Chapter 19

Be kept

Today's customers write the evidence tomorrow's buyers read.

Kept is the stage where a customer chooses you again. They can also choose someone else, and more of them now ask a machine before they decide.

Most brands treat Kept as the point where marketing's job ends. The customer passes to service, and nobody owns the moment they decide whether to stay.

That gap costs more than one renewal. A kept customer writes much of the evidence the next buyer reads: the referral, the review, the forum post. Lose customers quietly and the earlier stages weaken, often with a delay and no obvious cause.

Every renewal is a fresh sale

A date arrives, a price rises or a rival launches, and the customer asks whether you're still the best option. More and more, they ask an AI assistant.

The clearest evidence comes from car insurance. J.D. Power's 2026 U.S. Auto Insurance Study surveyed 52,216 customers, and 32% of shoppers had used AI tools. Those who did were "more than 1.3 times as likely to switch insurers" (J.D. Power).

That's a correlation. People who already meant to leave may be the ones who turned to AI. A third of the AI users also found the AI content unhelpful. The adviser doesn't have to be good to be consulted.

Business buyers are on the move too. Google and National Research Group surveyed 2,063 US business leaders. Of those who had bought in the past six months, 58% had also switched supplier in that time (Digital Commerce 360). Read it as a direction, not a precise rate.

So the renew-or-switch moment is now a sale with an adviser in the room. The answer to "alternatives to X" is built from what others have written about you, much of it by your own customers.

Being the incumbent doesn't exempt you. In 2025 a state government agency we'd supported for more than four years put its website rebuild out to tender. We pitched as the incumbent, and lost.

The debrief was fair. Our method was well defined and our build was good value. But our case studies didn't show enough of the accessibility outcomes the panel cared about. Our team looked small for the risk. And our hosting and maintenance costs weren't set out clearly enough to compare. A newcomer would have had to answer every one of those questions. We'd assumed our history answered them for us. It didn't.

A renewal is judged like a first sale, against everyone else's evidence. We kept their campaign work, and we kept the lesson.

The loop

Kept customers produce the material the earlier stages run on. The seven stages run as a loop. Here's how it closes, one link at a time.

Referrals build Known. A happy customer puts your name in front of someone who has never heard it. At a German bank, researchers compared 5,181 referred customers with 4,633 others (Schmitt, Skiera and Van den Bulte).

Referred customers had an 18% lower risk of leaving, and that gap persisted. Their margin started about 25% higher, but that edge faded after about 29 months. Over six years they were worth about 16% more.

So a referral pays mostly through retention. It's one bank, but the direction is clear.

Your best spenders aren't your best referrers. Kumar, Petersen and Leone studied a telecom firm and a financial services firm. Customer lifetime value and customer referral value picked out different people (Kumar, Petersen and Leone). The work appeared in HBR, not a peer-reviewed journal.

Reviews build Believed. Chapter 16 set out how reviews lift purchase odds. What matters here is who writes them, and that depends on whether you ask.

Askalidis, Kim and Malthouse studied data from four retailers (Askalidis, Kim and Malthouse). Customers prompted by email gave ratings up to 0.5 stars higher than people who reviewed unprompted. Their ratings stayed stable while unprompted ones drifted down. Unless you ask, your quiet, satisfied customers are missing from the record.

The folklore says angry customers tell everyone. The unhappy do talk more, "but the difference is smaller than commonly believed" (Anderson).

Forums and review sites feed Found and Preferred. Reddit, Trustpilot and G2 all appear among the sources AI assistants cite. The vendors who count them disagree on how much (Chapter 22). When a buyer asks a machine to compare suppliers, a thread from a disappointed customer can sit next to your own pricing page.

Two companies show the loop working. In 2010 Dropbox's co-founder, Drew Houston, said a two-sided referral programme "permanently increased signups by 60%" (Houston). These are the company's own figures.

Tally shows the newer version. It grew for years on word of mouth, and now credits AI assistants with much of its discovery. It says "forums, Reddit posts, blog mentions, and authentic UGC" fed that (Tally).

No study measures the whole loop. Nobody has shown that churn or service failures led, later, to fewer AI recommendations. So the loop is my reading, not a measured effect. I act on it because each link pays for itself, as the referral sum below shows, even if the whole loop never closes.

Delivery is the marketing

Customers judge you against the promise that won the sale. Richard Oliver showed in 1980 that satisfaction depends on what a customer expected, and how far performance confirmed it (Oliver). Satisfaction then shapes whether they buy again.

The distance between the two is where the next review comes from. A promise kept is a referral in waiting. A promise broken is a forum post in waiting.

