Recommendation Engine

Cold start, which is where most engines fail

Collaborative filtering needs history. A first-time visitor has none, and a new product has never been bought by anyone.

For visitors, session behaviour fills the gap — what they searched, what they viewed, how long they stayed on each. Within three or four page views there is enough signal to personalise meaningfully, which covers most of a first session.

For products, content-based similarity carries them until behavioural data accumulates. A new item with good attributes gets recommended alongside similar items from day one rather than sitting invisible for two months, which is what happens with a purely behavioural engine.

Where recommendations belong

The same engine performs very differently depending on where its output appears, and placement decisions are worth as much attention as the model.

• Product page, below the fold. Complementary items rather than substitutes — recommending three similar sofas to someone looking at a sofa mostly creates hesitation.
• Cart page. Genuine add-ons, small enough to be an easy yes.
• Home page for returning visitors. Continue where they left off, which is the single highest-performing placement in most deployments.
• Email. Personalised selections outperform batch campaigns substantially and cost nothing extra to send.
• Post-purchase. Often the best-converting placement and the most neglected one.

Business rules on top of the model

A pure similarity model will happily recommend out-of-stock items, products from a discontinued line, or an accessory that costs more than the item it accompanies.

Real deployments need a rules layer: exclude out of stock, respect category rules, avoid recommending something the customer bought last week unless it is consumable, cap how often any single product appears, and honour any merchandising priorities.

We build this as configuration your merchandising team controls, not code. The rules change with the season and with stock position, and a system that needs a developer for that becomes stale quickly.

Privacy and what you should not do

Recommendation depends on behavioural data, which under the DPDP Act needs a lawful basis, a stated purpose and reasonable retention. In practice this means telling visitors that browsing is used to personalise, offering a way to opt out, and not retaining individual behaviour indefinitely.

There are also recommendations that are technically possible and commercially unwise. Inferring health conditions, pregnancy, financial distress or religious observance from browsing and acting on it visibly makes customers uncomfortable even when it is accurate — particularly when the device is shared, as many Indian household devices are.

We build in category sensitivity rules for exactly this reason. The engine can know something and still not display it on a home page.

Build scope, cost and rollout

Eight to fourteen weeks from ₹3,50,000: event tracking, the hybrid engine, business rules layer, placement integration, the incrementality measurement framework and reporting.

You receive the engine on your infrastructure, the rules interface, the measurement dashboard and full source. Running costs are largely infrastructure rather than model usage and are modest at typical traffic.

We would deliberately start with two placements measured properly rather than eight placed everywhere. Proving lift on one page and then extending is both faster to value and easier to defend internally.