Marketing analytics a practical guide to real marketing science

1. 1Separate the four levels before you collect anything: primary motivations, experiential motivations, behaviours, results. Results (typically financial) are caused by behaviours, and behaviours are caused by one or both kinds of motivation. Tag every candidate variable with exactly one level.
2. 2Classify primary motivations (price valuation, attitudes about lifestyle, tastes and preferences, searching, need arousal) as psychographic and mostly unseen, with no brand interaction. Plan to infer them through survey or primary market research, not the transaction database.
3. 3Classify experiential motivations (loyalty, engagement, satisfaction) as states that arise after brand interaction. Treat engagement as an experiential cause, not a behaviour, and measure it through recency and frequency.
4. 4Build behavioural variables from the two main dimensions, transactions and marcom responses. For transactions: number of purchases, amount each time, products and categories, net margin, cost of goods sold, units, discounts applied, time between purchases. For marcom: vehicle type (direct mail, e-mail), opens, clicks, website visits, store purchases, discounts used, and when each vehicle was sent.
5. 5Map the consumer decision process to find where need arousal sits. Need recognition is the gap, or 'cognitive dissonance', between the ideal state and the current state. Decide whether your lever educates on real needs (survival, satisfaction) or on pseudo-needs ('all your friends have already done it').
6. 6Segment consumers into those still considering and those who have 'already decided'. Those who have decided use brand equity or satisfaction as a shortcut and skip critical evaluation. Don't spend persuasion on them.
7. 7Treat different customers as having different sensitivities, which means different motivations and different behaviours. Segment from the customer's point of view, not from a financial value tier like RFM. Then model each segment separately, because different variables drive the outcome in each one.
8. 8Use only data you can tie to an individual customer. Drop social media/in-bound marketing and web banners from the engagement dataset when they cannot be matched to a customer, and use unique identifiers (a per-catalogue 800 number, unique promo codes) so responses trace to a specific contact.

The source's computer-hardware business case follows Scott, the analyst, as he builds a customer database for a computer and consumer-electronics marketer. He and his boss decide that transactions need only broad product granularity: laptops, desktops and workstations, one level below (high-end versus scaled-back desktop), plus consumer electronics. They record gross and net revenue, discounts, number of purchases, time between purchases and purchase months. For marcom responses, which the source calls 'a sign of behaviour and an indication of engagement', they use direct mail and e-mail. They know who got each catalogue, when, what was on the cover and what offers were inside. Each catalogue has a unique 800 number and unique online promo codes, so calls and redemptions trace to a catalogue. E-mail opens and clicks are tracked by address. They ignore social media and web banners because those can't be matched to customers. They then add derived variables, such as monthly seasonality dummies and time between purchases, from the database, and consider overlay and primary research data for the unseen primary motivations. Behaviour supplies the engagement metrics; research supplies the motivation.

• If a variable involves brand interaction (loyalty, satisfaction, engagement), then label it experiential. If it is a pre-brand attitude or need (price valuation, lifestyle, tastes), then label it primary and plan to survey for it.
• If a metric is recency or frequency, then call it engagement and an experiential cause, and keep it out of your 'behaviour causes results' count.
• If a response can't be tied to an individual customer (banners, social), then exclude it from the individual-level engagement model, or add a tracking identifier first.
• If a model's drivers differ across customer groups (for example churn from dropped calls vs price vs bill sensitivity), then segment on the shared attributes and fit a separate model per segment.
• If a customer has 'already decided' on brand equity or satisfaction, then don't spend persuasion or discount on them. Aim need arousal and information at those still considering.

• Treating engagement as a behaviour. The tell: opens, clicks or purchases are lumped into one 'behaviour' bucket and the experiential motivation level has no variables.
• Inferring primary motivation straight from transactions. The tell: price sensitivity or lifestyle attitude is asserted with no survey or research data behind it.
• Using one average model for everyone. The tell: low explanatory power and recommendations that don't fit any group, because each group has different reasons for the behaviour.
• Segmenting only on financial value (RFM tiers) and calling it customer-centric. The tell: tiers are defined by spend, with no view of why customers differ in motivation or sensitivity.
• Counting unmatched channel data as engagement. The tell: social or banner exposure is in the file but can't be joined to a customer ID.

• ▸Outcome (result): the financial result you want to explain, such as revenue, margin or growth.
• ▸Behaviours tracked: transactions (count, amount, categories, discounts, margin) and marcom responses (opens, clicks, visits, redemptions).
• ▸Experiential motivations: loyalty, satisfaction, engagement. Name the metric for each, e.g. recency and frequency for engagement.
• ▸Primary motivations: price valuation, lifestyle attitudes, tastes and preferences. Name the survey or research source for each.
• ▸Need arousal type: real need (survival, satisfaction) or pseudo-need, and the message that triggers it.
• ▸Decision-stage split: share considering vs 'already decided', and how you identify each.
• ▸Customer-matching key: the identifier tying each response to a person (unique 800 number, promo code, e-mail address).
• ▸Excluded channels: sources dropped because they can't be matched, with the reason.
• ▸Segmentation basis and models: the customer-view variables you segment on and which model each segment gets.

