Instamart, India's Leading On-Demand Grocery Delivery Platform from Swiggy, Cut Cost Per Acquisition (CPA) by 51% with Signal Engineering

Predict who will become a regular at the first order

Each part of Instamart's business runs in its own growth context, with different customers, competitors, and goals. A single model couldn't serve both, so Voyantis built one for each, shaped around the question that mattered for its side of the business.

1. Instamart's model answers a question of magnitude. How much will a customer be worth? It predicts their LTV after thirty days.
2. Food delivery's model answers a behavioral question. Will a new customer become a regular? It forecasts propensity to place three orders within thirty days, the threshold that marks a forming Instamart delivery habit.

Both models read that early behavior, but each learned to weigh it based on thousands of different contextual and behavioral insights to capture each business nuance, including:

• Whether they join Swiggy One, and how fast
• How much they lean on discounts for food delivery
• What they build into a basket on Instamart delivery

The strongest early patterns weren't always the obvious ones. For example, commitment showed up before volume did. In the first hours after acquisition, how quickly a customer joined Swiggy One told the Food model more than how much they had spent. The full picture only emerges when these insights are weighed together, for every individual customer.

The result was a dependable read on who would become a regular, and by how much, so the auction could bid accordingly rather than treat every new customer as the same average bet.

Engineer the prediction into an auction-ready signal

Google's auction bids in currency and takes events given at face value. A prediction, which is always evolving, is treated as ground truth, steering the algorithms based on an incomplete view of customer value.

To bridge the gap, Voyantis engineered the model's scores into a signal fit to Swiggy's growth context and the learning logic of Google's algorithm:

• Timing: A prediction is generated at the moment of a customer's first order and delivered as a signal within hours – as soon as enough behavior has accumulated – to anchor Google's learning.
• Cadence: As the customer takes more action over the following days, the prediction sharpens and enhanced signals follow, when the change is meaningful enough.
• Value: For Instamart, every prediction is capped so the signals delivered to the auction keep a full spread of values to learn from, instead of piling budget onto a few inflated scores. For food delivery, the signal blends revenue earned from the first order with a forward estimate of the next 30 days.

That engineering, not the prediction, is what sets Voyantis apart from a capable in-house model. Both models could tell who was likely to keep ordering, but only one could turn that knowledge into a signal that moves the auction toward performance.

Activate in a live auction, market-by-market

To prove that engineered signals would drive better performance than a raw prediction, Voyantis ran a head-to-head for each arm of the business – one powered by its own predictive signals and the other by Swiggy's in-house model.

Instamart's business is significantly different city by city, so its test was built as a geographic split, with a matched set of markets running Voyantis signals against a matched set running the internal model at the same time. This isolated the signal's effect, without one market contaminating another. Food delivery ran the same design as a separate test.

Success was measured against Swiggy's own numbers, not Google's reported conversions. That distinction matters because Google's reporting runs on its own attribution and doesn't carry Swiggy's real KPIs, the revenue and retention that show whether an acquired customer was worth it.

Learn and evolve to keep the signal pointed at the right customers

Left alone, a signal drifts. The customers it wins today become the data the model learns from tomorrow. Meanwhile, Swiggy's business – its pricing, products, and markets – keeps moving underneath it. To keep the model accurate and the signals chasing the right customers, Voyantis has 24/7 monitoring across three fronts:

• Pipeline Integrity: Autonomous monitoring flags when data pipelines break before the issue impacts performance.
• Model Health: Predictions are checked against real outcomes as cohorts mature and the model is retrained as needed.
• Learning Preservation: Changes are made gradually so Google's accumulated learning is always preserved and never reset.