Author: sagi

Predictive member intelligence combines three systems: Predictive Lifetime Value, the Next-Purchase Window, and the Prediction-to-Action Loop. Together they identify who to prioritize, when to reach them, and how to make the prediction land as a member action, which is why top-quartile wineries grew DTC revenue 22% last year while the bottom quartile fell 13%.

Picture two DTC programs with the same stack, the same subscriber count, and the same data flowing in. One reports its numbers; the other predicts them and acts. Over a year, the second pulls ahead on the exact line both Directors are measured on, and not because it spent more or bought a better tool. It closed the distance between what its data already knew and what its members actually experienced. The spread is not hypothetical: last year top-quartile wineries grew DTC revenue 22 percent while the bottom quartile fell 13 percent (Silicon Valley Bank, DtC Wine Report 2026), on broadly similar tools.

This week covered the three systems that close that distance. Individually, most Directors recognize each. Together, they form something most programs have not yet built: member intelligence that reaches the member.

The Three Predictive Systems

Predictive Lifetime Value. Instead of waiting twelve to eighteen months for a trustworthy LTV number, you band members from their first 90 days, using order cadence, channel breadth, and engagement depth. You then point your scarcest attention at the predicted-high band while it can still shape behavior, which matters because roughly 40 percent of club members cancel inside their first year (SVB, 2026). Personalization tied to a member’s real standing carries a documented revenue lift of 5 to 15 percent (McKinsey), and early banding is how you aim it. The mechanism is the same onboarding budget, aimed, not a larger one.

Next-Purchase Window. Instead of a calendar release that interrupts everyone on the same Tuesday, you predict when each member is entering a buy decision, from consumption cadence, depletion timing, and category affinity, and reach them inside that window. The lift is in the published benchmarks: triggered, behavior-based emails click near 5 percent against roughly 1.5 to 2 percent for batch campaigns (Klaviyo, 2024), because relevance rises while total frequency holds or falls. The reorder revenue you currently lose in the gap between quarterly sends is the revenue this recovers.

Prediction-to-Action Loop. Instead of a score that sits on a dashboard, you resolve every prediction to one next-best-action, route it to an owner, measure the 90-day outcome, and feed that outcome back to sharpen the next prediction. Realistic churn intervention cuts losses 15 to 40 percent among targeted members (Gartner, McKinsey), and acting on member data compounds: members given control to edit their own packages show 20.7 percent higher AOV and roughly half the churn (Commerce7, 2025). The prediction reaches the member instead of the report.

How the Three Compound

The systems are not three separate projects. They feed one another.

Predictive Lifetime Value tells you who to prioritize. The Next-Purchase Window tells you when to reach them. The Prediction-to-Action Loop makes sure the prioritization and the timing become an action a member feels, and then learns from whether it worked. Value without timing reaches the right member on the wrong day. Timing without a loop produces a well-aimed message no one is assigned to send. The loop without value or timing acts on the wrong signal. Connected, they turn a stack that measures members into one that anticipates them.

This is the same principle behind a program we operate with 11,600 subscribers, where 48 percent of engaged subscribers have converted to buyers, sustained across more than four years. The 48 percent is the controllable part: disciplined synthesis of signals already present, turned into timed action. List growth around it is structural. The point for your program is not the number; it is that the number came from acting on data already collected, not from acquiring more of it.

That distinction matters more in 2026 than it did five years ago. Acquisition is harder and more expensive, and the channel is contracting: DTC shipments fell 15 percent in volume and 6 percent in value in 2025, the worst year since the report began in 2010, and the 11 percent rise in average bottle price was mix-shift as lower-priced buyers exited, not members trading up (Sovos ShipCompliant and WineBusiness Analytics, 2026). In that environment, the members you already have are the most defensible growth you own, and prediction is how you get more from them without spending into a headwind. None of the three systems asks you to buy a tool, migrate a platform, or add headcount. Each asks you to act earlier and more precisely on data your stack is already collecting and mostly wasting. That is the difference between a program that measures its members and one that anticipates them, and in a down market it is also the difference between a DTC line you defend and one that gets folded into someone else’s.