1. 1Define the behavior as a measured transaction field, not an attitude. Pull units, price paid and revenue per customer, per location and per product or service over a long window. The source example used two years. Add timing, channel and marcom response fields (opens, clicks, visits) where you have them.
2. 2Add competitor price to the same table. The source chain phoned three competing clinics around each of its own clinics for the 'shopped' products. This gives you own-price and cross-price on every row, so a unit change is not blamed on your price alone.
3. 3Segment before you model. Group locations by what causes behavior other than price: staffing, courtesy, location, growth, operational discounts, loyalty, defection. The source found a large, suburban, loyal segment and a small, urban, dissatisfied, high-defector segment. One pooled elasticity would hide that difference.
4. 4Fit a dependent-variable model of units on price, advertising spend and seasonal binaries, one equation per segment and product. In the source regression, q4 shifts units up by a fixed amount (about 3.8) whatever price or advertising does. Advertising is positive (0.0011, so +1,000 spend ≈ +1.1 units). Price is negative (-0.0275, so +100 price ≈ -2.75 units). Check that the signs match theory before going further.
5. 5Convert the price coefficient into elasticity. Use price coefficient × average price ÷ average units. If you model everything in natural logs, the coefficient is the elasticity, but this forces constant elasticity. Prefer the non-log form unless you accept the same response to a 5% and a 25% price change.
6. 6Translate elasticity into a revenue direction. Units always move opposite to price, but revenue follows price when demand is inelastic and follows units when demand is elastic. Raise price where inelastic and lower it where elastic. Do not apply one change to all regions, clinics and products.
7. 7Test before rollout. Pick shopped products, segments and regions, and run a test against control for a fixed window. The source ran 90 days and compared average net revenue. Judge the model by realized revenue in the test group, not by fit alone.
8. 8Where products' prices affect each other (desktop, notebook, workstation), move to a demand system. Each product's quantity is modelled on its own price and the other products' prices, solved simultaneously.

Scott's regression gives a price coefficient of -0.0275, an average price of 1,102 and an average of 63 units sold. Elasticity is -0.0275 × 1,102 ÷ 63 = -0.48. A 10% price increase therefore costs about 4.8% of units. Units fall less than price rises, so demand is inelastic and revenue follows the price up. A point-elasticity check shows the same logic. Price moves from 9 to 11 (+22%) and units from 1,000 to 850 (-15%), so units respond less than price and revenue rises from 9,000 to 9,350. The medical chain shows why segments matter. Its elasticity models, run per segment and product, found some segments and products sensitive to price and others not. Raising prices everywhere was therefore the wrong move. Management lowered prices on sensitive shopped products and ran a 90-day test against control. Test clinics beat control clinics on average net revenue by more than 10%. A follow-up read showed competitive reaction. One competitor needed only a 1% cut to offset the firm's 4% price decrease, because it was a strong substitute nearby.

• If a segment's response to a price change is smaller in percent than the price change (inelastic), then raise price. If larger (elastic), then lower price. Revenue follows whichever of price or units moves more in percent.
• If locations differ in loyalty, defection, staffing or setting, then segment first and estimate elasticity inside each segment. Never apply the chain-wide average.
• If a nearby competitor is a strong substitute, then expect small competitor price moves (1% against your 4%) to cancel your change. Model cross-price and expect retaliation in close locales.
• If the purchase is low-cost, short-lived or low-risk (limited problem solving), then expect rule-based, near-automatic choice. Price, prior experience and brand dislike dominate, so price tests have a clean signal. If it is high-cost, long-lived or a first purchase (extended problem solving), then expect more search and evaluation, and treat transaction response to price alone as incomplete.
• If you have real transaction and price history, then use it before surveys or conjoint. Surveys are self-reported and self-serving, and conjoint is still artificial.

• Raising every price across all regions, clinics and products. The tell: one price action is proposed for the whole chain with no segment-level elasticity behind it.
• Treating survey answers about price as price sensitivity. The tell: the evidence is stated opinion ('your prices are too high') while real price changes sit unused in the transactional database.
• Using an all-logs model and calling the coefficient the elasticity for every price range. The tell: the same elasticity is claimed for a 5% and a 25% change, and the response curve is concave everywhere.
• Omitting causes other than price, such as clinic differences, seasonality and competitor price. The tell: coefficients look odd, or the q4 effect gets blamed on price or advertising.
• Stopping at the sign of the effect. The tell: the report says 'units fall when price rises' (always true) and never says whether revenue rises or falls.

• ▸Product/service and segment: name the unit of analysis, e.g. one product within one location segment.
• ▸Segment drivers: list the non-price causes (loyalty, defection rate, setting, staffing, discounts) that define the segment.
• ▸Data window and fields: dates covered and whether units, price paid, revenue and competitor prices are present.
• ▸Price coefficient (units per unit of price): from the regression, with its sign.
• ▸Average price and average units: the means used in the elasticity formula.
• ▸Elasticity = coefficient × average price ÷ average units: record the value and whether |value| is below or above 1.
• ▸Revenue direction: if inelastic, raise price; if elastic, lower price. State the proposed change.
• ▸Competitor and cross-price check: nearby competitor prices, substitution strength and expected retaliation.
• ▸Test design: test and control locations, duration (e.g. 90 days), and the metric (average net revenue difference).

