Chapter 7: The Data That Predicts the Next Purchase

The Empty Bag, the Empty Bowl, and the Quiet Signal

A customer opens a bag of coffee beans on a Tuesday morning and notices there is only a handful left. The beans are a single-origin Ethiopian, roasted light, ordered three weeks ago from a small-batch roaster in Portland. The customer grinds the remaining beans, brews the pot, and makes a mental note to order more. The mental note dissolves in the day's meetings. A week later, the bag is empty. The customer opens the roaster's website, searches for the Ethiopian, and finds it is out of season. A replacement is chosen, but the delay has introduced friction. The customer, who once imagined a seamless rhythm of coffee delivery perfectly timed to consumption, is now drinking a supermarket blend and wondering whether the specialty roaster is worth the effort. The roaster, unaware of the empty bag, sends a promotional email for a new Kenyan release. The customer deletes it. The moment for an upsell — a larger bag, a subscription, a complementary coffee for the afternoon — has passed, not because the customer was unreceptive, but because the roaster did not know the bag was empty.

Now consider a different sequence. A pet owner orders a bag of grain-free dog food from a direct-to-consumer brand. The bag is a fifteen-pound size, and the product page, during the initial purchase, included a small, unobtrusive prompt: “Most customers with a medium-sized dog reorder in 28 days. Would you like us to send a reminder?” The pet owner checked the box. Twenty-eight days later, an email arrives. It does not say “Buy more food.” It says: “Charlie might be running low on his salmon and sweet potato recipe. We can ship another bag to arrive by the time the current one runs out. As a thank you for being a regular customer, here is a 10% discount on your next order — or try our new beef and pumpkin formula, on us, with your refill.” The pet owner opens the email, sees the offer, and clicks through. The reorder is placed in under a minute. The upsell — a new flavor, added to the refill order at no cost — is accepted not because of aggressive marketing but because the timing aligns precisely with the pet owner’s need state. The dog food brand knew when the bag would run out because it had analyzed the data from thousands of customers with the same dog size, the same bag size, and the same purchase history. The data predicted the next purchase, and the upsell rode on the back of the replenishment.

A third scene, in a different sector. A marketing manager at a mid-sized company is approaching the renewal date for a SaaS subscription. The platform, an analytics tool, tracks usage at the account level. Three months before renewal, the manager receives an email from the account executive, not with a generic “It’s time to renew” message, but with a customized usage summary: “Your team has run 43% more reports this quarter than last. The data suggests that adding our advanced attribution module would help you connect those reports to revenue outcomes. Here’s a proposal to add it to your renewal at a bundled rate.” The manager, who was already planning to renew, now considers the upsell. The timing is not accidental. The account executive’s outreach was triggered by an internal alert, generated by a cohort model that identified the usage threshold at which customers are most likely to accept an upsell — a moment when the product’s value is proven, the team’s investment is high, and the cost of switching away is rising. The data predicted the window of maximum receptivity, and the company acted on that prediction.

These scenes share a common structure. A purchase is made, a consumption pattern begins, and a clock starts ticking — a clock that measures the interval until the next purchase. The clock is not a guess. It is a statistical regularity that emerges when groups of customers are tracked over time. The customers who bought the same product, in the same quantity, for the same reason, tend to need it again in roughly the same timeframe. The coffee drinker who buys a twelve-ounce bag and brews two cups a day will empty the bag in about two weeks. The pet owner with a fifty-pound dog and a fifteen-pound bag will need a new bag in roughly four weeks. The SaaS team that adopted an analytics tool and grew its reporting volume by a certain percentage will hit the limits of its current plan in a predictable number of months. The regularity is not perfect — individual variation exists — but it is consistent enough at the cohort level to form the basis of a predictive upsell engine. This chapter examines that engine: how the data from repeat purchase behavior can be used to time an upsell not to a calendar event or a satisfaction spike, but to the moment when the customer is statistically most likely to need the product again.

Cohorts Are Calendars That Customers Write Themselves

Repeat purchase behavior is not random. It follows rhythms that are shaped by the nature of the product, the consumption rate of the customer, and the context in which the product is used. A consumable physical good — coffee, detergent, printer ink — depletes at a rate that can be estimated from the quantity purchased and the customer’s self-reported or inferred usage. A subscription service — software, streaming, membership — has a renewal cycle that is explicit, but the decision to renew or upgrade is influenced by usage intensity during the subscription period. A durable good — a laptop, a jacket, a piece of furniture — has a much longer repurchase cycle, but the purchase of complementary goods (accessories, add-ons, related items) often follows a shorter, predictable interval after the initial acquisition. In each case, the data from past purchases contains a signal about future purchasing behavior. The skill of the upsell architect lies in extracting that signal and converting it into a timed offer.