Which Prediction Is Your Gap

If you cannot say which of your current members will be worth the most, start with Predictive Lifetime Value. If you know their value but reach them on a calendar, start with the Next-Purchase Window. If you predict well but your scores never reach a member, start with the Prediction-to-Action Loop. The starting point is wherever the distance between your data and your members is widest.

The three-minute Winery Sales Growth Archetype quiz is built to locate that gap: whether your weak point is knowing a member’s value, knowing their timing, or acting on either, and which system will move your numbers by day 90 rather than in some distant year.

P.S. The fastest payback of the three is usually the Next-Purchase Window, because the data is already in your order history and the fix is a weekly segment, not a model. But it only holds if a loop carries it to the member every week. Prediction is cheap; the motion that acts on it is the scarce part, and it is the part that shows up on the line you defend.

A prediction only creates revenue once it resolves to one action, reaches an owner, and gets measured, not while it sits on a dashboard. The Prediction-to-Action Loop routes a next-best-action per member, ties it to a 90-day outcome, and feeds results back into next quarter’s predictions, which is how realistic churn intervention cuts losses 15 to 40% among targeted members.

Somewhere in your stack right now there is probably a score that could have saved a member, and didn’t, because no motion existed to carry it to the person who could act. This is the least discussed failure in DTC analytics. Teams invest in prediction, buy or build a model, stand up a dashboard, and then discover that a prediction with no attached action is just a more expensive way to watch the same members leave.

You have likely lived a version of this. The platform, or a tool you subscribed to, produces churn risk, or lifetime value, or purchase propensity. It renders as a number on a screen. And then the week fills with the pebbles it always fills with, and the number sits there, accurate and unused, until the flagged member cancels and the score gets to be right about a loss you could have prevented.

Why the Score Alone Does Nothing

The reason is structural, not a matter of discipline. A dashboard is a measurement layer. It tells you what is true. It does not tell anyone what to do about it, route that instruction to whoever executes it, or check afterward whether it worked. Those are three separate jobs, and no dashboard does them. This is exactly why the analytics tools most Directors have already tried stall: they solve “I can’t see my data,” and leave untouched the harder problem, “I don’t know what to do about what I’m seeing, and I don’t have time to do it every week.”

The fix is not a better model. Most programs already have predictions accurate enough to act on. The fix is the operating motion that sits between the prediction and the member.

It helps to see why the motion is the part that goes missing. A prediction becomes an action at an org-chart seam, the point where a number on your screen has to become something the club lead, or an automation, or the release process actually does. Nobody at your winery was hired to own that seam. The analyst, if there is one, owns the number. The club team owns the members. The handoff between them is unassigned, so it defaults to you, and your week is already full of the other unassigned handoffs. The prediction doesn’t fail because it was wrong. It fails because the calendar closed over it before anyone carried it across the seam. That is a coordination problem wearing an analytics costume, and no dashboard has ever solved a coordination problem.

The Prediction-to-Action Loop

Four steps turn a score into revenue, and each one is a habit, not a technology.

One next-best-action per member. The score has to resolve to a single instruction, not a range of possibilities. A member in the top churn-risk band and the top value band gets one action: prioritize her for the next allocation, or a personal call, whichever your data says works. Not a list of things someone could do. A decision, already made, so the person receiving it executes rather than deliberates.

Route it to an owner. The action lands with the channel or person responsible for it, on a schedule. The email action goes to the automation. The personal-touch action goes to the club lead’s Monday list. The allocation action goes to the release process. An instruction that lives in a report nobody is assigned to read is not routed; it is filed.

Measure the outcome. Tie the action to a number you already defend. Did the flagged member buy, stay, or upgrade inside 90 days? You are not inventing a new metric; you are connecting the action to DTC revenue, retention, or site-to-member conversion, the numbers already on your quarterly page. Measurement is what separates an action that works from one that merely feels responsible.

