An ecommerce order can increase turnover while weakening the business that fulfilled it. Discounting, free delivery, paid acquisition, returns and the cost of carrying slow stock can consume the apparent value of a sale before it reaches gross profit.
This is why the commercial case for autonomous product selection should not be judged by clicks, sends or attributed revenue alone. Its more valuable contribution is the ability to identify products that are sufficiently relevant to an individual without first making them cheaper for everybody.
The principle is not that discounts must disappear. Genuine clearance, seasonal trading and deliberate price investment remain valid retail decisions. The problem is habitual discounting used as a substitute for individual relevance.
A promotion cannot distinguish persuasion from subsidy
A blanket promotion treats very different customers as though they require the same reason to buy. The offer reaches people who would not have purchased, people who might purchase something else and people who were already prepared to buy at full price.
Only the first group may represent genuinely induced demand. For the last group, the discount subsidises an order that was already likely. Revenue is recorded, but margin has been surrendered unnecessarily.
Consider a product priced at £100 with a landed product cost of £55. At full price, its gross margin is £45. A 20% discount reduces the selling price to £80 and gross margin to £25. The retailer must sell 1.8 discounted units to create the same gross margin as one full-price unit, before considering extra fulfilment, payment, service and return costs.
That arithmetic is why campaign revenue can look healthy while profitability deteriorates. A promotion can raise the number of orders and still leave the retailer needing substantially more transactions merely to stand still.
Relevance changes the reason to buy
Price is one form of relevance: almost any product becomes more attractive when made sufficiently cheap. Individual product relevance is different. It asks which available item currently has the strongest purchasing probability for one identifiable customer, using that person’s changing behavioural and transaction evidence.
When autonomous product selection improves, the commercial proposition changes from “buy something because it is cheaper” to “consider this because it is unusually pertinent to you now.” The customer can still decline. The important difference is that the retailer has not reduced the price of thousands of products for thousands of people before learning whether a reduction was necessary.
This is the practical connection between SwiftERM’s next-purchase prediction and conventional product recommendation. Similarity, popularity and merchandising judgement can all supply useful evidence, but the final decision must concern the product this individual is most likely to buy next.
Full-price conversion is a margin measure
Conversion rate is often discussed without its price context. A campaign converting 5% of recipients after a deep promotion is not automatically more valuable than one converting 3% at full price. The useful comparison includes gross profit per eligible customer, not merely orders per recipient.
A simple assessment uses four figures:
- eligible customers exposed to the decision;
- net sales after cancellations and returns;
- product cost attached to those net sales; and
- the same measures for an equivalent holdout group.
The difference in gross profit per eligible customer shows whether the activity created additional margin. SwiftERM’s framework for measuring incremental ecommerce revenue explains why a controlled comparison is stronger than crediting every order that followed a click.
The calculation should be repeated for full-price and reduced-price orders separately. Otherwise a system may appear to improve performance while merely concentrating demand around products whose margin has already been surrendered.
Abandoned-basket incentives can discount resolved intent
An abandoned basket is not a diagnosis. The customer may have been comparing delivery dates, checking a size, waiting for payday, changing devices or intending to complete the order later. Automatically issuing an incentive assumes that price caused the pause.
That assumption can train customers to wait. It can also discount the people whose sequence of actions already indicated strong intent. The retailer pays to accelerate or claim credit for a purchase that might have occurred unaided.
Autonomous product selection takes a different route. It continues evaluating the individual across the active catalogue rather than making the abandoned item and a discount the only available response. The strongest next choice may remain the basket product, become a substitute, move to a complementary item or decay altogether.
This does not prohibit a basket incentive. It makes one a governed commercial exception rather than an automatic reflex.
Blanket discounts create a measurement illusion
Promotions are easy to attribute because the code, campaign and purchase can be connected. That traceability is not causality. Customers predisposed to buy are also the people most likely to open, click and redeem.
The result is an attractive report containing revenue that the retailer may have received anyway—at a higher price. Gross-margin protection begins by separating observed orders from caused orders.
A holdout should therefore remain eligible for the same underlying demand while receiving no autonomous intervention. Compare net revenue and gross profit per eligible person across sufficient time. If the exposed group produces more gross profit after product cost, reductions and returns, the intervention has demonstrated commercial value rather than merely collected attribution.
Long-tail productivity is not clearance by another name
Large catalogues usually contain a small group of products that receive disproportionate exposure. Human campaigns reinforce this pattern because teams understandably select recognisable products, new launches and dependable bestsellers.
Autonomous product selection can evaluate a much broader catalogue for each person. Within autonomous product selection, a less popular item can earn exposure when individual evidence makes it the strongest candidate—not because the retailer has decided to clear it indiscriminately.