That's the idea behind the promise ledger: what was promised, where, how delivery is measured, the gap and the owner. Every gap in it is a future review.

Most of what it tracks sits outside marketing, in product, delivery and service. Marketing's part is to know the promises word for word, and to hear early when one slips.

Two companies that cut the wrong thing

Two well-known firms cut service at the very stage where customers decide whether to stay. Both then backed away.

In February 2024, Klarna said its OpenAI-powered assistant was handling two-thirds of its service chats and did the work of 700 agents (Tech.eu). In May 2025, chief executive Sebastian Siemiatkowski told Bloomberg that cost had been "a too predominant evaluation factor", and that quality had suffered. "Really, investing in the quality of human support is the way of the future for us," he said (Entrepreneur). The quotes come from secondary reports of that interview.

Commonwealth Bank of Australia cut 45 customer-service roles in 2025, citing a voice bot. On 21 August 2025 the bank reversed the cuts. It said it "should have been more thorough in our assessment of the roles required" (ABC). The reversal followed union pressure, so it shows a misjudged cut, not proven customer loss.

Both firms treated service as a cost to shrink. In the loop, service is where the next review gets written.

How to tell it is failing

• Nobody can name the person who owns the renewal moment.
• The first renewal conversation happens when the invoice goes out.
• An AI assistant, asked whether your product is worth renewing, quotes complaints you have never answered.
• Your newest public reviews are months old, or come mostly from unhappy customers.
• Sales promises things the delivery team has never heard of.

The method: the promise ledger and the renewal moment map

The promise ledger catches delivery gaps before customers write them up. The renewal moment map shows what each customer, and the machine they ask, will say before they decide.

1. Collect the promises, word for word. For each kind of customer, list the three to five promises that won the sale.
2. Note where each one is made. Ask an AI assistant what you promise too, because its version is the one many buyers hear.
3. Give each promise a delivery measure. Pick something you can observe: a report sent on time, a reply within a day, a result in the customer's own data.
4. Name one owner and one early warning. The owner is usually outside marketing. The warning is a ticket theme or a review phrase.
5. Rate and write the gap plainly. Use the scale below. Each gap is a future review. Either fix the delivery or change the promise.
6. List the renewal triggers. With a contract: the renewal date, a price rise, a failure, a rival's launch. Without one, list the repeat-purchase triggers: the season, the next job, a bad experience, a rival's offer, a new person in charge, a long gap. For each, note what the customer will ask, often of a machine, and what the machine says now.
7. Send proof before the question. With a contract, plan what you send at 90, 60 and 30 days. Without one, time it to the trigger: before the season, after the job, when the usual gap has passed. Send proof of value in each decision-maker's own terms.
8. Review both every quarter. Check them against fresh reviews and fresh AI answers. Read renewal, referral and review numbers together, allowing for a lag.

A simple keep rating helps the ledger set priorities. Higher is better, like every other scale in this book. This is my own working rule, not tested against sales.

Reviews and referrals: ask everyone, reward only referrals

Most of the review method belongs to Believed. Three points belong to this stage.

• Ask every customer, at a fixed point. Pick one moment, such as the end of the first quarter, and ask everyone who reaches it, happy or not. Asking only the customers you expect to be pleased is review gating. The US Federal Trade Commission's guidance says "Don't ask for reviews only from customers you think will leave positive ones" (FTC). In the UK, the DMCC Act's review bans took effect on 6 April 2025. The CMA's guidance treats cherry-picking positive reviews as one of them (CMS).
• Never reward a review. Rewards are for introductions, not opinions. Keep the referral ask in a separate message, sent at a different time.
• Ask on the sites buyers and machines read, not only on your own site, and answer every negative review in public.

For referrals, pick referrers by referral value, not spend, and reward both sides, as Dropbox did. Expect the payback mostly as retention. In the bank study, the margin edge faded and the lower churn stayed.

The referral sum, for finance

A CFO will ask what a referral is worth. Every input below is invented except the bank study's 16%; replace them with yours.

The sum is arithmetic, not evidence, and the 16% comes from one bank. The "would have come anyway" line matters most, so test it: offer the reward to half your customers for a quarter, and compare.

What keeping cannot do

Keeping customers doesn't, on its own, grow a brand. Romaniuk, Dawes and Nenycz-Thiel studied 1,093 brand-share changes across 63 UK packaged-goods categories (Romaniuk, Dawes and Nenycz-Thiel). When brands grew, penetration changed about three times as much as loyalty. When manufacturer brands declined, they lost both.

I accept that. My claim is narrower: not that retention beats acquisition, but that kept customers produce it.

