Predictive Marketing: Easy Ways Every Marketer Can Use Customer Analytics and Big Data

1. 1Define trust as an observable behavior chain: return, spend more, advocate. The source's corner-store logic is that being known makes buying pleasant and raises the likelihood the customer will return, spend more, and develop brand loyalty and trust. Pick one metric per link (repeat purchase, revenue per customer and its trend, referral or advocacy) so you can tell whether trust is growing.
2. 2Build the 'corner store memory' first. Consolidate a complete, accurate profile per customer from web, email, social, mobile, and store data: name, what they bought, how long they have been a customer, and preferences. Start with digital sources (email and website) and add store transaction systems after you prove success. Without this profile, personalization is guesswork and feels generic.
3. 3Apply 'give to get' at every lifecycle stage. Before each message or offer, write down the specific value the customer receives (relevant recommendation, convenience, replenishment reminder). If you cannot name the value, do not send it. Customers buy more readily from brands they trust, and the way to earn that trust is to deliver an experience of value.
4. 4Start retention on day one. Treat the first transaction as the beginning of the relationship. Launch postpurchase campaigns, then replenishment, repeat-purchase, new-product, and customer appreciation programs, so engagement does not stop when a prospect converts into a buyer.
5. 5Differentiate by customer value. Spend more to appreciate and retain high-value customers. Upsell medium-value customers to migrate them upward. Reduce the cost of servicing low-value or unprofitable customers. Loyalty investment follows lifetime value, not equal treatment.
6. 6Measure retention in dollars, not heads. Track dollar-value retention so you do not retain customers while losing money. Separate churn by customer value, because losing an unprofitable customer is not nearly as bad as losing a best customer.
7. 7Prevent defection before reactivating. Set up proactive retention for at-risk valuable customers, because preventing departure is easier, cheaper, and more effective than winning back someone who has already stopped shopping.
8. 8Run a delightful-versus-creepy check on each data use. Consumers share preference information in exchange for apparent benefits such as convenience, but they feel differently about different types of information and different contexts. Review each personalization for the benefit the customer sees and the context it appears in before launch.

The source's anchor example is the corner store. The salesperson knows your name, what you bought, how long you have been a customer, and other things about your personality and behavior. That makes buying pleasant and raises the likelihood that you return, spend more, and develop brand loyalty and trust. Supermarkets and online shops lost that memory, and predictive marketing aims to restore it across websites, social media, mobile apps, and store visits. Zappos is the source's modern parallel: it knows your name, what shoes you like, and where you live. To apply this, consolidate that profile, then send a postpurchase or replenishment message that uses what you know and gives the customer something useful, such as a relevant recommendation. Concentrate appreciation spending on your highest-value customers, and check each data use against the convenience-versus-creepy test. The test of success is the chain the source names: return, spend more, loyalty.

• If a campaign cannot state the value the customer receives, then do not send it, because trust comes from delivering value first ('give to get').
• If a customer is high lifetime value, then increase appreciation and retention spend. If low value or unprofitable, then reduce servicing cost rather than matching the investment.
• If you must choose between a churn-reduction win and a dollar-value result, then judge by dollar-value retention, because you can retain customers and still lose money.
• If a customer is at risk but still active, then intervene proactively. Do not wait for reactivation, which is costlier and less effective.
• If a personalization uses a data type or context the customer would not expect, then drop it unless it delivers an apparent benefit such as convenience. When unsure, apply common sense: delightful or creepy.

• Treating the sale as the finish line. The tell: no postpurchase, replenishment, or repeat-purchase programs, and engagement stops at conversion.
• Measuring retention by customer count. The tell: retention looks healthy while revenue per customer or total dollars stagnate or fall.
• Treating all churn equally. The tell: equal win-back effort goes to unprofitable customers and best customers.
• Mass-marketing from a product-and-promotion mindset. The tell: the same one-size-fits-all promotions go to everyone, with no use of purchase history or preferences.
• Using customer data in ways that feel invasive. The tell: customers react to a message with discomfort instead of convenience, or the data type or context was never checked for acceptability.

• ▸Customer segment (value tier): high, medium, or low lifetime value, and how it is defined.
• ▸Profile data available: list what you know per customer (name, purchases, tenure, preferences) and the sources.
• ▸Value given: the specific benefit the customer receives from this touchpoint.
• ▸Lifecycle stage and program: acquisition, postpurchase, replenishment, repeat purchase, new product, appreciation, or retention.
• ▸Investment stance: spend more to retain, upsell to migrate, or reduce service cost.
• ▸Trust-chain metric: return rate, revenue per customer and trend, or advocacy, with the current baseline.
• ▸Dollar-value retention: dollars retained from this segment versus dollars at risk.
• ▸Churn-risk action: the proactive step for at-risk valuable customers before they leave.
• ▸Delightful-vs-creepy check: data type used, the customer-visible benefit, the context, and a pass or fail verdict.

1. 1Define the behavior you are predicting before you touch a model. The source frames it as 'the likelihood of a certain type of future behavior': buy, respond to an email promotion, or complete a transaction. Write one sentence naming the action and the window, e.g. 'purchase in the next period' or 'respond to this email'. Each action gets its own model.
2. 2Name the decision the score will change. The source gives three: how sales and customer success teams spend their time (B2B lead or customer scoring), how deep a discount goes, and how often you email (B2C). If no spend, discount, or frequency decision changes with the score, do not build it.
3. 3Get historical outcome data (did the customer buy or not) to use as a training set. Propensity models learn by comparing predictions with outcomes, so they need a training period and a testing period before you rely on them. A historical data set shortens that period. Keep a test slice the model never trained on.
4. 4Score every customer or prospect, then group the scores into deciles (top 10%, next 10%, down to the bottom 10%) instead of acting on individual scores. Read the average expected behavior per decile. The source's lifetime value example puts the top decile at about $1,000 and the bottom at $5. Your likelihood deciles should show a similar steep gradient.
5. 5Check the gradient before launching anything. If the top and bottom deciles behave about the same, the model is not ranking customers and you should not spend against it.
6. 6Assign a different action to each decile band. For remarketing (abandoned cart, browse, search), send likely returners a simple reminder, and reserve an extra incentive for those unlikely to come back. For B2B, point salespeople and customer success at the top-ranked accounts first.
7. 7Run the treatment against a holdout and let the model keep learning from the new outcomes. Because the model adjusts as results come in, feed the buy/no-buy results back so accuracy improves over time.
8. 8Start with one program. Mavi's head of CRM advises: 'Start small and pick just one program and build on that success.' Prove the ranking on one use, such as discount level or email frequency, before extending it.

Take the abandoned-cart campaign, the use the source names for splitting 'likely to come back' from 'unlikely to come back'. Without a score, every abandoner gets the same message, and likely buyers may receive a discount they did not need. With a likelihood-to-buy model, you score each abandoner and place them in deciles. The top deciles get a simple reminder with no discount, because they are likely to return anyway. The bottom deciles get an additional incentive, because a reminder alone is unlikely to move them. The source's decile logic applies here: the model is only useful if the top decile's average expected behavior is far from the bottom's, as in the $1,000 versus $5 lifetime value example. Mavi shows the payoff from precise targeting. Its persona clusters (a different technique from propensity) reactivated 20 percent of lapsed customers, and 80+ programs added 7 percentage points to revenue. A likelihood score adds a ranking so you can decide which customers get which offer.