1. 1Name the outcome before the model. Pick the one financial or strategic metric the work serves: revenue, net margin, growth, lifetime value, ROI, satisfaction, or loyalty. Write the question as a business question, such as 'to what extent does pricing affect total revenue?' or 'WHEN is a customer most likely to purchase?'. Analytics has no point without a strategic payoff, so a project with no named outcome does not start.
2. 2Trace the outcome back to consumer behaviour. Revenue, margin, and loyalty are produced by what customers do: purchase, defect, respond, buy sooner. List the behaviour that drives your metric (purchase timing, purchase amount, share of products, open/click response, price response). Model that behaviour at the consumer level, not the competitor level.
3. 3Pull your own-firm customer-level data and roll it up only as far as the decision needs. The clinic-chain data held products purchased, price paid, discount applied, total revenue, visits, time between visits, ailment, and clinic. That let the work run at customer level and roll up to clinic level.
4. 4Read the outcome metrics as a set, not one at a time. In the clinic chain, flat annual price increases coincided with falling satisfaction, more defections, lower loyalty, falling employee courtesy, harder price enforcement, heavier discounting, and minimal growth. Check whether revenue gains are being paid for by declines in satisfaction and loyalty.
5. 5Choose the technique that turns a driver into a lever. Use elasticity modelling for price response by segment or region. Use survival analysis (lifereg) for time-until-purchase, with independent variables as levers. Convert the coefficients into interpretable terms, for example 'every 1,000 e-mails, purchases happen three days sooner'.
6. 6Translate each lever into dollars. Timing gains are worth money even when purchase count does not change ('time is money'). Discount future net revenue (revenue minus costs) at a stated rate, 9% in the source example, to get NPV and LTV. Use this to compare strategies, including giving up margin to get revenue faster.
7. 7Segment by behaviour (purchase amount, timing, product share, marcom response), fit the model per segment, and score each customer. Then set a different strategy per segment, because needs and sensitivities differ.
8. 8Prove the result with a test versus control, then judge the work by the outcome metric that moved. A strong model that does not shift revenue, margin, LTV, satisfaction, or loyalty has not delivered.

Take Customer XXX from Table 5.4, a customer in one behavioural segment. XXX buys every 88 days, which is 4.148 purchases a year. Annual revenue is 43,958 and costs are 7,296, so net revenue is 36,662. Assume year two is identical. Year one discounted at 9% gives an NPV of 33,635 (36,662 / 1.09). By the same arithmetic, year two is about 30,857 (36,662 / 1.09²), so two-year value is about 64,492. The table's printed LTV is garbled in the source, so recompute it from your own discounting convention. The baseline is this time-until-purchase of 88 days. Suppose survival analysis shows a lever, such as e-mails or bundling, shortens that interval. Revenue arrives sooner, and the chance of added purchases rises, depending on the product. A business case then weighs any margin given up against the faster revenue. The source's clinic chain shows the opposite failure. Near-uniform price rises year after year came with falling satisfaction and loyalty, rising defections, and heavier discounting. That pushed the question to 'how much does price actually affect total revenue by segment and region?', answered through segmentation, elasticity modelling, and a test versus control.

• If a proposed analysis cannot name which outcome (revenue, margin, growth, LTV, ROI, satisfaction, loyalty) it moves, then drop it or reframe it. It is interesting to the analyst but not impactful to the business.
• If revenue is rising while satisfaction, loyalty, or discount dependence is worsening, then treat the revenue as borrowed and root-cause the policy (as pricing was in the clinic chain) before scaling it.
• If a tactic only shortens time until purchase and does not add purchases, then still value it as discounted revenue received sooner, and weigh it against any margin given up.
• If descriptive and predictive analysis give contradictory answers about a customer's value, then trust the predictive model, since it explains why rather than assuming past behaviour continues.
• If the temptation is to benchmark against competitors, then pull back to consumer behaviour, since competitive focus moves you from a marketing view to a financial/economic one. Competitor prices can inform but should not drive.

• Reporting model fit or output instead of financial results. The tell is a deck with coefficients and lift but no dollar, NPV, or LTV figure.
• Optimizing one metric in isolation, such as price increases for revenue. The tell is rising discounts, defections, and falling satisfaction or loyalty alongside 'growth'.
• Applying one rule uniformly across products, clinics, or regions. The tell is the same percentage increase everywhere, with no segment-level sensitivity estimate.
• Valuing customers without discounting or costs. The tell is gross revenue quoted as LTV, with no net revenue or NPV rate.
• Assuming that giving customers what they want equals a win. The tell is a strategy that ignores wants that diverge, conflict with the firm's minimum needs, or are unknown to the customer. The job is to incentivize behaviour toward a win-win.

