Blog: Pricing & AI—Retail Analytics and Strategies
The newspaper of the pricing
Many organizations receive a report on competitors’ price differences every morning, but few have a strategy. The difference comes down to three questions that must be asked before implementing the system: Why collect this data? What exactly should be tracked? And what decisions should be made once a price difference is identified?
In a network of 90 stores supported by Booper, the analysis showed that 10% of the products account for 80% of sales. Focusing monitoring and alert management on this core group yields better results than monitoring everything with the same intensity.
Web scraping measures the price displayed online; in-store surveys measure the price and product placement on the shelf; and consumer panels measure the price actually paid and the volume sold. These are three different pieces of information, not three different details of the same piece of information.
The most costly mistake isn't choosing the wrong source; it's combining them into a single price index. The correct breakdown is simple: scraping identifies items, the panel confirms them, and fieldwork verifies what the other two miss.
A stockout is almost never a logistical accident: it’s the result of a forecast that underestimated demand. It then skews the historical data, since lost sales appear as weak demand, and sets the stage for the next stockout.
Pricing is affected in two ways. A promotion on a sold-out item underestimates its own impact and makes demand appear less price-sensitive than it actually is. Price is also the quickest way to curb demand before the shelf runs out.
An exogenous event (weather, calendar, trend, competitor’s action) affects demand without appearing in the historical data. Not all of these events warrant inclusion in the model: their weight varies depending on the category and season.
Our position: An exogenous factor is only meaningful if it influences a decision—often a price. We use weather forecasts for the short term, adjust for movable holidays before making comparisons, and monitor trends and competitors.
The sales, procurement, supply chain, and pricing departments often each produce their own demand forecasts, each with legitimate biases: optimism, conservatism, smoothing, and direct translation into prices. The problem isn’t the bias; it’s the lack of coordination.
Our position: A forecast is valid only at a given price. The pricing and promotional plan must be a written assumption before the figure is finalized, and it is better to base decisions on outlined scenarios rather than on a single figure.
How much historical data is needed for forecasting? It depends on the category, but reliability matters more than depth: historical data in which breaks, promotions, and price changes are dated is better than a long, unprocessed data set.
Two factors are often overlooked: the history of prices paid—without which elasticity cannot be measured—and the years of inflation, as reflected in volume figures. The adjustment begins with the items that account for the bulk of sales.
The budget answers the question, “How much do we want to sell?” while the sales forecast answers, “How much are we likely to sell?” The former is an annual commitment; the latter is an estimate revised weekly: merging them undermines the value of both.
Pricing falls somewhere in between, with its own timeline: the price revision cycle. Our position: the price is set based on the forecast and evaluated against the budget, and it is never used to bridge the gap between the two.
Five features distinguish a pricing tool that is actually used from just another tool: price and margin analysis, elasticity, simulation, anomaly detection, and explainability. In France, native compliance with the “ below-cost resale threshold ” and reference pricing regulations is also a key factor.
Explainability is key to adoption: a recommendation that a product manager doesn't understand is quietly corrected. The best test before signing off is to have the system explain a recommendation generated from your own data.
Price collection is the goal; web scraping is the method that automates it on a large scale. A price collection is only useful if it specifies which price it captures: store-specific price, card price, promotional price, or regular price.
When it comes to pricing, the GDPR is rarely the real legal issue. The limitations stem primarily from websites’ terms of use, database law, and the load placed on servers—as the CNIL itself points out.
End-of-line inventory and dormant inventory tie up the same amount of capital, but the latter is unpredictable and gets diluted into the catalog average. The early warning sign is the Stock coverage, when compared to the remaining time until sale, rather than the number of weeks without a sale.
Before removing an item from the shelves, a gradual price reduction serves as a diagnostic tool. Since the AGEC law took effect, the destruction of unsold non-food items has been prohibited: the later the issue is detected, the fewer opportunities there are to sell the items at a profit.
A markdown policy combines two decisions—timing and depth—which must follow the same rule. Fixed dates are used for implementation, but it is the price difference at sell-through rate that determines which items to mark down and by how much.
Depth is calibrated by reference, never by a universal percentage grid, and the schedule must account for sales and the reference price over a 30-day period. A grid is evaluated based on the proportion of inventory that remains at the final level.
A markdown is a permanent price reduction and should not be treated as a promotion. It is triggered based on projected remaining inventory—which can be determined as early as the first few weeks of sales—rather than on the start date of the sales event.
The appropriate markdown strikes a balance between profit margin and " sell-through rate " by factoring in the cost of holding inventory and the residual value of the remaining stock. It is determined on a per-item and per-store basis, based on an acceptable residual value set in advance.
Cross-elasticity measures the effect of one product’s price on the sales of another. A positive value indicates cannibalization; a negative value indicates a halo effect. A price reduction that boosts sales of a particular item may reduce the department’s overall margin, because part of the sales volume is simply shifting to another product line.
It is not symmetrical: the effect of A on B is not the same as that of B on A. The halo effect, on the other hand, is better measured by the cross-purchase rate in receipts than by a regression analysis, and the decision is always made at the margin of the category.
A pricing tool centralizes sales, margins, and competition data to recommend, simulate, and sometimes execute pricing decisions, replacing manual management with large-scale analysis, pre-decision simulation, and full traceability.
There are several categories (analytics, AI optimization, competitive intelligence): clearly identifying which one meets your needs helps you avoid buying more—or less—than you actually need.
Having the same item listed under multiple SKUs distorts pricing, inventory, and product assortment. The least visible cost is elasticity: when a product’s SKU and price change, each product record has only ever had one price, and no variation can be measured.
Not all duplicates are merged. True data entry duplicates are merged, alternative formats are linked, and variants are grouped together: when in doubt, it’s better to link than to merge.