• If a customer sits in a top likelihood decile, then send a simple reminder and withhold the discount. If they sit in a low decile, then add an incentive.
• If the top and bottom deciles show similar average outcomes, then the model is not ranking customers. Stop and retrain with more history, and do not move spend.
• If you lack a training and testing period or a historical outcome set, then do not rely on the predictions yet. Load historical data or run a short test window first.
• If the audience is B2B, then use the score to order sales and customer success outreach. If it is B2C, then use it to set discount depth and email frequency.
• If you are new to this, then pick one program and prove it before building more.

• Acting on individual scores as if they were precise. The tell: debates about why one customer scored 0.43 versus 0.45. Group into deciles and act on the band average.
• Trusting the model before it is trained and tested. The tell: live spend shifts in week one, with no test slice and no historical data set.
• Discounting everyone equally. The tell: high-likelihood buyers get the same incentive as unlikely ones, so margin is given away on sales that were coming anyway.
• Optimizing for purchase likelihood alone and ignoring customer value. The tell: top-ranked customers are low-value. The source separates value-based marketing (predicted lifetime value) from likelihood, so check both.
• Building a score with no decision attached. The tell: a dashboard of scores while sales, discount, and email-frequency rules stay unchanged.

• ▸Behavior and window: write the one action you are predicting and the time period (e.g. purchase, email response).
• ▸Decision it changes: pick sales/CS prioritization, discount depth, or email frequency.
• ▸Training data source: name the historical set with buy/no-buy outcomes and its date range.
• ▸Test period or holdout: state the slice the model never trained on and when you read results.
• ▸Decile outcome table: record the average expected behavior for each decile, top 10% through bottom 10%.
• ▸Gradient check: note the top-versus-bottom gap and whether it is wide enough to proceed (yes/no).
• ▸Action by decile band: write the treatment per band (e.g. reminder only, reminder plus incentive, reduced frequency).
• ▸Control group: define the holdout that receives the old treatment for comparison.
• ▸Feedback loop: say how new buy/no-buy outcomes return to the model and how often.

1. 1Fix the unit first: one record per real person. The source says about 80 percent of predictive work is collecting and cleaning customer data. Merge duplicates and connect online and offline records (fuzzy matching helps) before you score anything. Duplicates split one customer's value across two records and push a high-value customer into a lower tier.
2. 2Define the value metric as customer lifetime value, with profitability as the stated aim. Write down which revenue, cost and time horizon feed it, so every team ranks customers the same way.
3. 3Rank every customer by lifetime value and cut tiers from the ranked list. The definition's illustrative split is top 10%, next 60%, bottom 30%. Treat those percentages as a starting point, not a rule, and record the cutoffs you choose.
4. 4Separate new customers from tenured ones before you apply tiers. The source says to take different actions and send different messages to brand-new customers than to those who have been with you a while. Keep a tenure flag alongside the tier.
5. 5Assign one job per tier. High-value: spend more to appreciate and retain. Medium-value: upsell to migrate them into higher segments. Low-value or unprofitable: reduce the cost to service them.
6. 6Set budgets per person, by tier, not per channel. The source's approach is to allocate dollars to the right people, rather than to the right products or channels. Set a per-customer spend ceiling for each tier and tie it to the tier's job.
7. 7Wire each tier to an action that a customer can actually experience (a different offer, service level or message). The source's test is whether analytics changes what customers experience. If a tier changes nothing at the customer level, the setup is wrong.
8. 8Re-score on a schedule and track migration. The medium tier is meant to move upward, so measure how many customers change tier and in which direction, and treat that as your outcome metric.

The source supplies no numbers, so this example is illustrative and uses the definition's 10/60/30 split. Take a 10,000-customer file. First merge duplicates and online/offline records so that you have 10,000 real people. Rank by lifetime value. The top 1,000 (10%) are high-value: they get appreciation and retention spend. The next 6,000 (60%) are medium-value: they get upsell programs aimed at migrating them into the top tier. The bottom 3,000 (30%) are low-value or unprofitable: you cut the cost of servicing them. You also flag brand-new customers, because their messaging differs from that of long-standing customers. The source's Best Buy anecdote is the check on the whole exercise. Best Buy had consultants, tools and talk of segmentation, but 'when you walked into a store, none of that had any impact at the customer level.' So you confirm that a top-tier customer is handled visibly differently from a bottom-tier one.

• If a customer's record may be duplicated or unlinked across online and offline sources, then fix identity resolution before assigning a tier.
• If a customer is brand-new, then apply new-customer actions and messages and put them in a tier only with the tenure flag attached.
• If a customer is high-value, then spend more to appreciate and retain them. If medium-value, then upsell to move them up. If low-value or unprofitable, then reduce service costs.
• If a tier produces no different customer-facing treatment, then the segmentation has failed the last-mile test, so change the treatment before you refine the model.
• If you are deciding where to put marginal budget, then allocate to people by tier before you allocate to channels or products.

• Tiering duplicated or fragmented records. The tell: a known loyal customer sits in a mid or low tier because their purchases are split across several profiles.
• Applying one program to everyone, or tiers that exist only in a slide. The tell: a high-value and a low-value customer receive the same offer, service and message, which is the Best Buy last-mile problem.
• Ignoring tenure and life cycle stage. The tell: brand-new customers get the same messages as long-time customers and are tiered on thin history.
• Reading 'reduce costs' as 'abandon.' The tell: low-value customers are cut from all contact with no plan, when the guidance is to lower the cost of servicing them.
• Treating the 10/60/30 cutoffs as fixed truth. The tell: nobody can explain why the cutoffs sit where they do, and you never check whether medium customers are migrating upward.

• ▸Value metric definition: state what feeds lifetime value (revenue, costs, horizon), with profitability as the goal.
• ▸Data readiness check: record that duplicates are merged and online and offline records linked, and note the date.
• ▸Tier cutoffs: write the percentage split (e.g., top 10% / next 60% / bottom 30%) and the rationale for it.
• ▸Tenure flag rule: define who counts as brand-new and which messages and actions differ for them.
• ▸High-value tier action and per-customer spend ceiling: describe the appreciation and retention treatment and the cap.
• ▸Medium-value tier action and per-customer spend ceiling: describe the upsell program that moves customers up, and the cap.
• ▸Low-value tier action: list the specific ways you reduce the cost to service these customers.
• ▸Customer-visible difference: describe what a customer in each tier experiences differently, to pass the last-mile test.
• ▸Review cadence and migration metric: record when you re-score and how many customers moved up or down a tier.