• ▸Business question: write it as a question about WHEN, WHO, or HOW MUCH, tied to one outcome.
• ▸Target outcome metric: revenue, net margin, growth, LTV, ROI, satisfaction, or loyalty, plus its current baseline.
• ▸Consumer behaviour driving it: the specific behaviour, such as purchase interval, price response, or defection.
• ▸Lever and measured effect: the independent variable and its translated effect, for example 'per 1,000 e-mails, purchase 3 days sooner'.
• ▸Segment(s) and baseline: behavioural segment definition and baseline time-until-purchase or elasticity.
• ▸Value calc: annual revenue, costs, net revenue, discount rate (e.g. 9%), NPV by year, LTV.
• ▸Trade-off check: margin given up versus revenue gained or accelerated, with the net result.
• ▸Guard metrics: satisfaction, defections, loyalty, discount usage, and their direction after the action.
• ▸Test design and verdict: test versus control cells, observed outcome movement, and scale / revise / stop.

1. 1Audit the current price policy first. Write down how prices were moved last cycle. The failing pattern in the source is raising nearly every product, in every location and region, by about the same amount every year. Its tells are falling satisfaction, rising defections, harder enforcement of increases and growing discount use.
2. 2Split the catalog into two groups. Cost-plus items are those thought insensitive to price, such as emergencies, radiology and major surgery, where price is input cost plus margin. Shopped items are those thought more price-sensitive. Spend elasticity work on the shopped group.
3. 3Segment customers or locations before modelling. Elasticity is estimated per segment, so segmentation comes first, then elasticity, then test vs control.
4. 4Fit a demand regression per segment and product. Include own price and competitor prices (cross-price). Own price gives customer sensitivity. Cross-price gives substitution and the likelihood of competitor retaliation.
5. 5Convert the price coefficient to elasticity: price coefficient × average price ÷ average quantity. Elasticity has no scale, so it can be compared across products. Classify |E| < 1.00 as inelastic (revenue follows price) and |E| > 1.00 as elastic.
6. 6Read the cross-price terms and the competitive set. Ask which competitors are substitutes and how strong. A strong substitute raises retaliation and defection risk. A weak or absent substitute lowers it. Add any market-research signal such as segment dissatisfaction and defector counts.
7. 7Move guidance down to the point of pricing. Take the segment price coefficient and apply it to each location's own average price and average quantity. State the move in terms of price position against surrounding competitors (highest, middle, lowest).
8. 8Run a 90-day test vs control. Use a matched control cell per tested product and location. Compare average net revenue. Roll out only what beats control, and keep testing cautious moves.

A chain of retail/medical clinics raised prices on almost everything by about the same amount every year. Satisfaction, loyalty and employee courtesy slipped, defections and discounting rose, and increases became hard to enforce. The analysts segmented the clinics and fitted elasticity models per segment. The method is shown on a simple unit model: coefficient −0.0275 × average price 1,102 ÷ average quantity 63 = −0.48. A 10% price rise then cuts units by about 4.8%, which is inelastic, so revenue follows price. One segment showed elasticity 0.095 (absolute) with dissatisfied customers and many defectors. Both competitors were substitutes and the high one was strong, so the call was to raise price only cautiously, as a test. At clinic level, the segment coefficient was applied to the clinic's own average price and quantity. For vaccine x at one clinic, guidance was a 6% price cut. That moved the clinic from the highest-priced option to a middle-priced one, and the high competitor was a weak substitute, so retaliation was judged unlikely. Over 90 days the clinic sold 84 units for 2,292 in revenue. That is about 27.28 average net revenue (printed as 2728 in the source) against 25.86 for the matched control, a 5.48% lift. Across the programme, test clinics beat control clinics by more than 10% on average net revenue.

• If |elasticity| < 1.00 and no strong substitute exists, then test a price increase, because revenue follows price.
• If |elasticity| < 1.00 but the segment is dissatisfied, has many defectors and a strong competitor substitute, then raise price only as a cautious test and watch defections.
• If a location sits at the top of its local price ladder, the segment is price-sensitive and the high-priced competitor is a weak substitute, then cut price toward the middle of the ladder. Retaliation is unlikely.
• If |elasticity| > 1.00, then do not raise price expecting more revenue. Units fall by more than the price rises.
• If the product is cost-plus (emergencies, radiology, major surgery), then keep margin-based pricing. Reserve elasticity-driven moves for shopped products.
• If the cross-price sign is positive, the competitor or channel is a substitute. A 10% rise in online price lifted offline demand 18% in the source. Price the two channels together, not separately.

• Applying one uniform increase everywhere. The tell is falling satisfaction, rising defections, growing discounts and increases that are hard to enforce.
• Reading the raw coefficient as elasticity. A coefficient of −0.0275 means little until it is scaled by average price over average quantity (−0.48). The tell is comparing coefficients across products with different price levels.
• Modelling own price only and ignoring cross-price. The tell is a cut or increase that triggers competitor retaliation or customer switching that the model never flagged.
• Using segment-average elasticity for every location. Local average price, local quantity and local competitor strength differ. The tell is a segment recommendation that fails at a location with strong nearby substitutes.
• Rolling out without a matched control. The tell is crediting a revenue change to price when seasonality or region drove it, and having no test-over-control figure to defend the move.

