the margin they bring in their first year, plus the margin of later years, counted for a little less, because a pound next year is worth less than a pound today.

$$V = M + M \times \frac{r}{1 + d - r}$$

Here $V$ is the value of a new customer in today's money, $M$ is the margin an active customer brings in a year, $r$ is the yearly retention rate (the share of customers who stay from one year to the next), and $d$ is the discount rate, the return you could earn on the money elsewhere. The second term is the standard lifetime value formula for the years after the first. Foundations: lifetime value and Foundations: discounting

Feed it margin, never revenue, and your own retention rate, never a figure from another industry. You will see headline numbers for what a small rise in retention does to profit. I leave them out: the formula gives you your own answer.

Worked example: what five points of retention are worth to Tom

Of the 900 customers Tom won last year, 360 ordered again this year, so $r = 360 / 900 = 0.40$. His margin is $M = £60$ a year and he uses $d = 0.10$.

$$V = 60 + 60 \times \frac{0.40}{1 + 0.10 - 0.40} = 60 + 60 \times \frac{0.40}{0.70} = 60 + 34.29 = £94.29$$

Here $V$ is the value of one of Tom's new customers, $M$ is his £60 of yearly margin per active customer, $r$ is his 40% yearly retention and $d$ is his 10% discount rate. Foundations: lifetime value

Now suppose he lifts retention by five percentage points, from 40% to 45%:

$$V = 60 + 60 \times \frac{0.45}{0.65} = 60 + 41.54 = £101.54$$

The symbols are the same as above, with $r = 0.45$. Foundations: lifetime value

Each new customer is now worth $£101.54 - £94.29 = £7.25$ more. Tom wins 80 customers a month, 960 a year, so a year's intake is worth $960 \times £7.25 = £6{,}960$ more, with nothing extra spent on advertising.

The horizontal axis is yearly retention, the vertical axis pounds. The rising line is a new customer's value, the flat line the £25 Tom pays to win one. The curve steepens: the move from 60% to 70% is worth £33 a customer, the move from 20% to 30% only £9.17. The better your retention already is, the more each further point is worth.

The first few weeks decide most of retention

In almost every shop and software product I have worked on, the largest loss is between the first purchase and the second. Someone who has bought twice has a small habit. Someone who has bought once has only tried you.

So the most valuable retention work happens early: getting a new customer to the moment the product feels worth it. In software that is the aha moment. In a shop it is more ordinary: the first cup tastes right, the jacket fits. To find what gets in the way, ask: a one line question in a follow up email, and a read through the customer service inbox. The article on onboarding and habit formation goes deeper into this for software products.

Worked example: Tom's first cup

Tom reads three months of customer emails. One complaint dominates: people brew ground coffee in a way that does not suit the grind, it tastes thin or bitter, and they blame the beans.

He makes two changes. At checkout, the grind choice now asks how the customer brews (cafetière, filter, espresso machine, stovetop) instead of "fine, medium, coarse". And every first order ships with a printed card for that method: how much coffee, how much water, how long.

He compares the share of new customers who reorder within 60 days, six months before and six months after: 144 of 480 before (30.0%) and 173 of 480 after (36.0%).

Could that be chance? A two proportion test answers it:

$$z = \frac{p_1 - p_2}{\sqrt{p(1-p)\left(\frac{1}{n_1} + \frac{1}{n_2}\right)}}$$

Here $z$ is the test statistic, $p_1$ and $p_2$ are the shares who reordered after and before the change, $n_1$ and $n_2$ are the number of new customers in each period, and $p$ is the pooled share across both periods together. A $z$ above 1.96 means the difference is unlikely to be chance at the usual 5% level. Foundations: hypothesis tests

The shares are $p_1 = 173 / 480 = 0.3604$ and $p_2 = 144 / 480 = 0.3000$, and the pooled share is $p = (173 + 144) / 960 = 0.3302$. Then:

$$z = \frac{0.3604 - 0.3000}{\sqrt{0.3302 \times 0.6698 \times \left(\frac{1}{480} + \frac{1}{480}\right)}} = \frac{0.0604}{0.0304} = 1.99$$

Here $z$ is the test statistic, $p_1 = 0.3604$ and $p_2 = 0.3000$ are the reorder shares after and before, $p = 0.3302$ is the pooled share, and $n_1 = n_2 = 480$ new customers in each period. Foundations: hypothesis tests

So 1.99, just over the line. After three months (72 of 240 against 86 of 240) it was 1.36, indistinguishable from noise. Since a before and after comparison can be fooled by the season, he also checks that last year showed no such jump. Statistical testing covers how much data a decision like this needs.