1. 1Fix the stage list before you touch data: prospect, new, repeat/active, at-risk/inactive, lapsed. The book frames the journey as acquisition, growth, and retention. Map your five stages onto those three objectives: convert (prospect, new), grow value (new to repeat/active), retain (at-risk, lapsed). Those correspond to the book's Plays Seven, Eight, and Nine.
2. 2Build the complete customer profile first. The book's rule is to know your customers before predicting anything, and stage assignment is only as good as the linked 360 profile. If purchases, engagement, and identity records are not joined to one customer, the stage is wrong before the campaign starts.
3. 3Cleanse before you classify. The book says data cleansing is 'the most important and most ignored stage', and that about 80 percent of predictive work goes to collecting and cleaning data. Fix inconsistent fields (for example, 'CA' versus California) and check missing values. Otherwise customers land in the wrong stage.
4. 4Remove outliers before you set stage thresholds. One customer who buys 50 televisions for $50,000 when the average is $500 skews any average you compute. Do not let that record set your definition of 'active'.
5. 5Transform skewed behavior variables before drawing stage boundaries. The book's example is a log of order count: 1 to 10 orders is the same distance as 10 to 100. Use a ratio where it carries more signal than a raw number, such as share of revenue returned. Then set your repeat and at-risk cut-offs on the transformed scale, with values you choose and validate yourself, since the passages give none.
6. 6Assign one stage per customer and attach one objective to it. Prospect and new customers get a conversion objective, active customers a value-growth objective, at-risk and lapsed customers a retention objective. Write the objective in one sentence per stage.
7. 7Apply 'give to get' at every stage. Each message must deliver value before it asks for something, because customers buy when they trust you and trust comes from an experience of value. Check each stage's campaign against the question 'what does the customer receive?'
8. 8Layer value on top of stage, using the book's value-based marketing logic. Spend more to appreciate and retain high-value customers, upsell medium-value customers, and cut service cost for low-value ones. Stage tells you what to attempt, and value tells you how much to spend.
9. 9Evaluate campaign results within each stage, never pooled across stages. The stage moderates the outcome, so a lift that exists in one stage can disappear or reverse in another. Report results per stage.

The passages give the stage principle but no single numeric life cycle case, so this illustration combines the book's own data-prep example with its stage logic. An electronics retailer averages $500 per customer. One buyer purchases 50 televisions for $50,000. If you compute 'typical active customer' spend or order behavior from raw averages, that buyer drags the benchmark upward and misplaces ordinary customers in the stages around it. You flag the record as an outlier, as the book advises, and exclude it when setting thresholds. Order counts are then log-transformed so the gap from 1 to 10 orders carries the same weight as the gap from 10 to 100. You use that scale to draw the new, repeat/active, and at-risk boundaries (cut-offs are yours to choose; the book does not specify them). A first-time buyer lands in 'new' and gets a trust-building, give-to-get message. A buyer past the at-risk boundary gets a retention message, and the bigger budget goes to those with high lifetime value. You then read each stage's response separately.

• If a customer's records are not linked into one profile, then do not assign a stage yet; fix identity resolution first.
• If a single extreme record (such as 50 TVs at $50,000 against a $500 average) is pulling your stage thresholds, then treat it as an outlier and exclude it before setting boundaries.
• If the customer is a prospect or new, then set the objective to convert and lead with a delivered experience of value, not a hard ask.
• If the customer is at-risk or lapsed and high lifetime value, then spend more to retain; if low value, then reduce the cost of serving them.
• If a campaign result looks strong in aggregate, then split it by stage before trusting it, because stage moderates the effect.

• Using one blended campaign for all customers. The tell is the same offer going to a first-time buyer and a long-lapsed one, with flat or mixed overall response.
• Setting stage cut-offs on raw averages polluted by outliers. The tell is a few giant buyers shifting the 'active' benchmark, so many ordinary customers read as inactive.
• Dirty or unlinked data driving stage assignment. The tell is customers with inconsistent fields or split profiles showing up in the wrong stage, or in two stages at once.
• Treating stage as a label and not an objective. The tell is a stage list with no distinct goal or message per stage.
• Asking before giving. The tell is acquisition or win-back messages that push an offer with no value delivered first, which ignores the trust principle.

• ▸Stage name: prospect, new, repeat/active, at-risk/inactive, or lapsed.
• ▸Entry rule: the behavior threshold that places a customer here (you set the value; validate it on cleansed, outlier-free, transformed data).
• ▸Data check: confirm profiles are linked, cleansed, and outliers removed for this stage.
• ▸Objective: convert, grow value, or retain, in one sentence.
• ▸Give-to-get offer: the specific value delivered to the customer before any ask.
• ▸Value tier treatment: spend level for high, medium, and low lifetime value customers in this stage.
• ▸Campaign and channel: the program and where it reaches the customer.
• ▸Success metric: the outcome measured within this stage only.
• ▸Result and next change: stage-level result and the one adjustment you make next.

1. 1Define the metric first: measure retention in dollars, not head count. The source warns that otherwise 'you could be retaining customers, but losing money anyway.' Report two figures side by side: customer-count retention and dollar-value retention. A gap between them means value is migrating.
2. 2Split dollar loss into two buckets: churn (customers who stopped buying) and value migration (retained customers whose spend is declining). Track both, because a retention rate based on active-versus-inactive hides the second bucket entirely.
3. 3Weight every churn event by customer value. Use predicted or historical lifetime value to tier customers into unprofitable, medium-value and high-value, as the portfolio chapter advises. Losing an unprofitable customer 'is not nearly as bad as losing one of your best.' Set retention effort and budget by tier.
4. 4Start retention on the day of acquisition. The source says the secret to retaining a customer is to start trying to keep her the day you acquire her. Build post-purchase, replenishment and repeat-purchase programs for new customers, and treat new customers differently from long-tenured ones.
5. 5Build a complete customer profile before building a churn program. Combine purchases, newsletter signups, product registrations and service contacts into one record per customer. Bosch and a New York boutique fashion brand both found this valuable even without advanced analytics.
6. 6Move from untargeted to targeted churn management. Untargeted programs apply equally to all customers. Targeted programs use a predictive model to flag customers at risk of leaving so you can act proactively. Earthlink identified dissatisfied customers about to churn and contacted them.
7. 7Prefer prevention to reactivation. Preventing a departure is described as easier, cheaper and more effective than winning back a customer who has already stopped shopping. Run proactive retention first and keep reactivation campaigns as a separate, lower-priority layer.
8. 8Prove each program with a test. As with the decile response model, run a/b tests within the at-risk groups to measure lift, and spend on retention offers only where the incremental profit justifies the cost.

Earthlink, cited in the source as an early adopter of predictive analytics, used it to identify dissatisfied customers who were about to churn. It then took proactive steps to contact and retain them, and reduced churn by 30 percent. Mechanically, the work has three parts: score customers for likelihood of leaving, select the flagged group, and intervene before departure rather than after. The source adds two supporting practices. A New York boutique fashion brand merged all customer data into a single profile, so service staff could see whether a complaining customer was high-value and respond accordingly. High performers in Mark Jeffrey's research also spent more on retention marketing (14 percent versus 11 percent) and less on demand generation (48 percent versus 58 percent). Combining these gives the workflow: unify the profile, rank at-risk customers by value, intervene early with the most valuable, and report the outcome in dollars.