Feed the outcome back. The results of last quarter’s actions sharpen next quarter’s predictions. Members who were flagged and saved teach the rule what a save looks like; members flagged and lost anyway teach it where it was wrong. A loop that learns beats a model that is merely accurate on the day it ships, because your members and your market keep moving.

The steps look obvious written down, and that is exactly why they get skipped. Each one is a small coordination act, not a technical build, and small coordination acts are the first things a busy week drops. The discipline is not difficulty; it is refusing to let the loop break at the handoff. A prediction that resolves to an action nobody owns, or an action nobody measures, quietly reverts to a number on a screen within a month. Naming the owner and the 90-day check is what keeps the loop a loop rather than a good intention that ran once.

What the Loop Produces

The gain is not from a more sophisticated model; it is from closing the distance between a correct prediction and a member who never felt its effect. Independent measurement keeps the claim honest: realistic churn-intervention effects cluster in the range of a 15 to 40 percent reduction among targeted members (Gartner, McKinsey), not the near-total saves some vendors imply, which is exactly why the measurement and feedback steps matter. They keep you honest about what the loop actually returns, and they compound the part that works.

The effect of acting on member data, rather than only displaying it, is visible in wine-specific numbers. Club members given the control to edit their own packages, a small operational act of letting the data drive a choice, show 20.7 percent higher average order value and roughly half the churn of members who cannot (Commerce7 Data Drop, 2025, across 1.4 million memberships and 17,000 clubs). The predictions differ, but the lesson is the same: the revenue lives in the motion that reaches the member, not in the score that describes her.

The defensibility benefit is as real as the revenue one. A loop produces a record: this member was flagged, this action was taken, this was the 90-day outcome. That record is a far stronger answer at a quarterly review than “our model has good accuracy.” You are showing decisions and results, not a capability.

The feedback step is the one most programs skip, and it is the one that separates a loop that improves from a routine that merely repeats. When you record which flagged members were saved and which left anyway, you are not just measuring; you are teaching the rule what a real save looks like at your winery, with your members, at your price points. A model bought off the shelf is accurate on the day it ships and slowly drifts as your market moves. A loop that feeds outcomes back moves with the market, because every quarter’s results are the next quarter’s training data. Over a year, the gap between a static model and a learning loop widens quietly, and it widens in your favor. This is also the honest guardrail: the same measurement that sharpens the model keeps you from overclaiming, so the effect you report is the effect you actually produced, which is the only kind that survives a second look.

This Quarter’s Action

Take one prediction you already generate, churn risk is the usual candidate, and build the loop around it for a single cohort. Resolve it to one action, route that action to one owner, and set a 90-day outcome check. Do not scale it. Run it once, measure it, and let the result tell you whether the prediction was worth the screen space it has been occupying.

Choose the smallest version that still closes the loop: one prediction, one segment, one action, one owner, one date to check. The instinct will be to do more, to loop three predictions at once because you can see all three on the dashboard. Resist it. A single loop that actually completes teaches you more than three that stall halfway, and it gives you a clean result to carry into your next review: here is the prediction, here is what we did, here is what happened in 90 days. That sentence is worth more than any accuracy statistic, because it is a decision with a measured outcome attached, and decisions with outcomes are what a Director gets credit for. Once one loop runs cleanly, the second is a copy, not a new build.

P.S. The tell that you have a score without a loop is simple: ask who acted on last month’s highest-risk member, and what happened. If the answer is a shrug, the model isn’t your problem. The motion is. And the motion is the cheaper thing to fix, because you already own the prediction.

Next-Purchase Propensity predicts when a specific member is entering a buy decision, so the reorder nudge arrives while the rack is empty instead of on a fixed calendar date. It combines consumption cadence, depletion timing, and category affinity, and it works because triggered emails click near 5% versus roughly 1.5-2% for batch campaigns.

Consider what a calendar release actually assumes. It treats a few thousand members as if they share a single buying rhythm, and it picks a single Tuesday to interrupt them all at once. A handful will be ready to buy. Most will not. And a quiet, expensive slice will have run empty weeks ago and already refilled the rack somewhere else, in the exact gap your quarterly schedule left open.