• ▸Product / pricing type: Mark cost-plus or shopped. Only shopped items proceed.
• ▸Segment: Name the segment and how it was defined.
• ▸Price coefficient: Enter the regression price coefficient for this segment and product.
• ▸Average price and average quantity: Enter the averages used (location-level if available).
• ▸Elasticity: Compute coefficient × average price ÷ average quantity. Mark inelastic (<1.00) or elastic (>1.00).
• ▸Competitor substitutes: List each nearby competitor, its cross-price sign, and whether it is a strong or weak substitute.
• ▸Current price position: Mark highest, middle or lowest versus surrounding competitors.
• ▸Proposed move and expected position: State the percent change and the resulting price position.
• ▸Satisfaction/defector signal: Record market-research evidence of dissatisfaction or defection in the segment.
• ▸Test design and result: Define the 90-day test and matched control cell. Record average net revenue for test vs control and the percent test-over-control.

1. 1Define the marcom question as timing and fit, not just who to target. Following Kristina's 'magic button' answer in the passages: for each customer, which product, at what time, with which promotion and messaging. Write it as: 'For [customer], send [vehicle] with [promotion] and [message] when [predicted time until purchase] is shortest.'
2. 2Segment before you message. The passages say some segments need a transactional style of communication and others a relationship style: 'One size does not fit all.' Give each segment its own story rationale and its own marcom strategy. If two segments would receive the same communication, they are not separate segments.
3. 3Match message to product penetration. If a segment is heavily penetrated in product X, send a product X message. If a segment may have a propensity for product Y given product X, test how to incentivize broader category purchases. Use a full factorial design to learn about vehicle, offer and message quickly, then aim marcom accordingly.
4. 4Score the database with time until event (TTE) per product using survival analysis. Do not use only an LTV or RFM list of high-valued customers. The passages call that descriptive targeting of the already engaged, with no implication for strategy.
5. 5Read the survival coefficients as lever sizes. Exponentiate each coefficient (e^B), subtract 1 for the percent change in TTE, and multiply by the average TTE to get weeks. Example: 'recent online visit' e^B = 0.987, so -1.3%; on an 11-week average that is -0.013 × 11 = -0.148 weeks.
6. 6Sort customers by predicted TTE for each product and send the communication to those nearest to purchase (Table 5.2 sorts by TTE, with the median at the 50% decile). Skip customers whose TTE is long, such as those who just bought that product, which saves catalogue and e-mail cost.
7. 7Build the business case: cost of marcom sent against revenue realized sooner because TTE shortened. Compare this per segment and per vehicle before scaling.
8. 8Account for lag. Treat marcom effect as acting over time on TTE, not only in the send week. Re-score the database on a regular cycle so timing lists stay current.

Scott's team (Kristina, Stacy, Mark) is asked what single piece of knowledge would improve their jobs. Kristina answers: if she knew WHEN each customer would buy a desktop or a notebook, she would stop sending useless catalogues and e-mails and send 'the most compelling marcom at just the right time with just the right promotion and just the right messaging.' The team defines the metric as days until each customer is likely to buy each product, and they sort by product so the soonest buyers get the communication. A survival model gives coefficients on TTE. 'Recent online visit' has e^B = 0.987, a 1.3% decrease; on an 11-week average that pulls purchase in by 0.148 weeks, which is real but not impactful. 'Purchased desktop < year' has a coefficient of 2.09, which pushes the next desktop purchase out by ((e^2.09)-1) × 11 = 77.934 weeks. So desktop marcom to a recent desktop buyer is wasted spend. The database is sorted by TTE (Table 5.2) and the list is built from those most likely to buy next.

• If a customer bought the product recently (e.g., a desktop within the year), then suppress that product's marcom, because TTE is pushed out by about 78 weeks.
• If a coefficient's weeks effect is tiny (0.148 weeks from an online visit), then do not build a campaign around that trigger. Use it only as a minor scoring input.
• If segments differ in how they respond (transactional vs relationship), then write separate messaging per segment. If you cannot state a different strategy for a segment, merge it.
• If a segment is heavily penetrated in product X, then send X messages. If it shows propensity for Y given X, then test incentives for Y before rolling out.
• If the proposed target list is simply the high-LTV or high-RFM customers, then treat it as descriptive and ask for TTE scoring before spending.

• Targeting only high-valued customers from LTV or RFM. Tell: the list never changes, and the only strategy is 'promote to the best customers' with no rationale for why they are valuable.
• Sending a product's marcom to customers who just bought it. Tell: catalogues and e-mails go to people with positive coefficients like 'purch desktop < year', whose TTE is far out.
• Sending one message to all segments. Tell: transactional and relationship segments receive the same creative and offer.
• Misreading coefficients as raw numbers. Tell: nobody exponentiates, subtracts 1 and converts to weeks, so a 1.3% effect gets treated as a large lever.
• Ignoring the cost side. Tell: the plan states response uplift but not marcom cost against revenue realized sooner.

• ▸Segment name and story rationale: one line on why this segment is distinct.
• ▸Communication style: transactional or relationship, with the reason.
• ▸Product in focus: the product X or Y being promoted.
• ▸Penetration status: heavily penetrated in X, or propensity for Y given X?
• ▸Median TTE (weeks): average time until event from the survival model.
• ▸Key TTE drivers: each variable with e^B, percent change (e^B - 1) and weeks (percent × average TTE).
• ▸Suppression rule: who is excluded (e.g., bought the product within the last year) and their TTE shift.
• ▸Vehicle, offer and message to test: factorial cells for channel, promotion and messaging.
• ▸Business case: cost of marcom sent vs revenue realized sooner.