• If a customer is flagged as likely to leave and is high-value, then intervene proactively and personally. If the customer is unprofitable, then do not spend retention budget, because losing that customer is far less costly.
• If customer-count retention is stable but revenue per retained customer is falling, then you have value migration. Treat it as a retention problem and build engagement and growth programs for that group.
• If a customer is still active, then prioritize prevention. Use reactivation campaigns only for customers who have already lapsed, since prevention is easier and cheaper.
• If the customer is new, then begin retention immediately with post-purchase, replenishment and repeat-purchase programs. Do not wait for signs of churn.
• If you can only afford an untargeted churn program, then run it, but plan to move to model-flagged targeting. Use a/b tests to confirm incremental lift before scaling the spend.

• Reporting only customer-count retention. The tell: the retention rate looks healthy while revenue from the retained base declines. The fix is to measure dollar-value retention.
• Treating all churn as equal. The tell: retention offers go to low-value or unprofitable customers at the same intensity as to top customers, and discount cost exceeds the value kept.
• Waiting until a customer lapses. The tell: most retention budget sits in win-back campaigns, and engagement with new buyers stops after the first transaction.
• Running retention without a unified customer profile. The tell: service and marketing staff cannot tell whether a complaining or departing customer is high-value, and data requests need an outside provider to build a new report.
• Judging campaigns by activity rather than lift. The tell: you can report sends and responses but cannot show incremental retained dollars against a control group.

• ▸Period and customer cohort: state the time window and which customers are included.
• ▸Customer-count retention: percentage of the cohort still buying at period end.
• ▸Dollar-value retention: share of the cohort's starting spend that remains at period end.
• ▸Churned dollars: spend lost from customers who stopped buying entirely.
• ▸Value-migration dollars: spend lost from retained customers who reduced spend.
• ▸Value tier of lost customers: split the loss into unprofitable, medium-value and high-value.
• ▸At-risk list: customers flagged as likely to leave, ranked by value, with the date flagged.
• ▸Proactive intervention and owner: the action taken, who is responsible, and the date contacted.
• ▸Test design: the control group and the measured lift in retained dollars for each intervention.
• ▸Reactivation backlog: customers already lapsed, kept separate from the proactive list.

1. 1Fix the unit first. The source treats the customer as the unit of value: 'Customers are the unit of value for any company.' Compute CLV per individual customer, not per product, channel or campaign. Company valuation is the sum of those individual values.
2. 2Choose profit over revenue as the target. The source frames the goal as maximizing 'the lifetime value, or profitability, of each and every customer.' Define the value as expected future profit: net of cost to serve and market, and discounted for time. This is what lets you identify 'unprofitable' customers at all.
3. 3Build complete customer profiles before modeling. The source says roughly 80 percent of predictive work is collecting and cleaning data, and that customer data is typically scattered, duplicated and error-filled. Consolidate it in one place. Use fuzzy matching to resolve identities across online and offline. Without this step, one person looks like several low-value customers.
4. 4Predict value forward-looking, not just historical. Chapter 8 is titled 'Predict Customer Value.' Score each customer on expected future value over the relationship, not only past spend, so that brand-new customers can be placed alongside long-tenured ones.
5. 5Segment the scored base into at least three tiers: high-value, medium-value, low or unprofitable. The source states every business has all three. Set tier cut-offs so each tier is a distinct population you will treat differently.
6. 6Assign a different play to each tier. High value: spend more to appreciate and retain. Medium value: upsell to migrate customers into higher-value segments. Low or unprofitable: reduce the cost to service them. This is value-based marketing.
7. 7Manage by life-cycle stage inside each tier. Treat new customers differently from those who have been buying a while. At every stage apply 'give to get': deliver an experience of value first, because trust drives purchase.
8. 8Allocate budget to people, not channels. Set spend by customer value and expected response. Shift effort toward retention and reactivation, which the source says drive growth more cost-effectively than acquisition-heavy mixes. Re-score periodically and review the tier mix as you would a portfolio.

The excerpt gives no company-level CLV figures, so this walk-through applies the source's value-based marketing logic to three hypothetical customers (the tiers are the source's; the customers are illustrative). Customer A is a long-tenured, high predicted-value buyer. Under the source's rule, you spend more to appreciate and retain A: earlier access, richer service, a personal touch. Customer B is mid-value with room to grow. You upsell B with recommendations and offers aimed at moving B into A's segment. Customer C is low-value and costs more to serve than C returns. You cut service and promotion cost for C, for example by moving C to lower-cost, lighter-touch channels, rather than spending acquisition-level dollars to retain C. Across all three you apply 'give to get': each receives something of value before you ask for more. The budget outcome mirrors the source's claim that the best way to maximize company value is to maximize each customer's lifetime value, with differentiated spend rather than an equal split.

• If a customer's predicted lifetime value is high, then increase spend on appreciation and retention rather than on acquiring lookalikes.
• If a customer is medium-value, then use upsell and cross-sell to migrate them upward, and judge the program by tier migration.
• If a customer's expected profit is low or negative, then cut the cost to serve them rather than adding promotional spend.
• If a customer is brand-new, then treat them as a different life-cycle stage from established buyers: lead with an experience of value (give to get) and re-score as behavior accumulates.
• If your budget is allocated by channel or product, then re-cut it by customer value, and favor retention and reactivation over pure acquisition where the data supports it.

• Measuring revenue instead of profit, so heavy-spending but costly customers land in the high-value tier. Tell: your top tier includes customers whose service or discount cost is rarely checked.
• Computing CLV on fragmented data. Duplicated identities split one valuable customer into several small ones. Tell: duplicate or mismatched profiles, and online and offline purchases that never appear under one ID.
• Using one message and budget for every customer. Tell: identical offers go to high-, medium- and low-value tiers, and spend is divided by channel, not by customer value.
• Over-weighting acquisition. Tell: nearly all budget goes to new customers while retention and reactivation of existing ones get little, even though the source calls these more cost-effective.
• Taking before giving. Tell: campaigns push the ask (sell, upsell) without first delivering value, which undercuts the trust that drives purchase.

• ▸Customer unit and ID rule: define the customer and how identities are matched across online and offline.
• ▸Value definition: state profit (not revenue), the costs netted out, and the time horizon and discounting used.
• ▸Data sources and cleanliness check: list the data merged into one profile and the duplicate/error rate you found.
• ▸Tier cut-offs: the predicted-value thresholds separating high, medium and low/unprofitable.
• ▸High-value play: the actions and extra spend to appreciate and retain this tier.
• ▸Medium-value play: the upsell actions intended to migrate customers upward, plus the migration target.
• ▸Low-value play: the specific cost-to-serve reductions.
• ▸Life-cycle 'give' step: the experience of value delivered to new customers before any ask.
• ▸Budget split and review date: spend by tier (retention/reactivation vs. acquisition) and when you will re-score.

