Score Promotions on Margin, Not ROAS

McKinsey's analysis of the average S&P 1500 company found a 1 percent price increase, with volume flat, lifts operating profit about 8 percent, and a 1 percent cut destroys the same 8 percent. A discount comes straight off contribution margin, so the only number that tells you whether a promotion paid is incremental contribution margin, not ROAS, revenue or conversion rate. Those three all rise during a discount by construction, which is why they mislead. On the honest scorecard most discounts lose money, because the breakeven volume is brutal and most of the lift was never incremental. AI can fix this by sending discounts only to the customers an offer moves and measuring against a holdout, but it optimizes the short term and ignores the brand and reference-price damage that compounds for years.

Why a discount punches so far above its weight

Price is the most powerful profit lever, which is what makes a discount so dangerous, and the same math runs in reverse on volume, you would need an 18.7 percent rise to offset a 5 percent price cut, a response that is extremely rare. The formula every marketer should have memorized is the breakeven, required volume increase equals the discount divided by contribution margin minus the discount. At a 40 percent margin a 20 percent discount needs unit volume to double to stand still, and at a 30 percent margin it needs a 200 percent lift, bars far above what discounting delivers.

Why most of the bump was never incremental

A promotion's dashboard counts every sale made with the code, but a large share of those sales would have happened anyway. John Dawes of the Ehrenberg-Bass Institute found that on average 58 percent of consumers who buy a brand on promotion had bought it in the prior 26 weeks, more so for large brands than small, so larger brands have even less to gain. The deeper result comes from van Heerde, Leeflang and Wittink, who split the promotion bump across four scanner datasets into three roughly equal thirds, brand-switching, purchases borrowed from other periods, and true category expansion. Only the last third is truly incremental, and the borrowed third is clawed back later as the post-promotion dip. The only honest way to separate the incremental sale from the subsidy is a holdout, a randomized control group that does not get the offer.

Why repeated discounting poisons full-price sales

Blanket discounting does not just fail to pay back today, it lowers what customers will pay tomorrow. Consumers hold an internal reference price, and the pricing literature is explicit that frequent discounting drags it down until buyers will not pay above the promotional price, because they underestimate the fair price of a product they keep seeing on sale. The long-run proof is Mela, Gupta and Lehmann, whose eight and a quarter years of panel data showed consumers grow more price- and promotion-sensitive over time as promotion rises and advertising falls. J.C. Penney is the natural experiment. When Ron Johnson replaced its high-low coupon model with everyday Fair and Square pricing in 2012, customers conditioned to coupons revolted, sales fell about 25 percent in a year, and he was gone by April 2013. His own post-mortem was that coupons were a drug that drove traffic. Once you have trained customers on discounts the reference price has already moved and you are trapped, the discount having destroyed the pricing power it meant to exploit.

How AI turns discounting from blanket to surgical

The throughline of every fix is the same, stop optimizing ROAS and optimize incremental contribution margin, and AI makes that switch operational in four steps. First, have an LLM build the true unit-economics model from list price, COGS and every variable selling cost, then compute the post-discount margin, the breakeven volume, and incremental margin against a no-promotion baseline. This kills most bad promotions before launch, which matters because a quarter to a third of direct-to-consumer SKUs already run negative contribution margin once fully loaded. Second, estimate discount elasticity from history with a demand regression that controls for seasonality and stockouts. Third, run a randomized holdout to measure incremental margin directly. Fourth, and this is where AI changes the game, target with an uplift model, not a propensity model. Propensity targets people likely to buy, many of whom would buy anyway, so you subsidize them. Uplift targets only persuadables, whose probability of buying rises because of the offer, while avoiding the sure things, the lost causes, and the sleeping dogs a discount pushes away. The open-source toolchain is CausalML, scikit-uplift and EconML, read off the Qini curve.

The evidence that this beats blanket discounting is peer-reviewed and large. Lemmens and Gupta found profit-based targeting delivered campaign profit 168 percent higher than conventional propensity targeting and 23 percent higher than standard uplift, and in one study 83 percent of customers were unprofitable to target at all. A self-reported practitioner case cut targeted customers by 80 percent and cost from 400,000 to 80,000 dollars while raising renewals, and in one campaign blanket treatment had the highest conversion at 18.04 percent but the lowest profit per customer, 5.18 dollars against 5.46 for uplift, the dashboard-versus-margin gap in one line.

The discount go/no-go check (copy this)

Run this before any promotion goes live.

• Score on incremental contribution margin, not ROAS, revenue or conversion, reporting post-discount margin per unit and incremental margin against a no-promotion baseline.
• Run the breakeven first. Required volume increase equals discount divided by contribution margin minus discount. At a 40 percent margin a 20 percent discount needs volume to double. If breakeven sits above your historical elasticity, kill it.
• Make a holdout the default. No recurring promotion launches without a randomized control group, and if incremental margin is at or below zero across two cycles, retire it.
• Target persuadables, not everyone. Move from propensity to uplift, and if treatment churn or return rate runs above control, you are waking sleeping dogs, stop.
• Re-engineer the structure, not just the depth. Prefer threshold offers, bundles and free-shipping just above order value, and whole-dollar or months-free framing, ranked by projected incremental margin.
• Protect the reference price. Cap promotional depth and frequency on hero and high-equity SKUs, and if full-price sell-through is falling quarter over quarter, pull back.

The full workflow, building the contribution-margin and breakeven model, estimating discount elasticity, designing the holdout, and targeting persuadables with an uplift model, is packaged as a reusable Claude skill. Get the free skill.

When a discount is the right call, and where AI misleads

The case against blanket discounting is not a case against all discounting. A discount is defensible when it reaches incremental, price-sensitive buyers without moving everyone's reference price, when it clears perishable inventory whose alternative is a write-off, when first-purchase economics are justified by measured lifetime value, and when competitive necessity demands it, though that last slides easily into a price war. Every one of these is targeted and measured against incremental margin or LTV, the discipline blanket discounting lacks. The AI workflow has its own limits. Elasticity models need real price variation and clean data, and if you have always discounted at the same times the model cannot separate the discount from the seasonal peak it rode on, because correlation is not causation. Uplift models need experimental data, you cannot observe both outcomes for the same customer without a holdout. Personalized pricing carries fairness, legal and trust risk, and peer-reviewed work finds it lowers perceived fairness even among the customers it benefits. A model maximizing this quarter's margin ignores the reference-price and brand erosion the long-run evidence warns about, so the strategy call stays with a human.

What to do Monday

Change the scorecard before you change the offers. Put incremental contribution margin on every promotion report and retire ROAS and CPA, then have an LLM build a reusable breakeven calculator and require sign-off on any discount whose breakeven runs above your real elasticity. Make a randomized holdout the default, retire the always-on coupon once two cycles show incremental margin at or below zero, and pilot uplift targeting where it beats both target-everyone and propensity on profit. Cap depth on the SKUs with the most brand equity to lose, because the reference price you protect today is the pricing power you keep tomorrow.

Sources: McKinsey, The Power of Pricing, S&P 1500 price-profit analysis; John Dawes, Ehrenberg-Bass Institute, 2013; van Heerde, Leeflang and Wittink, Marketing Science, 2004; Mela, Gupta and Lehmann, Journal of Marketing Research, 1997; Harvard Business School and reporting on J.C. Penney Fair and Square, 2012 to 2013; Lemmens and Gupta, Marketing Science, 2020; Schoder and Rossler, California Management Review Insights, 2025; CausalML, scikit-uplift and EconML.