Lifecycle emails work when they follow the customer's rhythm, with permission

The next lever is reminding people at the moment they need you. A reminder when the coffee runs out is a service. A week early it is noise. Coffee, contact lenses and dog food run out on a schedule you can calculate. Clothes do not, and there a reminder is about something new.

In the UK, regulation 22 of the Privacy and Electronic Communications Regulations (PECR) bans marketing emails without the recipient's consent, with one exception for your own customers, the soft opt in. It applies when you got the person's details in the course of a sale, you market only your own similar products, and you gave them a simple, free way to refuse when you collected the details and in every message since.² A refill reminder usually fits it well. The article on consent and email capture covers the rules in full.

Points, badges and progress bars are another lever, and a topic of their own: loyalty and earn and burn.

Worked example: Tom's refill reminder

Tom works out the rhythm. A typical customer makes two cups a day at about 12g of coffee each, so they use 24g a day. A 500g bag lasts

$$t = \frac{500}{24} = 20.8 \text{ days}$$

Here $t$ is the number of days a bag lasts, 500 is the grams in a bag and 24 is the grams used per day. It is a rate turned around: grams per day becomes days per bag. Foundations: ratios and rates

So he sends a short email on day 18 after delivery: "Your Harbourside beans will be running low about now. Same again?" with a single button that refills the last order. Anyone who has already reordered gets nothing. Every email carries a one click unsubscribe, and the checkout offered the same refusal, as regulation 22 requires.

Together with the brewing cards, his yearly retention rises from 40% to 45% over the following year. That is the £6,960 a year from earlier, earned with a card and one well timed email.

Referral turns customers into a channel you can measure

Referral turns a funnel into a loop: a customer brings a new customer, with no advertising bill. It rarely happens at scale on its own. Two numbers manage it: the referral rate, the share of new customers who came through a referral, and the viral coefficient, usually called $k$, the number of new customers each customer brings in on average.

$$k = i \times c$$

Here $k$ is the viral coefficient, $i$ is the average number of people each customer invites (in practice, distinct new people who open their referral link), and $c$ is the share of those invited people who become customers. Each is a ratio of two counts. Foundations: ratios and rates

Referred customers refer again, so a starting group produces a chain. When $k$ is below 1 the chain shrinks at every step and adds up to a fixed total.

$$C_{\text{total}} = C_0 \left(1 + k + k^2 + k^3 + \dots\right) = \frac{C_0}{1 - k}$$

Here $C_{\text{total}}$ is the total number of customers that a starting group eventually produces, $C_0$ is the size of the starting group, and $k$ is the viral coefficient from above. The sum is a geometric series: each term is the one before multiplied by $k$, the same repeated multiplication that drives compound growth. Foundations: compounding

At or above 1 each generation is larger than the last, which is what people mean by self sustaining. Two caveats: $k$ has a time attached (1.1 a year is far slower than 1.1 a week), and it falls as you run out of people who have not heard of you. For a shop, in my experience, 0.1 to 0.2 is a good result: a discount on every customer you buy, not a rocket. The article on referral and viral mechanics covers the products where $k$ really does get large.

Worked example: Tom's referral numbers

Tom gives every customer a personal referral link and code. In a quarter, his 240 new customers' links are opened by 144 distinct new people, so $i = 144 / 240 = 0.6$. Of those 144, 22 place an order, so $c = 22 / 144 = 0.153$.

$$k = 0.6 \times 0.153 = 0.092$$

Here $k$ is Tom's viral coefficient, $i = 0.6$ is the invited people per customer and $c = 0.153$ is the share of them who order. Foundations: ratios and rates

So each 80 customers from advertising eventually become

$$C_{\text{total}} = \frac{80}{1 - 0.092} = \frac{80}{0.908} = 88.1$$

The symbols are as in the series above, with $C_0 = 80$. Foundations: compounding

About eight extra customers a month, which lowers his real cost per customer from £25 to $£2{,}000 / 88.1 = £22.70$. If he could double $k$ to 0.184, the same 80 would become $80 / 0.816 = 98.0$, and the cost per customer would fall to £20.40.