1. 1State the premise to your team: customers are the unit of value, so enterprise value is the sum of individual customer lifetime values (customer equity). Maximizing the whole means maximizing each customer, not maximizing sales, channels, or products.
2. 2Build the single customer view first. Consolidate customer data into one place, clean errors and duplicates, and use fuzzy matching to resolve identities across online and offline. Budget for this: the book says data preparation can be 80 to 95 percent of the work, and without it any value estimate is wrong.
3. 3Make the profiles accessible to customer-facing staff and everyday marketers, not only analysts. The book calls this a great first step toward better experiences for every customer.
4. 4Segment the customer base into unprofitable, low-, medium-, and high-value customers using predicted lifetime value, and treat the base as a stock portfolio. Customers' values rise and fall over time, so re-score them on a schedule rather than once.
5. 5Assign a different strategy to each tier (value-based marketing). Spend more to appreciate and retain high-value customers. Upsell medium-value customers to migrate them into higher segments. Cut cost-to-serve for low-value or unprofitable customers.
6. 6Differentiate by life-cycle stage as well as value. Brand-new customers get different actions and messages than long-standing ones. Apply the principle 'give to get': deliver an experience of value to earn the trust that produces customer value.
7. 7Allocate budget to people, not to products or channels. Shift time and money from acquisition-only thinking toward retention and reactivation, which the book says grows the business more cost-effectively.
8. 8Track the portfolio with the measures the book names: lifetime value and share of wallet (rather than raw sales), plus relevance (rather than reach) for communications. Judge each program by whether it moves customer equity: more profitable customers acquired, each customer's value grown, relationships retained longer.

The book's own data-quality example shows how a bad customer view corrupts customer equity. Suppose you see only a person's in-store purchases, but that person makes 90 percent of purchases online. Your records label this customer unprofitable when they may be a VIP, so you cut spending on them and erode equity. The reverse also happens: someone who browses the website often but always buys in the store looks like a 'low value' website customer. A third case is a customer who spends a lot and buys often but returns items just as frequently or calls support constantly; revenue alone overstates their value. Fix the identity resolution, add online, offline, return, and service data to one profile, then recompute lifetime value. Now the VIP gets retention spend, the store-buyer is correctly valued, and the heavy returner is serviced at lower cost. The portfolio's total value rises because each customer's value is managed on accurate data.

• If a customer's profile does not combine online, offline, returns, and service data, then do not assign a value tier yet; fix identity resolution first.
• If a customer is high lifetime value, then spend more to appreciate and retain them; if medium, then upsell to migrate them up; if low or unprofitable, then reduce cost to service.
• If you are choosing between acquisition and retention/reactivation spend, then check whether retention and reactivation can deliver growth more cost-effectively before adding acquisition budget.
• If a customer is brand-new versus long-standing, then use different actions and messages; apply 'give to get' by delivering value before asking for it.
• If a metric rewards sales volume or channel reach, then replace or supplement it with customer lifetime value, share of wallet, and relevance.

• Valuing customers on partial data (one channel only). Tell: known loyal omnichannel buyers show up as low value, or high-spend customers with heavy returns or support calls show up as top tier.
• Treating enterprise value as a product or channel outcome. Tell: budgets are allocated by channel and product ROI, and no one can name the lifetime value of the top 10 percent of customers.
• Spending the same on every customer. Tell: identical offers and messages go to new, mid-value, high-value, and unprofitable customers.
• Over-focusing on acquisition. Tell: new-customer counts rise while retention and reactivation programs are underfunded or absent.
• Setting valuations once and never revisiting. Tell: segments are static even though customer values rise and fall over time, so high-value customers are not recognized when they start to decline.

• ▸Single customer view status: note whether online, offline, returns, and service data are merged and de-duplicated (yes/no, gaps).
• ▸Value tier definitions: write the lifetime-value cutoffs for unprofitable, low, medium, and high customers.
• ▸Customers and share of base per tier: record counts and percent of customers in each tier.
• ▸Tier strategy: for each tier, write the action (appreciate/retain, upsell/migrate, reduce cost to serve).
• ▸Life-cycle messaging: note what new customers receive versus established customers, and the 'give' (value delivered) in each.
• ▸Budget split: list spend by tier and by acquisition, retention, and reactivation.
• ▸Portfolio metrics: record lifetime value, share of wallet, and relevance measures, with current values.
• ▸Re-scoring cadence: set how often values and tiers are recalculated and who owns it.
• ▸Next review decision: write which tier shifts or budget reallocations the data supports.

1. 1Inventory every silo before you touch a tool. The source lists the typical ones: online transactions, store transactions, phone transactions, web behavior, email behavior, social behavior, and service center interactions. Write one row per silo with its owner, its customer identifier (email, address, loyalty ID, etc.), and how often it updates.
2. 2Pick the route by budget and team, not by preference. The source gives three: build it yourself on a predictive analytics workbench (millions of dollars plus an in-house team), outsource to a marketing service provider (annual contracts probably start at a quarter million dollars, typically Fortune 500), or buy a predictive marketing cloud (about $50,000 a year at the time of writing). With a small marketing team and no data scientists, the cloud route is the one the source's examples use.
3. 3Standardize every channel into one data model. Automated solutions use standard data models so online, store, and phone records land in the same fields. Do this before matching. Matching across inconsistent formats is where integration stalls, and the source notes it once took months or years.
4. 4Clean mailing addresses. Code to USPS standards (complete, with a ZIP+4 append), run NCOA to catch people who moved, require CASS certification, and run DPV to confirm each address is an active delivery point. Append the residence/business flag and the region or longitude/latitude where relevant. The payoff: address validation can cut mailing costs by as much as 80 percent.
5. 5Clean email addresses. Run syntax correction (gmai1.com becomes gmail.com; gmail,com becomes gmail.com) and a mail test that pings the domain. This protects reach and your sender reputation.
6. 6Link and deduplicate into one profile per person. Merge records so that online, store, and phone purchases, plus web, email, social, and service behavior, attach to the same customer. Variant spellings and formats must resolve to one person, because unmatched variants stay split into separate records.
7. 7Keep the profile current. The source's point about clusters applies to the whole profile: analysis that takes months leaves a customer file that is 'ancient history' by delivery. Choose tooling that refreshes daily, not quarterly.
8. 8Prove the profile with one business question that needs both online and offline data, such as which products customers buy, how often, and how discounts motivate them. If the question can't be answered from the unified profile, the integration is incomplete.

A global vitamin and supplement retailer's European operations, with more than 1,000 stores, wanted to understand online and brick-and-mortar transactions together: which products customers buy, how frequently, and how discounts motivate them. That question requires store and web purchases linked to the same customer, so integration came first. Once the data was unified, the retailer found that its most valuable shoppers, only 20 percent of the customer base, drove more than 80 percent of profit. It also found that food drives purchase frequency and attracts those high-value shoppers. Certain food lines had been considered unprofitable, and the retailer now keeps carrying them for that reason. It also learned that most customers buy infrequently and that loyal customers shop more often, and it tailored email campaigns to those behaviors. None of these findings is visible inside a single silo; each depends on the linked profile.

• If fewer than all of your channel databases (online, store, phone, web, email, social, service) feed the profile, treat it as incomplete and don't run cross-channel analysis on it yet. 68 percent of marketers in AgilOne's survey of 132 executives lack a single customer view.
• If you have a small marketing team and no data scientists, buy a cloud solution with a standard data model rather than building in-house. Building costs millions plus a team; cloud starts around $50,000 a year.
• If you send direct mail, validate addresses before the next send (NCOA, CASS, DPV), since the source cites mailing-cost savings of up to 80 percent.
• If customer segments or clusters are refreshed quarterly or less, move to daily updates. Quarterly re-scoring is expensive and still gives sub-optimal results.
• If an email domain fails the ping test or has a syntax error you can't auto-correct, suppress or flag it rather than send, to protect sender reputation.

