Articles

  • What we find when we build retail data and AI systems on real data - the numbers, the queries, and the parts that do not work.

AI agents in retail

What AI agents can and cannot automate in a retail business, written from two we built and run ourselves: a conversational analytics assistant over a governed warehouse, and a voice and chat agent that answers a business's phone. Where we quote a number, it is measured on our own systems.

Part 1

Nine jobs a retailer can hand to an AI agent today

A buying list rather than a theory: what each one replaces, what it needs from you first, how you would prove it works — and the one question that disqualifies most projects.

Part 2

Let anyone in the business ask the data a question and get an answer in twenty seconds

What it displaces is not analyst hours — it is the questions nobody asks because the answer would arrive too late. One right answer in ten became nine, with no change of model.

Part 3

Every call you miss, answered — and the rule that makes it safe to let an agent act

The one channel where failure leaves no trace. Plus what has to be true before software does anything irreversible: enumerated actions, filters rather than instructions, a human always reachable.

Part 4

What we would not sell a retailer yet

The jobs where the economics or the guardrails are not there — autonomous pricing, stock commitments, anything that writes to the live catalogue unattended. Coming soon.

Recommendation systems for retail

Building a recommendation engine from scratch on Google's own store data - the public GA4 export of the Google Merchandise Store - and publishing every result, including the ones that did not work. The question behind the series: does a recommendation engine add 2% or 12% to a shop's revenue, and how would you know?

Part 1

Your analytics data is lying to your recommender

Product views over-counted 13.6 times, nine product IDs for one sweatshirt, and a crawler responsible for 29% of all product-page views. What has to be fixed before any model is trained.

Part 2

A recommendation engine written in 30 lines of SQL beat the bestseller list by 60%

Five models on the same held-out January sessions. The winner has no training step, the textbook ranking loses fifteen points, and the purchase-only model comes last.

Part 3

Your recommendation engine cannot surface a new product. Four things that fix it

For products with no history the model returns an empty list every time — 39 of the 418 on sale. Why nothing in a sales report reveals it, and the ladder of fixes from bestsellers to inheriting a similar product's statistics.

Part 4

Can a new product inherit a similar product's recommendations? We tested it

The fix we proposed in part 3, measured on 2,453 held-out sessions: it recovered an eighth of the gap. The matching was not at fault — similar products simply are not browsed together.

Customer lifetime value

Which customers are worth paying for? Worked through on two years of real order history from a UK online retailer, ending with the part that changes a budget: handing a predicted value back to the ad platform that bids on your behalf.

Part 1

Your customer acquisition cost is judged against the wrong number

The top 10% of customers produced 64% of two-year revenue, the first order was 14% of a customer's value, and first-order value identified one in three of the customers who ended up mattering.

Part 2

Customer value becomes predictable after 90 days, not after the first order

How long a retailer has to wait before an acquired customer stops being a guess: 42% of the eventual top decile visible from the first order, 67% after 90 days, and where the curve flattens.

Part 3

Teaching Google Ads what a good customer is — and when you can actually know it

Value-based bidding works on the number you send it. At every moment inside the upload window a trained model matched a plain sort; the accuracy only arrives at 90 days, by which point the click identifier has expired.

Part 4

Google Analytics already predicts churn. When is that enough?

Free predictive audiences against a model on your own warehouse, and the blind spots that decide between them. Coming soon.

Every figure in these articles is reproducible from public data: the ga4_obfuscated_sample_ecommerce dataset published by Google through the Cloud Public Datasets Program, and UCI Online Retail II (Chen 2012, CC BY 4.0). The queries are in the articles.

Start Your Data Transformation Journey

Discover practical, scalable solutions tailored to your business priorities.

Book Your Free Consultation!