Referred customers tend to be worth more, so you can afford a fair reward

Referred customers tend to be better customers, not just less expensive ones. Philipp Schmitt, Bernd Skiera and Christophe Van den Bulte tracked about 10,000 customers of a German bank for almost three years. Referred customers had a higher contribution margin (a gap that shrank over time), a higher retention rate (a gap that did not shrink), and an average value at least 16% above comparable customers who were not referred.³ The bank paid the referrer a €25 voucher, and the authors estimate the return on that reward at about 60% over six years.³

That is a bank, so I would not copy the 16% into a forecast. The direction is what matters. And what to pay is arithmetic: any reward up to your advertising cost per customer is still a better deal than the advert.

The bank rewarded only the referrer.³ I usually reward both sides: in the shops I have worked on, a reward for the friend gives them a reason to act now, and a shared reward stops the referrer feeling they are selling to friends. That is my experience, not a proven rule, so test it.

Ask when the customer is happiest: after a second order, or after they have actually used the thing. The confirmation page of a first order is too early.

Worked example: Tom's reward

Tom's ceiling is his £25 advertising cost. He chooses £5 off the friend's first order and £5 credit for the referrer, £10 per referred customer, paid only when one arrives.

His 22 referred customers in the quarter cost him $22 \times £10 = £220$ in rewards. Winning the same 22 through advertising would have cost $22 \times £25 = £550$. Even valued as ordinary customers, the 22 bring $22 \times £94.29 = £2{,}074.38$ of value.

He moves the ask to the email after a customer's second delivery and onto the brewing card. One rule keeps it fair: the friend must be new, at a new delivery address.

The UK rules decide how you may ask for referrals and reviews

Two sets of UK rules apply: who sends the email, and reviews.

The first is PECR again. Regulation 22 applies not only to someone who sends a marketing email but to anyone who instigates one.² The Information Commission (still the ICO on its website), which replaced the Information Commissioner on 30 September 2026,⁴ says refer a friend schemes follow the same rules as other email marketing. They cannot rely on the soft opt in, and valid consent from the friend is unlikely to be possible in practice.⁵ A retailer that drafts a referral email for customers to send to friends is instigating it, and it does not comply.⁵ By contrast, a business that explains its scheme on an account page, without telling customers how to share it or suggesting they email or text their contacts, is not responsible for messages customers choose to send on their own.⁵

So never collect friends' addresses, never write to a friend until they come to you, and never draft the message for the customer to forward.

The second is the law on reviews. Since 6 April 2025,⁶ Schedule 20 of the Digital Markets, Competition and Consumers Act 2024 bans submitting or commissioning fake reviews, and reviews that conceal the fact they were incentivised. It also bans publishing reviews in a misleading way, for example removing the negative ones and keeping the positive ones.⁷ The Competition and Markets Authority (CMA) adds that simply emailing customers to ask whether they would like to leave a review, without steering what they write, is not prohibited. If you do offer something in return for a review, you must tell readers that the review was incentivised, and the review must still reflect the reviewer's genuine experience.⁸

Keep referral rewards and reviews apart.

The flow runs top to bottom. The point is who does what: the business writes only to its own customers, and the friend enters by their own choice.

Worked example: Tom's flow, checked against the rules

Tom had planned a "tell three friends" form that emailed each friend. That is the ICO's example of what not to do,⁵ so he drops it.

The code now sits on the account page and the brewing card, and his email to his own customer mentions the scheme in one line, with no ready made message to forward.

For reviews, every customer gets the same email two weeks after delivery asking for an honest review. No reward, no asking only the happy ones, every review published.

Partners can refer customers too, if both sides are counted

A business with the same customers and a different product can send you warm customers: a bakery and a roaster, a cycle shop and a bike fitter. Give each partner its own code, write down what you pay, and remember that if the partner is paid per customer, their posts about you are advertising, and the Advertising Standards Authority's guide on making clear that ads are ads covers how to label them.⁹

Worked example: Tom and a bakery in Bristol

A bakery in Bristol sells Tom's beans by the bag and hands every buyer a card with the code BAKERY10, worth 10% off a first online order. Tom pays the bakery £3 for every new customer who uses the code.

In three months, 31 new customers use it. Tom pays $31 \times £3 = £93$, and gives away 10% of 31 first orders of £24, which is $31 \times £2.40 = £74.40$. Total: £167.40, or £5.40 per customer, against £25 through advertising.

Word of mouth you cannot track can still be measured

Codes catch only part of word of mouth. Plenty of people hear about you over a coffee, search your name and buy without a link. That never shows up in $k$. (Sharing you cannot see at all is the subject of the article on dark social.)

I use three measures, because none is enough alone. The first is asking: an optional "How did you hear about us?" at checkout, the most direct evidence you will get.