• Integrating only the easy silos (e.g., web and email) and skipping stores or phone. The tell: customers who buy in-store appear as 'new' or 'inactive' online, and your best shoppers look unremarkable.
• Matching on raw data before standardizing. The tell: the same person appears as multiple records, with misspelled domains, malformed addresses, or varied formats that never merge.
• Mailing unvalidated addresses. The tell: returned mail, deliveries to people who have moved, and rising per-piece cost with flat response.
• Treating the profile or segments as a one-time project. The tell: the analysis took so long that marketing receives a customer file that is already ancient history, or the team re-scores only quarterly.
• Underestimating infrastructure cost and scope by building in-house. The tell: a data warehouse budget that grows with store count (Jeffrey's cited range runs from $50,000–$250,000 for a 10-store retailer to $2.5 million for 400 stores) before a single marketing use case ships.

• ▸Silo name: list each source (online, store, phone, web, email, social, service center).
• ▸Customer identifier: the key that source uses (email, postal address, loyalty ID, phone).
• ▸Standard data model mapped (Y/N): note whether its fields are mapped to one common schema.
• ▸Address validation status: note which of USPS coding/ZIP+4, NCOA, CASS, DPV are done.
• ▸Email validation status: note whether syntax correction and the mail test are done.
• ▸Dedup/link status: record the match approach and where variant records still split one person.
• ▸Refresh cadence: state how often this source updates the unified profile (target: daily).
• ▸Route chosen: build / service provider / predictive marketing cloud, with the annual cost estimate.
• ▸Proof question: name one business question needing online plus offline data, and whether the profile answers it.

1. 1Build the single customer profile before you model anything. Link online and offline customer data into one record per customer. The source says about 80 percent of the work in predicting behavior is collecting and cleaning data, so budget for it. Run the checklist's data-quality items: standardize phone, email and address; de-dupe daily at both individual and household level; use fuzzy/phonetic partial matching.
2. 2Choose the model family by the question you need answered, using the five out-of-the-box models in the capability checklist. Clusters (brand, product, behavior) answer 'who is like whom'. Likelihood-to-buy and likelihood-to-engage answer 'who to rank first'. Predicted customer value answers 'who is worth more'. Product recommendations answer 'what does this person want next'. Name one question per model and do not start with all five.
3. 3Define the output as a field that lives on each customer record: a cluster label, a likelihood score, a predicted value, or a ranked product list. A model that exists only as a slide is the 'last mile problem' the authors describe: right answers that never reach day-to-day execution.
4. 4Wire each field to an action in campaign automation before you call the model done. Name the channel (email, web personalization, direct mail, call center, display, retargeting) and the audience built from the score, for example the top-ranked likelihood-to-buy customers or a specific behavior cluster.
5. 5Measure every predictive program against a hold-out group and judge it in dollars. Use the checklist's A/B testing (results measured in $) and measurable hold-out group performance. A score that lifts opens but not revenue has not worked.
6. 6Read the model through profitability and opportunity reporting. Look at profitability measures such as returns and discounts, and at revenue-opportunity benchmarks. Connect the data to Tableau, Excel or SQL so a marketer can inspect it without a data scientist.
7. 7Pressure-test the model with the data scientist using business questions, not math. The source's advice is that business understanding trumps math and that asking the right questions goes a long way. Ask what data feeds it, what customers it treats as alike, and what action each output should drive.
8. 8Move from campaign-centric to customer-centric thinking. Each customer's score, cluster and recommendations should shape what they receive, instead of one-size-fits-all channel campaigns.

Netflix is the source's clearest case of the recommendation model running as an operating system. Research Director Xavier Amatriain wrote in 2012 that '75% of what people watch [on Netflix] is from some sort of recommendation.' The model's output, a row of titles 'you might like', sits directly on the screen the customer opens, so it drives behavior without anyone building a campaign. Amazon does the same: recommendations appear under a product the shopper is thinking of adding to the cart. Contrast this with the author's first McKinsey project, where analytics matched a technology company's sales team to customers by customer need, team skill and experience. The CEO liked the results on paper but could not operationalize them in a repeatable way. The lesson for your build is that the model, the field on the customer record and the delivery point must ship together. Netflix and Amazon placed output where the customer acts. The sales-coverage project stopped at the analysis. Before approving any predictive model, ask where its output appears and who or what acts on it.

• If you cannot name the action a model's output will trigger, then do not build the model yet. Analytics creates value only when it becomes part of day-to-day execution.
• If customer records are not de-duplicated and standardized (individual and household), then fix the data before modeling. Cleaning is about 80 percent of the work, and models trained on split profiles misread customers.
• If the question is 'who should I contact first', then use a likelihood-to-buy or likelihood-to-engage ranking. If it is 'what should I show them', use product recommendations. If it is 'how do I describe my base', use clusters by brand, product or behavior.
• If a program reports lift without a hold-out group, then treat the result as unproven and re-run it with a measurable hold-out judged in dollars.
• If you are weighing a model against your team's math skills, then decide on business understanding and the right questions first. The source ranks these above technical depth.

• Modeling on unintegrated data. The tell: the same person appears as multiple customers, or household members are scored separately, and the scores contradict what the sales team knows.
• Producing an insight with no delivery path (the last mile problem). The tell: results look right in a deck, nobody can say which campaign or channel uses them, and nothing repeats next quarter.
• Judging success by activity instead of dollars. The tell: reports cite engagement but never compare revenue against a hold-out group, and A/B tests are not measured in $.
• Ignoring profitability in the reporting layer. The tell: high-scoring customers or segments are chosen without checking returns and discounts, so a 'top' group may be unprofitable.
• Treating the model as a black box and leaving it to specialists. The tell: marketers cannot explain what a cluster or score means, so they cannot ask the right questions or decide how to use it.

• ▸Business question: the one question this model must answer, in plain words.
• ▸Model type: cluster (brand/product/behavior), likelihood-to-buy, likelihood-to-engage, predicted customer value, or product recommendation.
• ▸Data readiness: single customer profile built? De-duping (individual and household) and standardization done? Note any gaps.
• ▸Output field: the name of the score, label or list that will sit on each customer record.
• ▸Action and channel: the campaign or experience the output triggers, and the channel (email, web, direct mail, call center, mobile, display, retargeting).
• ▸Audience definition: which customers are selected from the output (for example top-ranked, or a named cluster).
• ▸Hold-out design: how the hold-out group is chosen and how the A/B test is measured in dollars.
• ▸Profitability check: what returns and discounts data you will review, and which revenue-opportunity benchmark you compare against.
• ▸Owner and review: who acts on the output, and the date the data scientist and marketer review it together.