The primary tool for extracting this signal is cohort analysis. A cohort is a group of customers who share a common characteristic during a defined time period — typically, customers who made their first purchase in the same month, or who signed up for a subscription in the same week. By tracking the behavior of a cohort over time, an analyst can observe patterns that are invisible in aggregate data. One cohort may show that thirty percent of customers make a second purchase within thirty days, fifty percent within sixty days, and a plateau thereafter. Another cohort, defined by a different acquisition channel or a different first product, may show an entirely different pattern. The analysis reveals the prediction window: the timeframe during which a customer is statistically most likely to make their next purchase, and the point after which the probability of a repeat purchase declines sharply. The prediction window is not a fixed number for all customers. It is cohort-specific, and it changes as the business evolves.

Amazon Subscribe & Save is the most visible, large-scale implementation of cohort-driven replenishment timing. The program, launched in 2007, allows customers to schedule regular deliveries of thousands of eligible products — household essentials, groceries, pet supplies, personal care items — at a discounted price. The customer selects a product and a delivery frequency (e.g., every month, every two months), and Amazon ships the product on that schedule, applying a modest discount (typically five to fifteen percent) for the subscription commitment. The customer can adjust or cancel at any time. The program’s genius is not in the discount; it is in the removal of the reordering decision from the customer’s cognitive load. The customer does not need to remember to buy toilet paper; the toilet paper arrives on a schedule that, over time, Amazon’s algorithms can optimize based on the customer’s actual consumption patterns. If a customer consistently orders a six-pack of paper towels every three months but occasionally skips a delivery, the system can suggest a four-month interval or a larger pack size. The upsell — a larger quantity, a complementary product from the same category — is presented at the moment when the customer is already in a replenishment mindset, making the offer feel like a convenience rather than a pitch.

The data infrastructure behind Subscribe & Save is a cohort analysis engine operating at unprecedented scale. Amazon’s systems track not just individual purchase histories but also the consumption rates implied by those purchases. A customer who buys a thirty-two-ounce bottle of laundry detergent every forty-five days, with minimal variation, is assigned a predicted depletion date. A customer whose purchase intervals are erratic may be excluded from predictive upsells until a pattern emerges. The system uses these predictions to time reminder emails, suggest frequency adjustments, and cross-sell related products — dryer sheets to go with the detergent, for instance — at the moment of refill. The cross-sell conversion rates on these timed offers are significantly higher than on untriggered, batch promotions, because the offer arrives when the customer is already thinking about the category. Amazon does not publicly disclose Subscribe & Save conversion rates, but the program’s sustained growth — it accounted for a substantial and growing share of Amazon’s consumables revenue throughout the 2010s and 2020s, per annual reports — suggests that the model works at a level that justifies the infrastructure investment.

The underlying psychological mechanism is contextual priming. A customer who is about to run out of a product is mentally primed to consider purchases in that category. The brain’s attentional filters, tuned by the impending need, are more likely to notice and engage with offers related to that category. A coffee drinker with an empty bag is thinking about coffee. An offer for a new roast, a larger size, or a complementary brewing tool lands in a receptive mental field. The same offer, sent three weeks after the bag has been emptied and replaced with a competitor’s product, lands in an irrelevant field. The prediction window is not just about timing the offer to the need. It is about timing the offer to the moment when the need occupies the customer’s active consideration. The difference between the two is the difference between an offer that feels helpful and an offer that feels like spam.

The concept of the prediction window is closely tied to the economic value of a customer over time. CLV, introduced in Chapter 6 in the context of cancellation retention, is also the guiding metric for repeat purchase timing. A customer who makes a first purchase but not a second has a low CLV, and the business invests in win-back campaigns to recover them. A customer who makes a second purchase within the prediction window has a higher CLV, and the business can invest in upsells with confidence that the customer is likely to remain active. The prediction window segments the customer base by temporal probability: the near-term high-probability segment receives upsell offers that are aggressive but relevant; the medium-term moderate-probability segment receives gentle nudges; the long-term low-probability segment receives re-engagement campaigns rather than upsells. This segmentation is dynamic. A customer who misses the prediction window is moved to a different segment and a different treatment. The system learns from each customer’s behavior, refining the window for that individual while still anchoring it to the cohort baseline.

