Retail & Consumer Goods

The same jacket has one number at the register, a second in the online shop and a third in buying. Stores are grouped by region in one system and by sales banner in another. And when someone asks which channel earns the most gross margin, everyone calculates it differently and everyone is a little bit right.

We build the data foundation that reports, dashboards and forecasts run on for retailers and consumer goods companies. AI and machine learning come in where they prove themselves on the company’s own data.

Matching items and locations

The technology is rarely the problem. Registers, shop systems and marketplaces have interfaces or deliver exports. The effort goes into matching: which register item number belongs to which shop variant, how are sizes and colours represented, which store belongs to which region? Where a well-kept PIM exists, that is a good starting point. Where it does not, you end up with a mapping table that someone in buying has to maintain, otherwise it is out of date again after three months.

Only once this matching is in place is a channel comparison more than an estimate. That is why we start here and not with the dashboard.

What revenue and gross margin mean

Finance, buying and the online shop often quote the same figure and mean different things: gross or net, before or after returns, by order, delivery or invoice date. For gross margin there is the further question of whether markdowns, supplier terms and bonuses have already been deducted. Questions like these decide whether a store looks good per square foot or only on paper.

We agree each of these definitions once with the people involved and store it in the data model. Every report calculates from it, including the weekly reports for store managers that someone currently builds by hand in a spreadsheet.

Evaluating promotions

“The promotion brought in revenue” is quickly said. What gets measured instead is the sales you would have expected without the promotion, minus pull-forward effects, shifts to other items and the discount’s effect on margin. That only works if the period, discount and placement of every promotion were recorded. In practice that is where it fails more often than on method.

A legal duty fits in here and is worth using for analysis too. Anyone advertising a price reduction has to state the lowest price of the last 30 days under the German Price Indication Ordinance (Preisangabenverordnung). A price history per item has to exist anyway, and it makes promotion analysis more reliable as well.

Forecasting demand without fooling yourself

A forecast estimates sales by item and location from history, season, price, promotions and calendar, and turns that into order suggestions. For items with enough history this works well. For novelties, special lots and erratic demand it is less certain, and the suggestion should say so.

There is a trap in the data itself. Where goods were missing, nothing was sold, and a model that learns from that underestimates demand. Stockouts therefore need to be flagged in the history. Before go-live we test the suggestions on periods the model has not seen and compare them with the current approach. In operation, buyers and planners stay responsible. They know about the supply bottleneck and the weather forecast that appear in no table.

We treat generated product texts from PIM data just as soberly: spot checks before anything goes live.

Online shop, returns and data protection

For returns, the average tells you little. Broken down by item, size or supplier, it shows where cut, description or quality is the problem. Since 19 June 2026, online retailers must offer a withdrawal button for consumer contracts. Withdrawals and returns should be recorded separately in the data, otherwise two different processes get mixed together.

We measure behaviour in the shop with server-side tracking so that it can be linked to orders and returns. In Germany, the TDDDG applies to cookies and similar access to the device, and the GDPR applies to personal data. That includes loyalty cards and customer accounts. Many analyses, such as assortment and basket, work with pseudonymized or aggregated data. The changes to these rules proposed in the Digital Omnibus are so far only a proposal and have not been adopted.