That gap is where reorder revenue leaks. Not to a better wine or a lower price, but to a competitor whose message happened to arrive closer to the moment the member needed it. Timing, not quality, decided the sale. And timing is a signal you already have, sitting unused in your order history.

This is worth sitting with, because the usual response to soft reorder numbers is a discount, and a discount is the wrong tool for a timing problem. A member who ran out three weeks ago and already refilled does not need ten percent off; they needed to hear from you three weeks ago. A member who is fully stocked does not become more likely to buy because you shaved the price; you have simply trained your best buyers to wait for the markdown. When the real constraint is timing, price promotion spends margin to solve a problem it cannot reach. Predicting the window costs nothing but attention to data you already own, and it protects the price you worked to hold.

The Next-Purchase Window

Next-Purchase Propensity is the discipline of predicting when a specific member is entering a buy decision, and reaching them inside that window rather than on a broadcast date. Three signals carry it, and you collect all three today.

Consumption cadence. Every repeat buyer has an interval, the typical spacing between their orders, and it varies by format and price tier. A member who buys a case of everyday wine every eight weeks and a few reserve bottles twice a year has two distinct clocks. Cadence is the baseline: it tells you roughly how long a member’s purchase lasts them before the next one.

Depletion timing. Cadence only matters relative to where the member is inside it right now. A member two weeks past a purchase is not in a window; a member who is a week short of their usual reorder interval is. Depletion timing is the live position: it converts a static average into a this-week signal about who is approaching a decision and who just made one.

Category affinity. The third signal keeps the timing relevant. It reads what a member actually repurchases, the varietals, formats, and tiers that recur in their history, rather than what they clicked once and never bought. Affinity ensures the well-timed message is also the right message: the member entering a window for their regular Pinot hears about Pinot, not a blanket release of everything. It also protects you from the most common false signal in DTC, the browse that never becomes a buy. A member who clicked a reserve tier once but has only ever purchased everyday bottles is telling you where their curiosity is, not where their wallet is, and affinity keeps you from mistaking the first for the second.

Combined, these three answer a question a calendar can never answer: which members are in a buy window this week, and for what? That list is small, specific, and actionable, and it changes every week as members move through their intervals. Notice that all three signals are backward-looking and already in your possession. You are not buying new data or guessing at intent from a survey. You are reading the purchase history you already store and letting it tell you what it plainly knows: roughly when each member tends to buy, where they are in that rhythm now, and what they reliably reach for.

What Timing Produces

The counterintuitive part is that this is not more marketing. It is frequently less. A propensity-timed program often sends fewer total messages than a calendar program, because it stops interrupting members who are nowhere near a decision and concentrates on the ones who are.

The lift shows up in the numbers the platforms already publish. Triggered, behavior-based emails, the kind that fire when a member enters a buy window, typically earn click rates around 5 percent, against roughly 1.5 to 2 percent for one-off batch campaigns (Klaviyo, Email Benchmarks 2024; GetResponse, 2024). The gain comes from relevance and timing, not volume: the same offer, aimed at the moment it is useful, converts at a materially higher rate than the same offer broadcast to everyone on a fixed date. And because frequency holds or drops, you spend down less of your permission asset to get the lift.

There is a retention effect underneath the revenue one. A member who consistently hears from you right when they need to reorder learns that your messages are worth opening. A member who consistently hears from you two weeks late learns the opposite. Over a year, timing quietly trains your open rate in one direction or the other, and your open rate is the foundation every other campaign stands on. A well-timed program is not just selling more reorders; it is protecting the deliverability and attention that make the next release land at all.

The objection here is usually “we already send plenty of email.” That is exactly the point. Propensity timing is not permission to send more; it is a rule for sending the same volume to better-chosen recipients. In most programs it lets you retire a portion of the blanket calendar sends, the ones going to members nowhere near a decision, and reinvest that frequency into the members who are. The member with a full rack stops hearing from you about reorders they don’t need, which is its own form of respect, and the member about to run dry hears from you first. You are not adding to the noise; you are moving it to where it reads as service.