1. 1State the strategic payoff before touching data. The source is blunt: 'There is no point to analytics unless it has a strategic payoff.' Write one sentence naming the decision the segments will change (a price, a marcom vehicle, or a segment to DE-market for efficiency).
2. 2Pick the segmentation base: behavioral data only, meaning transactions and responses to marketing communications. Pull those fields at the customer level. Customer attributes that are not behavior stay out of the clustering and are used later to profile the segments.
3. 3Decide a priori or not. The source's segmentation chapter treats 'A priori or not?' as an explicit fork. A priori means you impose the segment definitions (for example, rule-based cuts). Not a priori means you let an algorithm find the groups. Choose before running anything and record why.
4. 4Choose the technique and run it. The source names K-means and latent class analysis (LCA) and devotes a section to comparing them. Run the candidate method(s) on the same behavioral inputs so the comparison is like-for-like. Do not treat RFM as the endpoint: the book's highlight 'Why Go Beyond RFM?' argues behavioral segmentation gives you something RFM does not.
5. 5Score the solution against the source's 'criteria for actionable segmentation' and 'metrics of successful segmentation'. The test is homogeneous within, heterogeneous between. Then ask whether each segment can be treated differently across the four Ps (product, price, place, promotion).
6. 6Describe every segment in plain business language. Give each a name, a size, and what it does (for example, price-sensitive, or prefers a given marcom vehicle). Write one differentiated action per segment, including the option of de-marketing a segment.
7. 7Model within segments, not across them. In the book's retail/medical clinic chain highlight the order is segmentation first, then elasticity modelling, then test vs control. Estimate the driver (such as price elasticity) per segment to quantify what moves each one.
8. 8Validate with a designed test versus a control group, and stay involved in the design: 'everyone wants to test, but almost no one knows how to design a statistical test.' Set check-ins at 30, 90 and 180 days. Report results in financial terms (for example 'net profit can increase by 2.5% next quarter'), not as a t-ratio.

The source's own case is the retail/medical clinic chain field test, summarised in the contents as three steps in a fixed order: 'First: segmentation. Then: elasticity modelling. Last: test vs control.' The passages give the structure but not the figures, so no numbers are reported here. Step 1: the chain's customers are divided into behavioral segments. Step 2: an elasticity model is fitted so that each segment gets its own estimate of how demand responds to price. This is the book's example of knowing that 'this segment is sensitive to price' and therefore which lever to pull. Step 3: segment-specific treatments are run against a control group to confirm that revenue actually rises. The book's claim is that segmentation plus elasticity modelling can maximize revenue, and the test-vs-control step is what turns that claim into evidence. The point to copy is the sequence: segment, then quantify each segment's response, then prove it with a test.

• If you can't name a different action for a segment (a different price, vehicle or de-marketing decision), then merge it into a neighbour; it fails the actionability criterion.
• If the only behavioral input is recency, frequency and monetary value, then test a fuller behavioral segmentation against it, since the source argues behavioral segmentation provides something RFM does not.
• If you need discovered groups rather than imposed rules, then work not-a-priori with a clustering method, and compare K-means with LCA rather than assuming one.
• If segments exist but their response to price or promotion is unknown, then fit the elasticity or response model per segment before assigning treatments.
• If a stakeholder asks whether it worked, then answer with a test-vs-control result in financial terms, and never with model statistics.

• Segments with no differentiated action. The tell: the segment profile page reads well, but the treatment column is identical across rows.
• Segmenting on non-behavioral attributes and calling it behavioral. The tell: the inputs are demographics or opinions, not transactions or marcom responses.
• Treating RFM as the whole answer. The tell: every segment is defined by the same three variables and no response to marketing communications is used.
• Deploying without a designed control. The tell: the evidence for success is a before/after comparison, not a test vs control group.
• Reporting technical output to decision makers. The tell: slides lead with significance and fit statistics, and nobody asks about the money.

• ▸Decision this segmentation changes: one sentence tying it to a strategic payoff.
• ▸Behavioral inputs: list the transactional and marcom-response fields used.
• ▸A priori or not: state which, and why.
• ▸Technique(s) compared: name the method(s), for example K-means and/or LCA.
• ▸Segment name and size: one row per segment.
• ▸Defining behavior: what this segment does that others do not.
• ▸Differentiated treatment: the price, product, place, promotion or de-marketing action for this segment.
• ▸Response model: the elasticity or response estimate to be fitted for this segment.
• ▸Test design and check-ins: the control group definition and the 30/90/180-day review dates.
• ▸Result in financial terms: the expected or measured profit or revenue effect, in plain language.