1. 1Start with the capability checklist, not the tool demo. Confirm the platform covers campaign execution (native email or via partners, native web personalization, social, offline direct mail/call center/store, mobile messaging, display ad targeting, web retargeting) and campaign design (segmentation UI, built-in audiences, email templates, A/B testing measured in dollars, measurable hold-out groups, list export by manual and API, attribution).
2. 2Wire insights to touch points before you design any campaign. Write down every touch point (email, web, mobile, store, call center) and mark how a predictive output (likelihood-to-buy, likelihood-to-engage, cluster, predicted value, recommendation) reaches it. A touch point with no feed is a gap, because insight alone is not enough.
3. 3Make the trigger event-driven. Define each campaign as 'insight or event, then action' (for example, a cart abandonment or a predicted lapse starts a campaign) and run it in real time or as close to it as the system allows. Require daily-refreshed profiles and segments at minimum, with automated, not manual, updates; manual steps push refreshes out to once a quarter.
4. 4Split every remarketing and lifecycle flow into two branches by likelihood to return. Customers likely to come back get a simple reminder. Customers unlikely to come back get an added incentive. Apply this to abandoned cart, browse, and search campaigns.
5. 5Set lifecycle timers from the product's own rhythm. For replenishable goods (face cream, water filters, pet medication), send a friendly reminder a week or more before supply runs out. If no repeat purchase occurs within 60 days, escalate to an incentive.
6. 6Personalize dynamic content from the same insights: email content and digital ads should draw on clusters or personas, not only on the trigger. Route personas to the channels they respond to (Mavi used email and SMS).
7. 7Build measurement in at launch. Hold out a group, run A/B tests judged on dollars, and set attribution before sending. A campaign without a measurable hold-out cannot show what automation added.
8. 8Plan the bandwidth explicitly. Marketers underestimate the effort to develop and automate targeted campaigns, so assign each task to a named owner, build a campaign roadmap, and phase delivery with realistic timelines. Begin with digital sources (email and web) and add store transaction data after you have proven success.

Mavi, the Istanbul fashion manufacturer and retailer with 350+ stores and sales channels, began with a single predictive campaign. Before that, marketing and IT each kept their own reports and customer data, and marketing could not reach customer data without IT. In 2009 Mavi deployed a cloud-based predictive marketing solution that consolidated, cleansed, and de-duped customer data daily. Predictive analytics then found product-based clusters: shoppers who favored woven shirts, shoppers who favored beachwear, and shoppers who bought new-season high fashion and accessories. Mavi used these personas to run more targeted campaigns via email and SMS, starting with a reengagement campaign. The sequence is the one to copy: fix data access first, derive the segments from predictions, then deliver by channel. The source gives no results for the reengagement campaign, so do not assume any.

• If a customer is likely to return (cart, browse, or search abandoner), then send a simple reminder; if unlikely, then add an incentive.
• If a replenishable product has no repeat purchase 60 days after purchase, then escalate from reminder to an incentive, since you have nothing to lose.
• If your segments or profiles refresh less than daily or need manual steps, then fix the integration before building more campaigns, because you need near real-time, automated updates.
• If you lack proven results, then launch on email and web data first and add store transaction data after the first campaigns show success.
• If a vendor cannot show hold-out group measurement and dollar-based A/B testing, then you cannot prove incremental value; treat that as a gap.

• Insight without activation: models and dashboards exist but nothing triggers a campaign. The tell is that reports are circulated but no touch point changes behavior (the 'last mile problem').
• Data warehouse that does not talk to execution tools: IT delivers integration without knowing your campaign stack. The tell is that you export lists by hand to send them.
• Stale audiences: manual refresh steps make updates cost-prohibitive more than once a quarter. The tell is that triggers fire on customers whose status changed weeks ago.
• One-size-fits-all remarketing: everyone gets the same reminder or the same discount. The tell is that incentives go to customers who would have returned anyway.
• Underestimated bandwidth: unclear task ownership and unrealistic timelines. The tell is a campaign roadmap with no owners that slips repeatedly.

• ▸Campaign name and lifecycle stage: name it and mark acquisition, growth, or retention.
• ▸Trigger insight or event: the model output or behavior that starts it (e.g., cart abandonment, likelihood-to-buy score, predicted lapse).
• ▸Data refresh cadence: how often the profile or segment updates (daily minimum) and whether it is automated.
• ▸Segment or persona: cluster or audience receiving it, and the likely-to-return versus unlikely-to-return split.
• ▸Channels and touch points: email, web, social, mobile, direct mail, call center, store, display ads, retargeting.
• ▸Message and incentive by branch: reminder for the likely-to-return, added incentive for the unlikely.
• ▸Timing rule: delays and escalation points (e.g., reminder one week before supply ends; incentive at 60 days).
• ▸Measurement design: hold-out group size, A/B test variants, dollar-based success metric, attribution method.
• ▸Owner and timeline: who builds, who approves, and the phased launch dates.

1. 1Build the individual profile before any campaign. Per the book's sequence (Chapter 3 precedes the plays), collect purchase history, behavior, and preferences into one record per customer. Add a field for which type of personalization that customer appreciates, because the passages say appreciation varies by age, location, gender, and other factors.
2. 2Pick the personalization type to deploy from the ones the Conlumino survey of 3,000+ online shoppers named: an alert about a new product matching their interests, a refill reminder, VIP recognition, loyalty rewards, personalized discounts. Start with loyalty rewards and personalized discounts, which were popular across the board. Test the others by segment.
3. 3Set the baseline expectation: more than half of US and UK consumers expect e-commerce sites to remember their past purchases. Audit every touch point and confirm it uses purchase history. A site that forgets a past purchase falls below the floor.
4. 4Predict personas (Play Two) and group customers into behavioral buying personas. Design acquisition and retention messages for a specific microsegment rather than the whole list. The book claims hypertargeting a microsegment can raise conversions four times or more.
5. 5Layer in life cycle stage (Play Three) and timing. Predict each customer's likelihood to buy or engage (Play Five), then match message, offer, and timing to that score. The book names relevance, personalization, and timing together as the levers that keep customers coming back.
6. 6Generate individual recommendations (Play Six) around what each customer will want in the future, not only what the product line pushes. This reverses the usual question from 'who will want our product' to 'which products will this customer want.'
7. 7Watch for behavior changes that signal the relationship needs a different message. The casino example: a gambler who wagers less than usual, or cashes in all player card points, triggers an intervention such as a letter or call offering a free meal, show ticket, or comps.
8. 8Give customers control over the data: let them edit data (erase a baby-shower purchase, turn off a data source), explain why a recommendation was made, and collect feedback on good and bad recommendations. Feed the feedback back into the profile as explicit learning alongside observed behavior.
9. 9Measure relevance, not reach. The book states the key metric is relevance. Track conversion and retention by microsegment and persona, and compare against your one-size-fits-all control.

Harrah's Entertainment's Total Rewards (launched as Total Gold in 1997, renamed a year later) is the book's clearest case of relevance at scale. Predictive algorithms track each gambler's behavior and calculated lifetime value. When a gambler wagers less than usual, perhaps because a monthly visit was skipped, the casino responds individually with a letter or phone call offering a free meal, show ticket, or gaming comps. If a long-time customer cashes in all their player card points, the system flags it as possible dissatisfaction and alerts management so someone can ask whether there is a problem. The offer and the timing both fit one person's situation, and the book says this can be the difference between retaining and losing the customer. Harrah's confidence in the program was strong enough that it cut traditional ad spending by more than 50%: measured media went from $106 million in 2008 to $52 million in the first half of the following year and $20 million in the first half of the year after. The sequence is signal, individual action, retention, then reduced reliance on mass advertising.

