Promotional effectiveness is measured against what actually ran, in which stores, on which days. A feed that reads promotions three days late does not measure effectiveness — it measures something adjacent to it, with an error nobody can quantify.
The source is named because it is public; the client is not. Every figure on this page is a measured production number the client agreed to publish. Have a source of your own? Send it over, whatever its size — you get an answer within 24 hours.
SKU-store rows a year
rows a day, every store
measured coverage
stores covered
Not more rows. Every one of these projects started with somebody who already had data and could not use it for the decision in front of them.
The problem
The client runs national promotions through a retailer and had no independent way to verify what ran where. Compliance reporting came from the retailer, arrived weeks later and was aggregated, so a promotion that failed to appear in four hundred stores looked, in the report, like a promotion that had underperformed.
Fixed on day one
Every one of these is measured continuously and reported on the same dashboard the client watches. A commitment nobody measures is a sentence in a proposal.
None of these is solved by better headers or a bigger proxy pool. Each needs a different piece of engineering, and working out which one you are actually facing is most of the job.
Promotion execution reported by the party executing it is not verification, and the aggregation hides the failures inside the average.
What we doIndependent per-store observation of the promotional state, delivered daily, so execution can be measured store by store rather than accepted in aggregate.
Promotions start and end on their own schedule, and a read one day late misattributes the effect to the wrong price.
What we doThe promotional surface is swept ahead of the main refresh so the file reflects what is live that morning, and start and end are recorded as dated events.
Out of stock, not ranged in this store, and read failure are entirely different facts, and collapsing them into a null makes promotional lift impossible to compute correctly.
What we doEach is classified explicitly with the evidence retained, and the classification is audited weekly against manual checks in a sample of stores.
The estate multiplied by the tracked assortment is a large matrix to sweep inside an overnight window without exceeding the load the source should carry.
What we doPer-store assortment models keep the sweep to what is actually ranged, and scheduling puts the slow tail first so the window is met.
The architecture is not interesting; every extraction system has a queue, a fetcher and a parser. These are the decisions that made this one work where the last one did not.
The whole point is that the observation does not come from the party being measured. Everything else in this pipeline is in service of that.
The promotional surface is read ahead of the main refresh, and start and end dates are recorded as events, so lift is computed against the price that was actually live.
Out of stock, not ranged and read failure are three fields. Weekly manual audits in a sample of stores keep the classification honest.
Sweeping the full assortment against every store is the naive plan. Learned per-store ranges are what make an overnight national sweep possible.
Scheduling is driven by the client's morning run, with the slowest stores started first rather than left until the end.
Production figures, not a benchmark run. Coverage is recounted daily against an independent sample of the live source rather than asserted, which is why the numbers are not round.
Daily profile
Twenty-seven million rows a day is 810 million a month, 9.9 billion a year and 20 billion since 2024, from 1.8 million requests at 21 a second. Inside that, 130 000 promotional changes a day — 47 million a year — is the layer the client had never seen: six hundred and forty execution failures a week, none of them visible in the aggregated compliance report they had been accepting.
Headline
These four sit in the support agreement. When one of them drifts outside its band, we are alerted within fifteen minutes and fixing it is routine work under the monthly arrangement, not a change request.
The columns are the client's own numbers from before the rebuild and the measured ones from production today. The left column is the part most vendors would rather not put on a page.
| Metric | Before | Today |
|---|---|---|
Promotion verification |
Retailer's own report, aggregated | Independent, per store, daily |
Timing |
Up to a week late | Same day |
Availability signal |
Absent | Three states, audited weekly |
Lift calculation |
Against assumed prices | Against observed live prices |
The first quarter's data changed how the client negotiated their next trade agreement. That is a commercial outcome, and it came from an availability field.
One case file per source, each opening the same way: what the client needed, what the source put in the way, how the pipeline is built and what the numbers are today.
2 800 stores, loyalty and shelf priced separately
Read the case →
580k merchants, the fee stack per address
Read the case →
420k vehicles, availability by the hour
Read the case →
Half a million events on sale, priced continuously
Read the case →
135M properties, and the whole live market daily
Read the case →
Start here
Send the source, the fields you need and roughly how often. You get a straight answer within 24 hours: whether it can be done, what makes it hard, what coverage is achievable and roughly what it costs to build and to run. Whatever the size.
Direct
NDA before technical detail, as always. If it isn't our kind of work, we say so in the first reply.