Baremetrics, a subscription analytics platform, has published case studies and benchmarks from its own data and from its customers that illustrate the power of cohort-based timing. In one publicly shared analysis, a SaaS company examined the conversion rate of upsell offers sent at different points in the customer lifecycle. Offers sent shortly before the customer was predicted to hit a usage limit — identified by a cohort model that tracked feature adoption trajectories — converted at nearly twice the rate of offers sent at a fixed calendar interval. The prediction window, in this case, was not based on a consumable’s depletion rate but on a usage threshold: the point at which the average customer in the cohort had exhausted the core features of their current plan and was likely to begin exploring upgrades. The company used this window to trigger a personalized upgrade email, framed around the specific features the customer was about to outgrow, and saw a measurable uplift in expansion revenue. The Baremetrics case underscores that the prediction window is not limited to physical replenishment. It applies wherever usage follows a trajectory that can be modeled, and it applies especially to subscription businesses where the upgrade decision is a function of growing engagement.

The technical requirements for implementing a prediction window are not negligible, but they have become increasingly accessible as analytics tools have matured. A small business can start with a simple spreadsheet: list all customers who made their first purchase in a given month, then count how many made a second purchase within thirty days, sixty days, ninety days. The resulting curve will be noisy with small samples, but the median interval will emerge. That median can be used as a first approximation of the prediction window. An email automation can then be set to trigger a reminder and an upsell offer at the median interval minus a few days, to arrive before the customer is likely to begin shopping elsewhere. The process can be refined over time as more data accumulates, and it can be automated using tools like Klaviyo, Mixpanel, or custom SQL queries on the transaction database. The important step is not the technical sophistication; it is the mental shift from treating all customers as if they are in the same buying state to recognizing that each customer is on their own clock, and the business that reads those clocks will sell more by selling at the right time.

The next section examines Amazon Subscribe & Save in more detail, as the primary case of a predictive replenishment engine operating at global scale. A secondary case, drawn from Baremetrics’ published analyses of SaaS expansion revenue, illustrates how the prediction window applies to non-physical subscription upgrades.

Amazon’s Subscribe & Save Engine — and Baremetrics’ Upgrade Curves

Amazon Subscribe & Save began not as a grand strategic initiative but as a response to a structural inefficiency in the sale of consumable goods. The inefficiency was simple: customers who purchased products they would predictably need again — diapers, detergent, coffee, pet food — were forced to make a new decision each time the product ran out. Each decision was an opportunity for the customer to choose a different retailer, a different brand, or a different product. Each decision was also a small cognitive burden, a recurring to-do item that added friction to the customer’s life. From Amazon’s perspective, each decision was a risk of leakage — a customer who might not return, might comparison-shop, might switch to a competitor’s subscription program. The program, launched quietly in 2007, addressed this inefficiency by collapsing the recurring decision into a single upfront choice: select a product, select a delivery frequency, and let the system handle the rest. The customer could adjust, skip, or cancel at any time. The promise was convenience, and the discount — initially modest, varying between five and fifteen percent depending on the number of items in the subscription — was the incentive to commit.

The architectural foundation of Subscribe & Save is a data model that connects purchase quantity to consumption rate. When a customer places a first Subscribe & Save order for, say, a 120-count box of baby wipes with a delivery frequency of every two months, the system records the product’s unit count, the customer’s self-selected frequency, and the shipment history. Over time, the system observes whether the customer adjusts the frequency, skips deliveries, or returns to order ahead of schedule. These behaviors are signals. A customer who consistently skips every other delivery is signaling that the frequency is too aggressive; the system can suggest a longer interval. A customer who orders an unscheduled refill between deliveries is signaling that the frequency is too conservative; the system can suggest a shorter interval or a larger pack size. The system is not merely executing a static subscription. It is learning the customer’s actual consumption rhythm from the customer’s own behavior, and it uses that learning to optimize the replenishment schedule and the accompanying offers.