The second is branded search. Google Search Console's Performance report shows the queries people typed before they clicked through to your site, with clicks and impressions for each.¹⁰ Searches for your name signal talk, unless a campaign explains them.

The third is the Net Promoter Score (NPS). It asks "How likely are you to recommend us to a friend or colleague?" on a scale from 0 to 10, and subtracts the percentage of detractors (0 to 6) from the percentage of promoters (9 or 10).¹¹ It measures what people say they would do, not what they do, so I treat it as a smoke alarm, not as a referral count.

Worked example: Tom's word of mouth multiplier

In a quarter, 300 new customers answer the checkout question and 54 (18%) say a friend or family member sent them. His codes credited 22. So for every tracked referral, $54 / 22 = 2.5$ customers say a friend sent them (an upper bound, as some of the 54 used a code).

The horizontal axis lists the answers, the vertical axis the number of new customers giving each, out of 300. Friends and family are the second largest source, more than twice what the codes show. So Tom's real $k$ is higher than 0.092, and a more visible code would capture a habit that already exists. His branded search clicks rose by about a fifth in the same quarter, with advertising flat.

Retention first, referral second, measured the same way every month

Referral multiplies whatever you have, so referral on a leaky first month mostly brings in more people who leave. Fix the step from first order to second, then ask where happy customers will see it.

The steps run top to bottom. The decision box is where most shops land on yes, in my experience, and it is why retention comes before referral.

Worked example: Tom's year

Twelve months on, Tom's advertising is unchanged at £2,000 a month. Retention is 45%, so each new customer is worth £101.54 instead of £94.29. His bakery sends about ten customers a month, and referrals turn every 80 customers from advertising into 88.

None of it was clever: a definition, a cohort table, a card in the box, an email on day 18, a code in the right place, counted the same way every month. Keep the customers you have, then give them an easy, honest way to bring their friends.

Footnotes

1. Amplitude, Retention analytics: retention analytics for stopping churn in its tracks, Amplitude, https://amplitude.com/guides/retention-analytics

2. The Privacy and Electronic Communications (EC Directive) Regulations 2003, regulation 22, Use of electronic mail for direct marketing purposes, as amended, legislation.gov.uk, https://www.legislation.gov.uk/uksi/2003/2426/regulation/22

3. Philipp Schmitt, Bernd Skiera and Christophe Van den Bulte, Referral programs and customer value, Journal of Marketing, volume 75, issue 1, pages 46 to 59, American Marketing Association, 2011, https://faculty.wharton.upenn.edu/wp-content/uploads/2012/04/Schmitt-Skiera-vandenBulte-2011-Referral-Programs-Customer-Value.pdf

4. Data (Use and Access) Act 2025, section 118, Abolition of the office of Information Commissioner, in force 30 September 2026 (SI 2026/1015), legislation.gov.uk, https://www.legislation.gov.uk/ukpga/2025/18/section/118

5. Information Commissioner's Office, Guidance on direct marketing using electronic mail: How do we comply with the PECR electronic mail marketing rules? (Can we ask people to pass on our electronic mail marketing?), ICO, https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/guidance-on-direct-marketing-using-electronic-mail/how-do-we-comply-with-the-pecr-electronic-mail-marketing-rules/

6. Department for Business and Trade, Fake reviews and sneaky hidden fees banned once and for all, GOV.UK, 6 April 2025, https://www.gov.uk/government/news/fake-reviews-and-sneaky-hidden-fees-banned-once-and-for-all

7. Digital Markets, Competition and Consumers Act 2024, Schedule 20, paragraph 13, legislation.gov.uk, 2024, https://www.legislation.gov.uk/ukpga/2024/13/schedule/20

8. Competition and Markets Authority, Fake reviews: guidance on the protection from unfair trading provisions in the Digital Markets, Competition and Consumers Act 2024 (CMA208), paragraphs 3.4 to 3.7, GOV.UK, 4 April 2025, https://assets.publishing.service.gov.uk/media/67eeb64fe9c76fa33048c790/CMA208_-_Fake_reviews_guidance.pdf

9. Advertising Standards Authority, Influencers' guide to making clear that ads are ads, ASA, fourth edition, 2026, https://www.asa.org.uk/resource/influencers-guide.html

10. Google, Performance report (Search results): overview and basic setup, Search Console Help, https://support.google.com/webmasters/answer/7576553

11. Bain and Company, Measuring your Net Promoter Score, Net Promoter System, https://www.netpromotersystem.com/about/measuring-your-net-promoter-score/