The Promo & Discount Playbook — Stop Burning Margin on Customers Who'd Buy Anyway

Promotions are usually the largest discretionary spend in an e-commerce P&L after marketing — and the worst measured. Marketing has attribution debates; promo often has no measurement at all beyond "GMV went up during the sale." Of course it did. The question was never whether GMV goes up. It's whether the gross profit from incremental sales exceeds the margin given away on sales that would have happened anyway. Everything in this playbook serves that one sentence.

1. The Margin Math Nobody Runs

Start with the brutal arithmetic of discounting. When you cut price, you sacrifice margin on every unit to gain volume. The required volume lift just to break even on gross profit:

Required volume lift = 1 ÷ (1 − discount ÷ margin) − 1

Read that table again. A 20% discount on a 30%-margin product needs to triple volume just to hold gross profit flat — and that's before counting the promo's operational costs, and critically, before the incrementality haircut below. In electronics or grocery, where margins run 10–25%, deep sitewide discounts are mathematically incapable of paying for themselves on margin alone. They must be justified by something else: customer acquisition, inventory clearance (see the inventory playbook), or competitive defense — and each of those justifications has its own metric you should be tracking instead.

2. Incrementality — The Only Question That Matters

Every order placed during a promotion is one of these:

The promo P&L in one line:

Promo profit = (incremental units × unit margin after discount) − (non-incremental units × discount) − operational costs

The typical unmeasured promo shows a beautiful GMV spike where 60–80% of redeemed discounts went to customers who were buying anyway. That's how "50–60% of promotions lose money" happens while every campaign readout claims success — the readouts measure participation, not incrementality.

3. Holdout Design — Your Measurement Infrastructure

Incrementality can only be measured against a counterfactual, and the gold standard is embarrassingly simple: randomly withhold the promo from a slice of eligible customers and compare.

# The entire incrementality readout, conceptually
# treated:  eligible customers who could see/use the promo
# holdout:  randomly selected eligible customers who could not

incremental_rate   = treated.cvr - holdout.cvr
incremental_orders = incremental_rate * n_treated
true_cost_per_inc  = total_discount_spend / incremental_orders

# Compare true_cost_per_inc against unit margin →
# the promo is profitable iff margin_per_order > true_cost_per_incpython

Design rules that make or break the measurement:

• Randomize at customer level, before the campaign. Comparing redeemers vs non-redeemers is selection bias in its purest form — redeemers were already your most purchase-ready users. This error alone accounts for most inflated promo ROI numbers in the industry.
• Size the holdout with power analysis (see the experimentation playbook). 5–10% of the eligible base is typical — small enough that the business barely notices, large enough to measure.
• Keep a permanent "universal holdout" — 2–5% of customers excluded from all CRM promos for a quarter. This measures your entire promo program's incrementality, catching the interaction effects single-campaign holdouts miss (customers trained to wait for vouchers).
• Measure over a window longer than the campaign — 2–4 weeks after, to catch pull-forward. If treated customers stop buying right after the promo while holdout keeps buying, you borrowed demand, not created it.

4. Cannibalization, Pull-Forward & Stockpiling

Three demand-shifting effects that quietly rewrite your promo P&L:

Product cannibalization

Discounting Pepsi steals from Coke on your own shelf. If you carry both, your incremental category volume is far below the promoted SKU's lift. Measure at category level, not SKU level: the promoted SKU's +80% means little if category volume moved +6%. This is the single most common way merchandising teams fool themselves.

Pull-forward (inter-temporal cannibalization)

Diapers, detergent, pet food: customers don't consume more because it's discounted — they buy the same annual volume earlier and cheaper. The tell-tale signature is a post-promo trough in the sales curve. Always plot 4 weeks after the campaign; if the area of the trough approaches the area of the spike, the promo was a financing operation, not a growth one.

Stockpiling & deal-training

Run the same "20% off first Friday of the month" long enough and customers learn to never buy at full price. Deal-trained customers are structurally lower-LTV. Track full-price purchase rate by cohort — if customers acquired via deep discounts never convert to full-price buyers, your acquisition promos are buying revenue, not customers. Irregular timing and targeted (rather than public) offers slow the training.

5. Uplift Targeting — Who Gets the Discount

Holdouts tell you if a campaign was incremental. Uplift modeling makes the next campaign incremental by design — by predicting, per customer, how much the discount changes their behavior. The four types every promo owner should internalize:

The practical implementation is less exotic than the name suggests. From any past campaign that had a randomized holdout, train a T-learner:

# T-learner uplift: two models, subtract predictions
import lightgbm as lgb

m_treated = lgb.LGBMClassifier().fit(X[treated], y[treated])
m_control = lgb.LGBMClassifier().fit(X[holdout], y[holdout])

uplift = m_treated.predict_proba(X_all)[:, 1] \
       - m_control.predict_proba(X_all)[:, 1]

# Rank customers by uplift; send vouchers to the top decile(s).
# Negative uplift = sleeping dogs → exclude entirely.
# Evaluate with Qini curves / uplift@k, NOT accuracy or AUC.python

Features that typically carry uplift models: discount dependency (share of past orders with a voucher), recency, price sensitivity proxies (brand mix, price-band mix), engagement without purchase (sessions, carts), and lifecycle stage. Libraries: causalml (Uber) or scikit-uplift. The realistic payoff, consistently reported across the industry: same incremental revenue at 30–50% less promo spend, simply by not paying the sure things.

6. The Segment Playbook

While your uplift model trains, RFM segments plus discipline get you 70% of the value:

7. Discount Depth & Mechanics

• Depth: use the smallest effective incentive. Test 5/10/15% ladders with holdouts per arm. The persuadable who needed 15% off is often equally persuaded at 10% — response curves flatten fast, and every extra point is pure margin. Where elasticity data exists, size depth to it per category.
• Threshold offers ("X off orders above Y") protect margin structurally: set Y at 110–120% of the segment's typical AOV and the discount funds its own basket growth. Watch the basket-padding-then-return pattern in guardrails.
• Percentage vs fixed amount: % reads better on low prices ("20% off" > "AED 6 off"), fixed reads better on high prices ("AED 100 off" > "4% off") — and fixed amounts cap your exposure on expensive baskets.
• Vouchers beat price cuts for measurement: a voucher is targetable, holdout-able, and trackable per customer. A public price cut can't be randomized at customer level (that's a pricing test — use region or time splits). Structural implication: shift budget from sitewide events toward targeted CRM offers as your measurement matures.
• Funded vs unfunded: in marketplaces, who pays the discount — you, the seller, or the brand — changes the P&L completely. Track incrementality on your contribution margin, not the face value of the discount.

8. The Promo P&L Dashboard

If your promo reporting shows GMV and redemptions, it's a participation tracker. The operator's version:

9. Implementation Roadmap

Month 1 — Measure

• Add a 5–10% randomized holdout to your next two CRM campaigns. Report incrementality next to the naive readout.
• Build the margin-math table (section 1) for your top categories — circulate it. It changes conversations immediately.

Month 2 — Stop the bleeding

• Suppress obvious sure things (recent frequent full-price buyers) from the next campaign. Measure the spend drop vs revenue drop.
• Stand up the promo P&L dashboard; add the post-promo trough view.

Month 3 — Target

• Train a T-learner on the holdout campaigns from month 1. Deploy top-decile targeting on one campaign, holdout-validated.
• Propose the universal holdout for next quarter.

Sources & Further Reading:
Glencoyne: Promotion ROI Modeling • CausalML (Uber) • scikit-uplift • RELEX: Promotion Optimization