1. 1List the external drivers you hypothesize move demand before touching data. The source names seasonality, competition and consumer confidence alongside pricing, campaigns and engagement. Mark each as external (outside firm control) or internal (a marketing lever), and as blatant or latent.
2. 2Score competitive intensity with Porter's five forces: bargaining power of buyers, bargaining power of suppliers, threat of new entrants, rivalry among existing firms, and threat of substitutes. Write one line per force saying whether it pushes price down or margin down.
3. 3Translate each force into consumer behaviour. Powerful buyers mean price-sensitive consumers. New entrants mean consumers who are informed about other offers. Rivalry is a zero-sum fight over a fixed pool of potential loyal customers. Substitutes mean consumers can trade down to lower-priced products. Competing with a firm means incentivizing its consumers, so measure the consumer response, not just the rival's moves.
4. 4Rate market power. The more intense the competition, the more your firm is a price taker with little control over price, and the less profit you can expect. This sets how much price headroom your model and plan may assume.
5. 5Enter the external variables in the demand or engagement model next to pricing, marcom and engagement, so the model shows how much explanatory power engagement has given seasonality, competition and the rest. Do not model engagement in isolation.
6. 6Test the full variable set for collinearity. Use VIF first, then condition indexes. Count the indexes above 30. In the variance-proportion table, treat any proportion above 0.50 as a red flag and note which variables share it.
7. 7Fix collinearity by understanding how the data are generated, not by deleting variables. In order: combine the offending variables into a factor (factors are orthogonal), transform one, take differences of the raw measure, get a larger sample. Consider ridge regression, which trades collinearity for biased coefficients and needs special software. Drop a variable only as a last resort, and only with a theoretical reason.
8. 8Pick one strategic response to the conditions: low cost (operational excellence), differentiation (product leadership), or a niche segment (customer intimacy). Porter and Treacy and Wiersema both say to concentrate on one, because firms that try to do everything are diluted.

The source's collinearity diagnostic shows how contextual drivers can contaminate each other in a model. The correlation matrix has an intercept eigenvalue of 6.86, so the first condition index is the square root of 6.86/6.86 = 1.00. The second is the square root of 6.86/0.082 = 9.142. Three condition indexes exceed 30 (230.42, 1048.1 and 432750), so there may be three collinearity problems. In the variance-proportion table, X6 loads on the intercept, X1 and X4, and X5 loads on X2 (0.8306) and X6 (0.504). Several proportions exceed 0.50, so X6 is flagged as the most problematic variable. The remedies follow in order. Combine X5 and X6 into one orthogonal factor. Transform X6 (exponent, square root) to find an X6-like variable that is correlated with the dependent variable but less correlated with X5. Take differences of X6 instead of the raw measure. Try a larger sample. Only then, with a theoretical reason, drop X6 and re-run. The same logic applies when seasonality, competition, consumer confidence and pricing share variance. Read the diagnostics before you interpret any external-factor coefficient.

• If competitive intensity is high (strong buyers, easy entrants, close substitutes, heavy rivalry), then treat the firm as a price taker, avoid price-increase assumptions, and plan around cost leadership, differentiation or a niche.
• If a competitor's move looks like firm-versus-firm, then reframe it as how it incentivizes your consumers, and measure that consumer response.
• If any condition index exceeds 30 and a variance proportion exceeds 0.50, then fix the correlated variables (factor, transform, difference, more data) before reading coefficients.
• If a context variable looks insignificant or unstable, then check collinearity and business logic before dropping it. Dropping is a last resort and requires a theoretical reason.
• If you consider ridge regression to fix collinearity, then accept that the coefficients become biased. Whether the trade is worth it depends on the case.

• Treating competition as a firm-versus-firm game (game theory) and forgetting consumer behaviour. The tell: the competitive analysis contains rival actions but no consumer price sensitivity or loyalty data.
• Leaving external drivers out of the engagement or demand model. The tell: engagement gets credit for lift that seasonality, competition or consumer confidence produced.
• Dropping a collinear variable by reflex. The tell: the variable list changes on statistical diagnostics alone, with no strategic reason recorded.
• Ignoring collinearity diagnostics because VIF looked fine. The tell: unstable or wrongly signed coefficients, or condition indexes far above 30 that nobody checked.
• Trying to be low-cost, differentiated and niche at once. The tell: the strategy document lists all three and the offers diverge across channels.

• ▸Outcome being explained: name the demand, purchase or engagement metric the context affects.
• ▸External drivers: list seasonality, consumer confidence, competition and any other uncontrolled factors, each marked blatant or latent.
• ▸Five-forces read: one line each on buyer power, supplier power, new entrants, rivalry and substitutes, with the direction of the price or margin effect.
• ▸Consumer-behaviour translation: for each force, state the consumer response it implies (price sensitivity, informedness, switching to lower-priced products).
• ▸Market power rating: price taker, partial or strong, with the evidence behind the rating.
• ▸Model inclusion: the variables entered alongside pricing, marcom and engagement, with the measure used for each.
• ▸Collinearity check: VIF, condition indexes above 30, and variance proportions above 0.50 with the variable pairs they point to.
• ▸Remedy applied: factor, transform, difference, larger sample or ridge, plus the reason. If a variable is dropped, state the theoretical justification.
• ▸Strategic stance: low cost, differentiation or niche, with the single focus stated in one sentence.