The upsell mechanism is embedded in this rhythm. When a Subscribe & Save customer logs in to manage their subscriptions — to confirm an upcoming delivery, to adjust a frequency, to review what is arriving next month — the interface presents products that are complementary to the items in the subscription. The customer who subscribes to coffee pods might see a recommendation for a descaling solution, a milk frother, or a variety pack of limited-edition flavors. The customer who subscribes to dog food might see treats from the same brand, a dental chew, or a supplement. The recommendations are not generated by a generic collaborative filter. They are generated by an attribute-based system that understands the relationship between the subscribed product and the complementary product, a system built on the kind of data granularity discussed in Chapter 3. The timing of the upsell — during the subscription management session, when the customer is already thinking about replenishment — capitalizes on the contextual priming described in the previous section. The customer is in a replenishment mindset, and the offer arrives as a natural extension of the task at hand.

The quantitative impact of Subscribe & Save on Amazon’s business is difficult to isolate precisely because Amazon does not break out the program’s revenue in its financial reports. However, external analysts have estimated the program’s significance. In 2021, market research firm Consumer Intelligence Research Partners (CIRP) estimated that Subscribe & Save accounted for a substantial portion of Amazon’s consumables sales in the United States, with tens of millions of customers enrolled. The program’s stickiness — the tendency of customers to remain enrolled and continue receiving deliveries — was cited in Amazon’s annual letters to shareholders as a driver of repeat purchase behavior and customer loyalty. The discount offered to subscribers was not a cost to Amazon; it was an investment in retention and in upsell exposure. Every delivery confirmation email, every subscription management session, every pre-shipment notification was an opportunity to present a relevant complementary product to a customer who had already demonstrated loyalty to the platform and the category.

The predictive capability of the Subscribe & Save infrastructure extends beyond individual customers to the supply chain. By aggregating the scheduled deliveries across millions of subscribers, Amazon can forecast demand for specific products with a precision that far exceeds traditional retail forecasting based on point-of-sale data. This demand visibility reduces inventory costs, improves warehouse efficiency, and enables negotiated volume discounts from suppliers, some of which are passed back to the customer in the form of the subscribe-and-save discount. The program, in this sense, is not just a customer-facing convenience. It is a supply-chain optimization tool that funds its own incentives through operational savings. The upsells that emerge from the program — the complementary products added to refill orders — carry a marginal cost that is lower than the marginal cost of a standalone purchase, because they ride on the existing fulfillment and shipping infrastructure of the scheduled delivery.

A secondary case, drawn from the world of B2B SaaS, illustrates how the same predictive logic applies when the product is not a consumable physical good but a software subscription whose value grows with usage. Baremetrics, a subscription analytics and revenue recovery platform, has published analyses of expansion revenue — the revenue generated when existing customers upgrade to higher-tier plans — based on its own customer data and on benchmarks shared by its clients. One analysis, published on the Baremetrics blog in 2020 and updated in subsequent years, examined the upgrade patterns of SaaS companies and identified a consistent finding: customers who upgraded their subscriptions typically did so shortly after a usage threshold was crossed, not at a random point in the billing cycle. The threshold varied by company and product, but the pattern was robust. A project management tool might see upgrades spike when a team added a tenth project. A social media scheduling tool might see upgrades spike when a user attempted to connect a fifth social account. A customer support platform might see upgrades spike when ticket volume exceeded a certain monthly count. In each case, the upgrade was triggered by a functional constraint — a limit on projects, accounts, or usage — that the customer hit as a natural consequence of using the product successfully.

The Baremetrics data showed that SaaS companies that instrumented these usage thresholds and triggered upgrade prompts at the moment the threshold was approached — not after it was already a source of frustration — achieved higher expansion revenue than companies that waited for the customer to initiate the upgrade or that sent upgrade prompts on a fixed calendar schedule. One anonymized case study described a B2B SaaS tool that implemented a predictive upgrade trigger based on cohort usage trajectories. The company analyzed the usage data of customers who had upgraded in the past and identified the point in the usage curve — typically a month after crossing a specific feature adoption milestone — where the probability of upgrade acceptance peaked. The company then automated an email sequence that launched at that point, with messaging tailored to the specific feature the customer was likely to need next. The result, reported by the company to Baremetrics, was a twenty-three percent increase in upgrade conversion relative to the previous approach of sending upgrade offers at fixed intervals during the subscription term.

The Baremetrics case extends the Amazon Subscribe & Save logic to the non-physical domain. The “consumption” being tracked is not the depletion of a physical product but the adoption and deepening of software usage. The “prediction window” is not the interval until the next bag of dog food is needed but the point at which the customer is statistically most likely to be receptive to an expanded feature set. In both domains, the principle is the same: past behavior within a cohort predicts future receptivity, and timing the offer to the predicted moment of need — whether for replenishment or for expanded capability — yields higher conversion than timing it to a calendar event. The data does not need to be perfect. It needs to be directionally accurate enough to move the offer from a random point in the customer’s journey to a point that aligns with the customer’s own internal trajectory.