Stated intentions are weak evidence too. Keiningham and colleagues found Net Promoter Score was rarely the best predictor of revenue growth among Norwegian firms (MeasuringU). Count who renewed and who referred, not who said they would. And loyal customers aren't better advocates: repeat customers' reviews didn't beat first-timers' at attracting new customers (Shin et al.).

So the loop runs on volume and freshness of honest evidence. You need many kept customers saying true things recently, not a few loyalists saying glowing ones.

The bookkeeper's promise ledger

The bookkeeper for cafés is an invented example. Here's her ledger after one quarter.

Three late Mondays are three chances for a review that says "great, when it arrives". She moved the send time earlier in holiday weeks and told clients why.

Her clients have no renewal date. They stop after a bad month, or when a new manager arrives. So she sends a three-month review before the summer holidays, and a note to any new manager on day one.

Fieldline's renewal moment map

Fieldline, the invented scheduling-software firm, lost Northgate to doing nothing. Meanwhile, Harbour Plumbing, an invented customer of similar size, reached its renewal.

At 60 days the finance lead did ask an assistant for cheaper alternatives. But he already had the 90-day review, in his own numbers. Harbour renewed, though finance pushed the price down first. Keeping a customer isn't the same as keeping your margin.

Fieldline asks every customer for a review at the end of their first year. Harbour's operations manager wrote a long one on G2. It gave the travel-time figure and a detail Fieldline never led with: job costs now matched invoices without a weekly spreadsheet. She told the same story at a regional trade forum.

Nine months after Northgate said no, Ravi was planning next year's budget. He asked an assistant: which field-service software makes job costs match invoices? This time Fieldline was named, and the answer cited Harbour's review.

Ravi forwarded it to Dana. Dana forwarded it to the CEO, with her own line about the vans on top. Northgate asked for a second meeting, and Chapter 27 picks up from there.

The story is invented. It shows the loop as I believe it works. A kept customer, asked like everyone else, wrote the evidence the next buyer read.

By mode

The renewal moment looks different in each mode. Habitual buyers re-order on autopilot until a stock-out or a bad batch prompts a look, so consistency and availability come first. Quick-choice buyers come back only if they remember you: a good first use, then a review ask at a fixed point. Considered personal renewals, such as insurance, are often checked with AI, so send proof of value before the notice arrives. Owners have no date and wonder after a bad month. Committee members each ask their own question as the contract end nears, so map the renewal moment per person.

How Kept connects to the other stages

• It needs Bought and Preferred. The promise made at the sale, and the reason behind it, are what delivery is measured against.
• It feeds Known, Found and Believed. Referrals spread your name, and reviews and forum posts are the evidence others and machines cite.
• The builds carry it. Part four covers the email, automation, PR and review work that runs this stage.

What not to do

• Don't cut service to fund acquisition. Klarna and Commonwealth Bank of Australia both cut where customers decide to stay, and both reversed.
• Don't ask only the happy ones. It skews the record, and regulators in the US and UK treat it as deceptive.
• Don't promise what delivery can't measure. If nobody can check it, nobody will catch the gap until a customer posts it.

How you know it worked

The Selectability Score (Part five) asks the same checks at the renew-or-switch moment. Does the machine name you when a customer asks for alternatives, for what reason, and accurately? Your own repeat and renewal data is the truth.

Then read four numbers of your own, together and allowing for a lag:

• Renewal or repeat rate, by kind of customer.
• The share of new customers who came by referral.
• Fresh public reviews per month, and their average rating.
• The share of AI answers about you that cite your customers' words.

Toolkit

The selection loop. Walk one real customer round it, and see where the loop leaks.

Read it clockwise from the next buyer: each promise kept feeds referrals and reviews, which recruit the buyer after that.

Every template for this stage is in the free Selectability Playbooks at rocmode.com/playbooks.

What we don't know

• Whether churn later weakens AI recommendations. No study links a firm's churn or service failures to a later fall in how often AI cites or recommends it. That's the central link in this chapter, and it's untested.
• How machines weigh customer text. How AI training and retrieval weight review and forum text isn't public. The citation shares used in this book come from vendor trackers.
• Business renewals and AI. I found no B2B evidence on AI use at renewal. The J.D. Power study covers US car insurance only.

Whose evidence this is. J.D. Power sells benchmarking and data to insurers. Google sells advertising, and G2 sells visibility on its review site. Kumar consults on customer value. Keiningham worked at Ipsos Loyalty, which sells rival measures. Dropbox's and Tally's figures are their own. Peer-reviewed sources: Schmitt, Skiera and Van den Bulte; Anderson; Oliver; Romaniuk, Dawes and Nenycz-Thiel; Shin and colleagues.