Putting It Into Practice

• For your top repeat buyers, calculate the median interval between their orders. That single number is a usable cadence baseline; you do not need a model to start.
• Flag members who are currently near or past that interval without a reorder. That is your live buy-window list this week.
• Match each flagged member to their most-repurchased category, and send the release or reorder nudge for that category, not a blanket announcement.
• Hold your calendar release as the control for one cycle, and compare conversion on the propensity-timed send against it. Let the numbers, not the theory, earn the change.

Two cautions keep this honest. First, cadence is a starting estimate, not a guarantee. A member’s interval shifts with seasons, gifting, and life, so treat the window as a reason to reach out, not a certainty to over-message against; if a flagged member doesn’t buy, that is information, not a failure. Second, resist the urge to widen the net. The value of a propensity list is its narrowness: twenty-five members genuinely near a decision will outperform a thousand-member blast, and the moment you loosen the criteria to feel productive, you are back to the calendar with extra steps. Precision is the product, and the discipline is holding the list small even when a bigger one feels busier.

This Week’s Action

Pull the 25 members who last ordered closest to one full cadence-interval ago. They are, statistically, your most likely buyers this week. Send them a category-matched message now and tag it. When you compare its conversion to your last calendar send, the difference is the size of the timing leak you have been paying for all year. Do this for three cycles before you judge it: one send can flatter or disappoint on luck, but three will show you whether the window is real. It almost always is, because you are no longer guessing when your members buy; you are reading the record of when they already have.

P.S. The fastest version of this needs no new tools: one saved segment of “members past their median reorder interval,” refreshed weekly, matched to category. It is the same custom-segment work you already do by hand for releases, pointed at timing instead of a date. The members it surfaces are the ones a competitor is otherwise happy to reach first.

Predictive Lifetime Value forecasts a member’s value band from their first 90 days instead of waiting 12-18 months for a trustworthy trailing LTV number. It reads order cadence, channel breadth, and engagement depth to sort new members into bands you can defend in a sentence, so you can aim hospitality and allocation at the predicted high band early, while roughly 40% of club members are still deciding whether to stay past their first year.

You already know your subscribers are not worth the same, and you already do the math to prove it. Most Directors at your tier rebuild lifetime value by hand every month in a spreadsheet from raw order data, because the platform’s LTV methodology doesn’t align with what you present to ownership. That instinct is correct. The problem isn’t that you calculate the value. It’s that you calculate it too late to act on it.

By the time a trailing LTV number is trustworthy, the member is twelve to eighteen months in, and their habits are set. Roughly 40 percent of club members cancel within their first year (Silicon Valley Bank, State of the US Wine Industry 2026), so the number that would let you allocate precisely often arrives after the member has already gone quiet. You end up spreading hospitality, allocation priority, and personal outreach evenly across a base you know is uneven, because the read you need lands after the moment that mattered.

The cost is not dramatic in any single case. It is a slow, distributed leak: a high-potential member who got a generic welcome and quietly never returned; a top-band member who would have responded to an early invitation that you instead spread evenly across everyone. Multiplied over a year, that spread alone becomes one of the larger uncaptured revenue lines in an otherwise mature DTC program. It hides in plain sight precisely because nothing about it looks like a mistake. Every member got something; no member got the wrong thing; the numbers simply came in lower than the base was capable of, and no single decision is there to point at.

The Early Value Signal

Predictive Lifetime Value advances the estimate. Instead of waiting for a member to accumulate spending history, you forecast their value band based on the behavior they show in their first 90 days. Three inputs, all of which you already collect, carry most of the predictive weight.

Order cadence. The interval between the first, second, and third purchases is one of the strongest early tells. A member who reorders within six weeks of a first purchase is on a different trajectory than one who takes five months, regardless of order size. Cadence tends to be stable: the rhythm a member sets early is close to the rhythm they keep. You are reading the tempo, not the total.

Channel breadth. Members who engage across surfaces, a tasting room visit plus email clicks plus a web order, resolve to a higher band than single-lane buyers of the same early spend. Breadth signals that the relationship has more than one point of contact, and multi-contact members are structurally harder to lose and easier to grow. A member who only ever opens email is more fragile than their order total suggests.