1. 1Write the question as one sentence and name its output type before choosing a technique. The book's chapter structure maps question to method: what drives demand or elasticity (single-equation dependent-variable model), who is most likely to buy (logistic regression with lift charts), when customers are most likely to buy (survival analysis), outcomes that feed each other (simultaneous equations), what the market looks like (segmentation), and unobservable constructs such as loyalty (SEM).
2. 2Check whether real behavioural data exists for the question. For an existing product, use real transaction data, because the author holds that real data is far superior to survey-derived estimates. Reserve conjoint for new products or services where no real data exists. If you only need a price range, consider a van Westendorp survey (the 'too expensive' and 'too cheap' price curves cross at the hypothesized optimal price) rather than forcing price into a conjoint design.
3. 3Test the data source against the author's four objections to conjoint-derived elasticity. (1) Are the estimates contrived rather than real economic data? (2) Is the sample too small for corporate strategic choices? (3) Is the data self-reported, so respondents spend no real money? (4) Does real data exist that would do the job better? If any answer is yes, downgrade the claim or switch the data.
4. 4Choose the model structure that matches the causal story. If the outcome is driven by one equation, use a single-equation model. If variables influence each other, use simultaneous equations. If the key driver is latent (loyalty, like intelligence, is only indirectly measured), use SEM, where manifest variables such as attitudes, convenience, share of voice, competitive density, transactions and responses reveal the latent variable.
5. 5Run the diagnostics the technique requires before you read coefficients. For logistic targeting models this includes the collinearity and variable diagnostics the book covers. For estimators, check the desirable properties (the book lists them in the simultaneous-equations chapter) rather than assuming they hold. Correct the violation; do not alter or discard the data to get a cleaner fit.
6. 6Evaluate the model with a design that can disprove it. Use lift charts for propensity models and a test-versus-control design for pricing or elasticity work, as in the retail/medical clinic chain field test, which ran segmentation, then elasticity modelling, then test versus control.
7. 7Read results by segment and by path, not just in aggregate. Look at which coefficients are large and significant and which are insignificant, then name the lever the result points to (for example, a path with a controllable driver).
8. 8Prefer predictive over descriptive output when the decision is about the future. The book's lifetime-value highlight argues predictive analysis is superior to descriptive analysis, so do not hand over a historical average as a forecast.

Take the book's loyalty model. Behaviour (transactions and marcom responses) is treated as caused by a latent loyalty spectrum, from none through transactional to emotional. That spectrum is in turn driven by attitudes (price, value, satisfaction, quality) and by operational metrics (convenience, share of voice, competitive density). Loyalty has no direct measure, so SEM with a factor-analysis-type step teases it out from manifest variables. Segments X and Y were defined by behaviour and have roughly the same transaction and response metrics. A behavioural view alone would call them equally loyal. The SEM separates them. Segment X shows very high, significant parameter estimates for convenience and competitive density, a strong negative effect of share of voice, and high positive effects of price and quality attitudes. Most emotional-path variables are insignificant. That is a transactional-only pattern, and the author calls it a red flag, especially when the segment looks loyal from purchase counts. The model also gives the lever. The emotional path's single largest impact is share of voice, which is somewhat controllable, so that is where to push to move X from transactional to emotional loyalty. A purchase-count summary would have missed this. The technique fit is the reason the lever translates into a predicted outcome.

• If the product is new and no real transaction data exists, use conjoint. If the product already sells, use real data and treat conjoint-derived price elasticity as contrived.
• If the question is price response for an existing product and only survey data is on offer, prefer a van Westendorp-style price question for a price range, and estimate elasticity from actual sales.
• If the quantity you care about cannot be measured directly (loyalty), use SEM with a latent variable instead of a proxy such as spend or transaction count.
• If two behaviours look identical in the data but may differ in cause, model the drivers before choosing the lever.
• If the question is about timing ('when'), use survival analysis. If it is about likelihood ('who'), use logistic regression. Do not substitute one for the other.

• Reading conjoint price levels as an elasticity curve. The tell: the 'elasticity' rests on self-reported choices from a small survey sample, not on observed purchases.
• Treating behaviourally similar segments as equally loyal. The tell: spend and transaction counts match, yet no attitudinal or operational drivers were modelled, so a transactional-only segment gets an emotional-loyalty strategy.
• Selecting the technique first and the question second. The tell: the model type is named in the brief before the decision it must inform is stated.
• Interpreting coefficients before checking collinearity and other diagnostics. The tell: unstable or counterintuitive signs on correlated variables that are explained away rather than investigated.
• Reporting a descriptive history as a prediction. The tell: lifetime value quoted as a past average with no forward model.

• ▸Decision to inform: state in one sentence what will change depending on the answer.
• ▸Question type: tick demand/elasticity, who will buy, when will they buy, simultaneous drivers, market structure, or latent construct.
• ▸Candidate technique: name the method that matches the question type.
• ▸Real data available?: yes/no, and the source. If no, note whether the product is new (the only case where conjoint is defended).
• ▸Data-quality objections: record any of the four conjoint objections (contrived, small sample, self-reported, better real data exists) that apply.
• ▸Diagnostics to run: list the checks required (for example collinearity, variable diagnostics, estimator properties) and the result of each.
• ▸Alteration log: list any data removed or changed, and why. 'None' is the target.
• ▸Validation design: lift chart, test versus control, or other, with the pass/fail threshold set in advance.
• ▸Lever and predicted outcome: name the controllable driver the model points to and the outcome you expect from moving it.