PriceDesk — case study
A pricing tool that starts from three questions. Product by product, it shows what volume and gross margin did when a price moved. Version of 21 November 2026.
The problem
Prices change every year, some up and some down. Few companies go back, product by product, to see whether each change paid off.
“Price up, volume down” is not the answer. A rise that loses some volume can still earn more gross margin. It depends on the margin the product started with.
Volume also moves for other reasons. When the whole range sells less, a product that fell after a price rise may have done no worse than the rest.
In eight years in the agricultural industry, six of them in global product management, I handled many pricing cases. I wrote down about twenty of them before building anything, and built the tool around a few.
The user
A commercial or pricing lead at a company with many products and many customers, its invoice lines in an export, and no pricing software.
They know their products and their customers. They are not trained in pricing theory, and they have a few minutes.
The test for every screen: someone who opens the site cold gets one useful answer in under two minutes, with nobody there to explain.
The second reader is a pricing expert, who asks “based on what?” of every figure.
What I built
Three questions, each opening one page with the answer on top:
- Price up, price down: what happened to my volume and gross margin?
- Where am I below my minimum margin?
- Which customers pay less than similar ones?
Behind them: an Overview and nine more analyses, from returns and credit notes to prices across countries.
It runs on real public data: 1.04 million sales lines of a UK online giftware retailer, December 2009 to December 2011.
It also runs on your own file: CSV or Excel, five columns, read only in your browser. A sample file comes with five figures to find, so you can check that the tool read it correctly.
Key choices
- Questions before pages. The menu shows three questions. Everything else is one click away, under “More analyses”.
- Break-even as the yardstick. Each price move is set against the volume it could lose, or had to win, before gross margin falls.
- A comparison before a conclusion. Each product is compared with the products whose price did not change.
- Observed, not explained. The screens say what volume and gross margin did. They never say “because”.
- Every amount says what it rests on. Costs on the public data are simulated, and labelled. On your file, without a cost column, every margin carries the word “assumption”.
- Your file stays in your browser. Once you choose a file, the page refuses its own network requests.
- Rules write every figure and every summary. Each formula is written out in Calculations.
What I cut, and why
- Most pages, from the first menu. It showed sixteen. It now shows three questions and the Overview; the rest is folded under “More analyses”, not deleted.
- A second demo on a generated company. It allowed more advanced analyses, but the real data was enough, and the tool was already complex. Its data now makes the sample file.
- AI-written summaries. Rules write them now, so a sentence cannot drift from the number beside it.
- Three large “opportunity” amounts on the first screen. They could not be added together. The Overview now shows what was observed: sales, gross margin, who bought more and who bought less.
- A score against a general AI. I had scored it myself, on a grid drawn from my own rules. A high score was expected, so it proved little.
- Accounts, a database, anything paid. This is a portfolio demo, not a product.
What the real data showed
December 2010 – November 2011 against the year before. € at a stated display rate of 1 £ = 1.15 €. Gross margin = net sales minus the cost of the goods, on simulated costs.
- The 1,287 products whose price did not change sold 23.1 % fewer units. Without that comparison, the tool would have blamed price for a falling year.
- 151 products had a price rise of 5 % or more. Gross margin fell on 81 of them (−€70,416 together) and held or rose on 70 (+€111,215).
- The typical rise cost little volume: a median of −0.3 % of volume for each +1 % of price, beyond the unchanged products.
- One example: a peg bag went up 18.9 % and lost 27.4 % of its volume. With a 29.3 % gross margin it could lose 39.2 % before gross margin fell. It earned €547 more.
- Timing can fake a volume change. 656 products were sold in at least three fewer months in one of the two years; for 217 of them the direction reverses on the months sold in both.
- Sales rose 1.4 %, which hides what moved underneath: customer accounts bought €3.31M more and €3.56M less.
- 1,177 products had sales below a 6 % minimum gross margin, €84,832 short of it. On simulated costs, the pattern counts more than the amount.
- On the best-selling product, sales below the usual price add up to €17,510. Most of it is quantity discounts: 47 of 873 customers pay at least 3 % less than customers who buy similar quantities, and that gap is €833.
From sixteen pages to three questions
By November 2026 there was a lot in the tool. A visitor could not find their way through it alone, so I reworked how it opens.
- I measured before fixing. After “Open the tool” the screen stayed blank for about 25 seconds on a slowed-down test machine. The page was hidden until its slowest part was computed. Showing each part when it is ready cut the blank to about 1 second. The Overview's 14 figures were compared before and after: none changed.
- Seven mock-ups, one menu. I compared seven versions of the menu and kept the one with three questions.
- The headline asks instead of claiming. A first version gave counts: so many price rises, so many margins down. It put the result before the question and covered only rises. It became “Your prices moved. Did it pay off?”, then “Your prices moved. What happened to volume and gross margin?”: the first still hinted that the price had caused the result, the second only asks what the tool can show.
- The tool says what it does not prove. A box lists six things it doesn't do. The first: it does not prove that a price change caused a volume change.
“Open the tool” now opens the first answer instead of the whole Overview: about 7 seconds in the same kind of test. The full Overview still takes about 17 seconds, but the wait is no longer blank.
Limits
- It does not show cause. It compares two years; promotions, competitors and stock-outs also move volume.
- Costs on the public data are simulated. The margin amounts there show the method, not that company's real margin.
- One dataset, from 2009 to 2011.
- No user testing yet.
- On your own file, the three questions and the Overview run. Seven other pages do not yet, and say so.
- Not built for scale. The public data was prepared in advance; an uploaded file is capped at 500,000 lines.
- On a phone, the Margin dashboard scrolls sideways.
How it was built
The code was written with an AI assistant (Claude). I chose the questions, the rules and the thresholds, and decided what shipped.
Python prepared the public data once. JavaScript in the browser computes everything that moves with a filter or a threshold, and reads an uploaded file.
Each formula was checked on two or three cases, by hand or by a separate calculation, before it shipped.
Hosting runs on a free tier.
Built by Alexis De Schrevel. Hosted by Cloudflare, Inc., 101 Townsend St, San Francisco, CA 94107, USA.
Data: Chen, D. (2012), Online Retail II, UCI Machine Learning Repository, doi.org/10.24432/C5CG6D, licensed under CC BY 4.0; cleaned, aggregated and analysed by PriceDesk; the original company is not identified. Costs are simulated, and shown as such in the tool. € at 1 £ = 1.15 € (display only). A portfolio demo — not financial or professional advice.