Engagement depth. Not opens for their own sake, opens and clicks that sit in front of a purchase. The question is whether a member’s attention converts to action. A member who opens every email and never buys is a different kind of member than one who opens selectively and buys most of the time. Depth separates genuine intent from list noise.

Read together, these three inputs place a new member into a value band inside their first quarter: a range you can state plainly, defend in a sentence, and revise as more behavior arrives. A band, not a false-precision point estimate. “This cohort is tracking toward the top value tier” is a claim you can carry into a review and stand behind. A single black-box score to three decimals is not.

What Early Banding Produces

The value of predicting early is entirely about what you do with the extra time. When you can identify a probable high-value member in month two rather than month twelve, you can point your scarcest resources, allocation priority, a personal note from the club team, an early invitation, at the members most likely to compound, while the attention still shapes the relationship.

Directors who allocate based on predicted value early may see more of their top band reach the top of its range because recognition arrives during the formative window rather than after it. The mechanism is not a bigger onboarding budget; it is the same budget, aimed. Personalization that reflects a member’s real standing has a documented revenue effect in the 5 to 15 percent range (McKinsey); early value banding is what lets you apply that lift to the members where it pays back most.

There is a compounding effect worth naming. A member who reaches your top band is not simply worth more in isolation. High-value members tend to refer; they anchor the demand that makes your allocations feel scarce, and they are the most forgiving of an occasional misstep. Moving even a few points of each new cohort into that band changes the shape of the base you carry into next year, and the year after. Early banding is about influencing that shape while it is still forming, rather than documenting it once it has already set. Reporting tells you what your base became; prediction gives you a hand in what it becomes.

None of this requires you to be right about every member. Prediction at this stage is about being right on average, early. If your predicted high band converts to an actual high value more often than an even guess would, you are already allocating better than a flat spread, and the rule sharpens every quarter you check it against what actually happened. The bar is not perfection; it is better than treating everyone as identical, which is the standard you are actually competing with. That is a low bar to clear and a high one to ignore.

There is a second benefit specific to your quarterly review. A value-band forecast, tied to the three inputs that produced it, is a defensible artifact. When ownership asks why the club team spent time on a particular cohort, “they scored in our top predicted-value band on cadence, breadth, and depth” is an answer that holds up in sixty seconds. You are no longer defending an even spread; you are defending a decision.

Putting It Into Practice

You can build the first version of this without hiring a data scientist and without a new platform.

• Pull your members from the last two years who reached your top value tier. Look only at their first-90-day behavior: cadence, breadth, depth. Write down what the high-value members had in common at day 90.
• Do the same for members who stalled or churned early. The contrast gives you a rough banding rule you can apply to new members by hand this quarter.
• Apply the rule to members who are currently within their first 90 days. Tag each into a predicted band.
• Route one concrete action to the predicted-high band now: an allocation priority, an early invitation, a personal outreach. Change nothing else, so the effect is attributable.

The most common objection is that a winery your size does not have enough data to make predictions. The opposite is usually true. You do not need a large model; you need the last two years of your own members, which you already hold. The banding rule that comes out of that history is specific to your brand, your price points, and your buying patterns, which makes it more useful than any borrowed benchmark, not less. Start manual, start small, and let the rule earn trust before you automate it. The first version can live in the same spreadsheet where you already rebuild LTV each month; you are simply moving the estimate earlier in the member’s life, where it can still do work.

This Month’s Action

Take the ten members who joined in the last 60 days. Score each on the three inputs and assign a predicted value band. Then compare your banding to how those members are actually being treated by your current onboarding. If your highest-predicted members are receiving the identical sequence as everyone else, you have found the gap, and closing it costs attention, not money.

P.S. The reason for banding rather than scoring is defensibility. A band with three named inputs survives the question “how do you know?”; a single decimal from a model you didn’t build does not. You already rebuilt LTV by hand because you need to trust the number. Predictive banding gives you a number you can trust earlier, which is the only version that lets you act while acting still matters.