Research on long-tail recommendations shows the opportunity and the danger. Work on recommendation technologies and sales diversity found that recommendations can also concentrate purchases around popular items, depending on how the system is designed. Catalogue breadth therefore needs to be measured rather than assumed.
Useful measures include the proportion of active products selected, the share of incremental gross profit coming from outside the bestseller group, the age of stock sold without reductions and whether returns rise when catalogue breadth increases.
Stock carrying cost changes the real margin
Gross margin at the order line is incomplete when stock remains unsold for long periods. Capital is tied up; warehousing, insurance, handling, shrinkage and obsolescence continue; and the probability of a later markdown rises.
Autonomous product selection can improve catalogue productivity before clearance becomes necessary. It exposes appropriate long-tail products to the individuals for whom they have a defensible purchase probability while retaining normal commercial safeguards for availability, suitability and brand presentation. Research into relevance-based recommendation for long-tail and overstock products illustrates why relevance must remain part of the decision even when stock productivity is a commercial objective.
This should not become a hidden instruction to push overstock. Customer relevance remains the governing condition. Forcing unwanted products may improve a short-term stock measure while damaging returns, trust and future value.
Margin rules belong around the decision, not instead of it
Retailers need commercial boundaries. Products may be excluded because of availability, regulation, returns risk or brand commitments. Minimum margin, frequency and stock rules may also be appropriate.
But a margin rule should not quietly become a merchandising campaign that predetermines the answer. First determine which products are plausible for the individual; then apply legitimate commercial constraints transparently. If the retailer chooses the collection in advance and the system merely orders it, product selection is not autonomous.
The ecommerce personalisation due-diligence checklist sets out the questions needed to distinguish catalogue-wide individual selection from campaign automation.
Guard against false urgency and misleading reductions
Margin protection must remain compatible with fair treatment. The UK Competition and Markets Authority’s guidance on urgency and price-reduction claims warns businesses about misleading countdowns, scarcity statements and comparison prices. Artificial pressure is not a sound substitute for relevance.
The governance questions are straightforward:
- Is a reduced price genuine and correctly compared?
- Is a scarcity statement supported by the stock and eligibility that apply to this customer?
- Would the communication remain defensible if the customer understood how the decision was made?
- Can the retailer explain why an incentive was offered and whether it created incremental gross profit?
These controls protect both the customer and the credibility of the margin result.
A finance-ready measurement framework
For every controlled period, record eligible customers, net units, net revenue, product cost, reductions, returns and contribution before fixed overhead. Report the exposed and holdout groups using the same definitions.
Then calculate:
- net revenue per eligible customer;
- gross profit per eligible customer;
- gross margin percentage;
- full-price order and unit share;
- average reduction per net order;
- return-adjusted gross profit;
- catalogue coverage; and
- incremental gross profit attributable to the intervention.
No single figure should dominate. A higher margin percentage can accompany lower total gross profit; greater catalogue coverage can be accompanied by more returns; and increased full-price sales can simply shift demand between channels. The controlled, customer-level comparison keeps those trade-offs visible.
What autonomous product selection means for SwiftERM
SwiftERM does not set the retailer’s prices or replace its commercial governance. It continually determines which available products each identifiable customer is most likely to buy next, using the retailer’s first-party behavioural, product and transaction evidence.
That separation matters. With autonomous product selection, the retailer retains control of price, stock, consent, exclusions and brand. SwiftERM supplies the changing individual decision that can reduce dependence on blanket incentives and manual product campaigns.
The defensible claim is not that autonomous product selection guarantees margin growth. It is that it creates a testable route to more full-price relevance: select better, discount deliberately, preserve a holdout and measure additional gross profit rather than attributed turnover.
References and accreditations
- Competition and Markets Authority, Using urgency and price reduction claims online — guidance on transparent scarcity, urgency and comparison-price claims.
- Competition and Markets Authority, Evidence review of online choice architecture and consumer and competition harm — evidence concerning reference prices, time pressure, scarcity and consumer decisions.
- Yi, Kim and Ju, Recommendation technologies and consumption diversity — experimental evidence on product recommendations, search and sales diversity.
- Alcalá-Fdez et al., Item-based relevance modelling of recommendations for getting rid of long-tail products — relevance-based treatment of long-tail and overstock products.
- Dubé and Misra, Personalized Pricing and Consumer Welfare — field evidence on pricing, firm profit and consumer welfare; included to distinguish personalised pricing from individual product selection.
- Rabinovich, Sinha and Laseter, Unlimited shelf space in Internet supply chains: Treasure trove or wasteland? — evidence on online assortment breadth, sales concentration and returns.