From Replenishment Rhythm to Expansion Trigger

The Amazon Subscribe & Save and Baremetrics cases operate on different product types, at different scales, and with different technical infrastructures. Yet they share a common logical spine: the data that predicts the next purchase is already inside the business, waiting to be structured. That data consists of the time intervals between a customer’s successive purchases, or between the adoption of one feature and the need for the next. When these intervals are grouped by cohort — customers who share a first-purchase month, a plan type, a product category, or a usage trajectory — they reveal a distribution. The distribution has a median, a spread, and a tail. The median is the prediction window. The spread tells the business how much variation to expect. The tail tells the business when to shift from upsell mode to re-engagement mode. None of this requires a data science team. It requires a willingness to look at the existing transaction log not as a record of what happened, but as a map of what will happen.

The principle that underpins prediction-driven upselling is that the customer’s future behavior is a continuation of the patterns already present in the customer’s cohort. This is not a psychological claim about individual constancy. Individuals change their habits, their budgets, and their preferences. But at the level of the cohort, the aggregate patterns are stable enough to be commercially useful. A cohort of customers who bought a twelve-ounce bag of specialty coffee will, on average, reorder in a certain number of days. The coffee roaster does not need to predict the exact day for each customer. It needs to predict a window during which a reminder and an upsell offer will, on average, find a receptive audience. If the window is twenty-one to twenty-eight days after the first purchase, an email sent on day twenty that says “Your coffee may be running low — here’s a new roast to try with your refill” will convert a measurable fraction of recipients. That fraction, applied across the customer base, generates revenue that is directly attributable to the timing choice.

The structural beauty of this approach is that it aligns the commercial offer with the customer’s own sense of need. The customer with an empty coffee bag is already planning to buy coffee. The upsell — a larger bag, a subscription, a complementary item — is presented as an option within a decision the customer was already going to make. The conversion lift comes not from persuading the customer to do something new, but from removing the effort of having to search, compare, and decide from scratch. The convenience is the value, and the upsell is the beneficiary of that value.

Extending this logic across three distinct business environments demonstrates its flexibility and its limits. For the small consumable-goods ecommerce business, the starting point is embarrassingly simple. The owner exports a list of all customers who have made at least two purchases, calculates the days between the first and second purchase for each, and finds the median. That median, even with a sample of fifty customers, is a working hypothesis. The owner then sets up an automated email — using any of the widely available marketing platforms — that triggers at the median interval minus a few days, reminding the customer to reorder and offering a complementary product at a modest discount or as a free sample with the refill. The offer is not a cold pitch; it is a service. The customer who was about to run out of the product sees the email and feels that the business is attentive, not intrusive. The owner can refine the window over time, segmenting by product category or by customer’s order size, but the first approximation costs an afternoon of spreadsheet work and a few hours of email configuration. The return, in the form of a permanently lifted repeat purchase rate, continues for as long as the automation runs.

For a B2B SaaS company, the prediction window is not about days until depletion but about the accumulation of usage that pushes the customer toward a plan limit. The data is already present in the product telemetry: the number of projects created, the number of reports run, the number of seats actively used, the number of integrations connected. The company can define a usage milestone that historically correlates with upgrade acceptance — for instance, the point at which a team has created fifteen projects and invited five members, beyond which the free tier’s project cap becomes a binding constraint. The analytics team then builds a simple model: among customers who reached that milestone in the past, what percentage upgraded within thirty days, and what was the median time to upgrade? That median becomes the prediction window. An automated in-app message or email, triggered when the customer crosses the milestone, can present the upgrade as a natural next step rather than a sales interruption. The framing matters: “You’re building momentum — here’s what you can unlock next” works because it acknowledges the customer’s progress and ties the upsell to that progress. The B2B context adds a layer of complexity because the user hitting the milestone may not be the person with purchasing authority. The system must either route the upgrade prompt to the account admin or equip the user with language to advocate internally — “Your team has hit the limit on projects; share this summary with your manager to unlock unlimited projects.” The data that predicts the next purchase can also predict the next internal conversation, and the system that facilitates that conversation multiplies its impact.