• If a customer's behavior drops below their own norm (smaller wagers, skipped visits, cashing out all points), then send an individual retention offer or contact rather than the standard campaign.
• If a segment's appreciation of a personalization type is unknown, then start with loyalty rewards or personalized discounts, which were popular across the board, and test the other types by age, location, and gender.
• If a recommendation could be wrong or feel intrusive, then explain why it was made and offer a way to give feedback. The explanation tends to make even a mistake look logical.
• If a customer's data includes one-off or gift purchases, then let them erase that history so it does not distort future recommendations.
• If you must choose between reaching more people and matching fewer people better, then choose relevance: target the microsegment, because relevance, not reach, is the key metric.

• Treating personalization as recommendations only. The book says predictive marketing is much more: timing, lifecycle stage, offers, and retention triggers matter too. The tell: personalization appears only in a product widget while emails and offers are uniform.
• Forgetting what the customer already did. More than half of US and UK consumers expect sites to remember past purchases. The tell: you promote products a customer already bought, or ignore their history at a touch point.
• Assuming one personalization type fits everyone. Appreciation differs by age, location, gender, and other factors. The tell: a single tactic, such as VIP recognition, gets strong response in one segment and indifference or negative response in another.
• Opaque or uncorrectable data use. Without edit, explain, and feedback controls, a single mistake reads as creepy and erodes trust. The tell: customers complain about odd recommendations, or opt out of data sources, with no way to correct them.
• Optimizing for reach and volume. Blasting the full list keeps the old one-size-fits-all model. The tell: success is reported as sends and impressions rather than conversion or retention by persona or microsegment.

• ▸Customer segment/persona: name the microsegment or behavioral persona this audit covers.
• ▸Life cycle stage: state where this segment sits (acquiring, growing, at risk).
• ▸Data in profile: list purchase history, behavior, and stated preferences you actually hold for them.
• ▸Personalization type to test: choose from new-product alert, refill reminder, VIP recognition, loyalty reward, personalized discount.
• ▸Trigger signal: describe the behavior change or prediction (likelihood to buy, drop in activity) that fires the message.
• ▸Message, offer, and timing: write the individual action and when it goes out.
• ▸Customer controls provided: note whether they can edit data, see why, and give feedback.
• ▸Relevance metric: define conversion or retention measured for this segment versus a generic control.
• ▸Result and next change: record the outcome and what you adjust.

1. 1Define engagement as a per-customer set of observable signals before you score anything. List the signals you can capture: email opens and clicks, website visits, purchases, social interaction, and product usage. Mark which ones you can tie to one customer record today.
2. 2Consolidate, cleanse, and de-dupe customer data into one profile per person, and refresh it daily. Mavi did this first, with a cloud-based predictive solution in 2009. Engagement scored on fragmented or duplicate records misreads who is active.
3. 3Set engagement against a baseline of recency and consistency, not a single event. The definition stresses level and consistency. Record how often each customer interacts, and flag when the pattern slows. That is the early sign of a lapsing relationship.
4. 4Rank customers by predicted likelihood to buy or engage (Play Five) instead of by past activity alone. Use the ranking to separate customers likely to come back from those unlikely to, so each group gets a different treatment.
5. 5Cluster customers into personas by product, brand, or behavior. Match content to each cluster's preferences. Mavi found woven-shirt, beachwear, and new-season high-fashion shoppers, and sent each group the right products.
6. 6Treat engagement as continuing after the first purchase. Run postpurchase, replenishment, repeat-purchase, new-product, and customer appreciation campaigns. Apply give to get: deliver an experience of value before you ask for the next purchase.
7. 7Differentiate the response by likelihood. Send a simple reminder to customers likely to return, and add an incentive only for those unlikely to return. This applies to abandoned cart, browse, and search campaigns.
8. 8Weight engagement by customer value. Spend more to appreciate and retain high-value customers, upsell medium-value customers, and cut service cost for low-value ones. Judge retention in dollar value, not headcount.
9. 9Start with one program and build on it. Account for every discount and test it. Mavi's Elif Oner advises: 'Start small and pick just one program and build on that success.'

Mavi, a clothing retailer, consolidated, cleansed, and de-duped its customer data daily on a cloud-based predictive marketing solution in 2009. It then used predictive clustering to find groups with distinct product preferences. It found at least three personas: customers who favored mostly woven shirts, beachwear shoppers, and a group who mostly bought new-season high fashion and accessories. Mavi ran a reengagement campaign for lapsed customers by email and SMS, featuring the right products for each persona. The campaign reactivated 20 percent of lapsed customers, and each reactivated customer reduces the need to acquire a new one. Mavi now runs more than 80 predictive programs a year. Collectively they added 7 percentage points to revenues in the first few years. The book estimates that at about $52 million on 2014 revenues of $747 million, using a Wikipedia figure. The lesson is to clean the data first, segment by behavior, and run one targeted engagement program at a time.

• If a customer is likely to return, send a simple reminder. If the customer is unlikely to return, add an incentive.
• If a customer's interactions have lapsed, run a reengagement campaign built on that customer's persona rather than a generic blast.
• If a customer is high value, spend more to appreciate and retain them. If low value or unprofitable, reduce the cost to serve rather than chase engagement.
• If you must choose between retaining a customer before or after they leave, prevent the departure. Prevention is easier, cheaper, and more effective than reactivation.
• If you are starting out, pick one program, measure it, and expand only after it works.

• Measuring one channel in isolation, such as email opens only. The tell: a customer looks inactive in the email report but is buying or visiting the site.
• Scoring engagement on duplicate or unconsolidated records. The tell: the same person appears as several partial profiles, each with thin activity.
• Stopping engagement after the first purchase. The tell: new buyers receive no postpurchase or replenishment contact and drift away.
• Giving every lapsed customer the same incentive or the same message. The tell: discounts go to people who would have returned anyway, and reactivation stays low.
• Counting retained customers instead of retained dollars. The tell: retention looks healthy while revenue falls, because you kept low-value customers and lost your best ones.

• ▸Segment or persona: name the cluster (product-, brand-, or behavior-based) you are scoring.
• ▸Signals tracked: list which of email opens/clicks, site visits, purchases, social interaction, and product usage are captured per customer.
• ▸Data consolidation status: confirm records are consolidated, cleansed, de-duped, and refreshed daily.
• ▸Baseline engagement level: record typical interaction frequency and recency for this segment.
• ▸Consistency trend: note whether interactions are rising, flat, or slowing versus baseline.
• ▸Likelihood to return or engage: classify as likely or unlikely to come back.
• ▸Customer value tier: mark high, medium, or low lifetime value.
• ▸Chosen action: reminder, incentive, reengagement, appreciation, or upsell, matched to likelihood and value.
• ▸Result and next step: record reactivation or response rate, and what single program to build on next.