For a marketplace, the prediction window often applies to both sides of the transaction but with different rhythms. On the buyer side, a food delivery marketplace can predict that a customer who ordered dinner on a Friday night is more likely to order again the following Friday than on a Tuesday, because Friday dinner is a habit anchored to the end of the workweek. The marketplace can send a push notification on Thursday evening with a weekend offer, timing the upsell — a larger order, a premium restaurant, a dessert add-on — to the habit window. On the provider side, a freelance platform can predict that a new freelancer who completes five projects within the first month is likely to be receptive to a premium membership offer that promises better visibility, because the freelancer’s earnings trajectory has crossed a threshold where the membership fee represents a smaller percentage of income and a larger potential return. The platform’s data tracks both the transaction history and the engagement signals, and the prediction window for the provider upsell is defined by the point at which the freelancer’s activity pattern resembles that of previously upgraded providers. The marketplace’s advantage is its visibility into the entire economic relationship; its challenge is the complexity of modeling two distinct sets of behaviors simultaneously.

None of this requires the business to claim more knowledge than it actually has. The prediction window is a statistical construct, not a psychic insight. A well-designed upsell that arrives on day twenty-one of a coffee customer’s cycle opens with humility, not presumption: “If you’re running low on your Ethiopian, we have a fresh batch ready.” The “if” is essential. It acknowledges that the prediction may be wrong, that the customer may have brewed less coffee this month, may have bought a backup bag from a local shop, may have switched to tea. The offer that leaves room for the customer’s variability is trusted. The offer that presumes to know the customer’s kitchen inventory is resented. The distinction between the two is a single word, but it reflects a deeper orientation: the data is a guide, not a leash. The business that uses its prediction window to serve the customer, rather than to corner the customer, builds a relationship in which the upsell becomes part of the service rather than a separate commercial act.

Do You Know the Median Interval Between Purchase One and Purchase Two?

For the small business owner, the first question is practical and narrow. If you look at your last one hundred customers who made a second purchase, can you calculate the median number of days between their first and second purchase? If not, what is the single piece of data you are missing — a purchase date field in a spreadsheet, a timestamp in an order management system, a report in an ecommerce platform — and can you obtain it within the next business day? The question is not about building a predictive model. It is about establishing a baseline against which to test whether timed upsells outperform untimed ones. The second question is about the nature of the offer that should accompany the prediction. If you knew the exact day a customer would start thinking about reordering, what would you say to them that would feel like a genuine improvement to their experience rather than a promotional intrusion? The offer that adds a free sample, a relevant complementary product, or a flexible frequency adjustment is more likely to be accepted than a simple discount, because it respects the customer’s original motivation for buying from you — the product itself, not the price.

For the consultant, the first question concerns the design of a measurement framework. A client who invests in a predictive upsell system will eventually ask whether the investment is paying off. How would you isolate the incremental revenue attributable to the prediction window from the revenue that would have occurred anyway? One approach is a holdout group: randomly exclude a percentage of customers from the predictive offers and compare their repeat purchase behavior to the group that received them, over a period long enough to capture at least two purchase cycles. The challenge is that the holdout group is being denied a potentially beneficial experience, which raises ethical and business concerns. The consultant must design a test that is both rigorous and reversible — perhaps a short-duration experiment with a small holdout, followed by a full rollout with a post-hoc analysis comparing the behavior of customers who received the offers at different window settings. The second question concerns the organizational capabilities required to maintain a prediction-driven upsell system over time. Cohorts drift. Products change. Consumption patterns shift with seasons, with economic conditions, with competitive entries. Who in the organization is responsible for recalibrating the prediction windows, and how is that recalibration funded and staffed? Without an owner, the system becomes a legacy automation that fires on outdated intervals, and the uplift erodes. The consultant who helps the client establish this maintenance function is providing a structural solution, not a one-time optimization.

The practical one-week implication for a reader who wants to act is this: choose one product that customers buy repeatedly — a consumable, a replenishable, a subscription-eligible item — and calculate the median repurchase interval from the existing transaction data. If the data is insufficient, start collecting it now with a simple field in the order system. Then, set up one automated email or notification that triggers at that median interval minus a short buffer, offering a complementary product or an upgrade to a larger size. Measure the conversion rate on that email against the baseline of untriggered promotional emails over the following two weeks. The numbers will be small, but the direction will be informative. The purpose of the experiment is not to prove the principle definitively. It is to build the habit of looking at the customer’s purchase history not as a series of concluded transactions, but as the leading edge of a pattern that, with a little structure, can be read and responded to. The data that predicts the next purchase is already there. The question is whether the business chooses to listen.