Our Intent model: Purpose built for eCommerce

We hold a single intent model purpose built for eCommerce that holds a contextual layer that fine-tunes to the context of your website.

It is purpose-built from the ground up to address the unique challenges of e-commerce: volatile user journeys, sparse signals, and contextual relevance. The LEM combines large-scale pre-training with on-site fine-tuning, adapting continuously to each retailer’s specific site behaviour while benefiting from a shared intelligence layer drawn from a vast corpus of retail data.

This is a deliberate, superior choice that has been iterated and tested on over 145+ retailers over the past three years. Why? With this, you benefit from the collective learning of one of the largest specific ecommerce customer behavioural models in the world and the precision of your own site context.

We have tested custom models previously and they don’t perform as well as our now global model, trained on pure eCommerce behaviours, fine-tuned to your website.

We can split our model into two functions:

Our LEM (large event model) is built from (as of Oct 25) over 50 billion of eCommerce events across various categories, refined by years of observing what truly signals intent versus what’s just noise. This has created one of the largest specific ecommerce customer behavioural models in the world giving you comparison against other brands.

This means you can be up and running on day one. That foundation gives every new brand a head start, day-one predictive power, not week-one training data. It already “understands” ecommerce patterns before seeing your data.

The models contextual layer then fine-tunes the model to your site’s nuances, without throwing away that global knowledge.

🧠 … This combination matters. A model trained solely on your site’s data quickly overfits to its quirks and blind spots, it only learns what your visitors already do. We know this because we started here, too, three years ago. We tested this multiple times. We found:

Each site must collect enough data, choose the right features, train, and continuously validate its own predictions.

By contrast, our model has already seen billions of real ecommerce journeys, so it recognises patterns your site alone would never surface. It’s the difference between training a model in isolation versus training it on the collective experience of the entire industry. That global foundation prevents misbehaviour, accelerates learning, and makes predictions more resilient to change, reflective of the market (trends, patterns, seasonality) while the contextual layer ensures they’re still uniquely yours.

We run a live feed, not a snapshot. Our model reads intent as it evolves, giving your brand seconds, not pages, to respond.

We predict intent on every event, analysing over 800 signals per event, every few seconds, not once per page view. That cadence means you’re never more than a few seconds away from knowing a visitor’s current intent and being able to act on it. This is capability comes from our real-time inference architecture feeding into our Intent LEM model purpose build for eCommerce. Because of our unique approach to a universal deep learning neural network fine-tuned with your sites context, opposed to just the latter, it means that we’re high-resolution, low-latency, high-accuracy and infrastructure costs are low.

Bespoke-model predictions can’t operate at this frequency so usually score on page load - missing key signals, feelings and response that happen within a page. Such comparison is inadequate, solving a simpler, slower problem as updates per page miss those critical moments. In short, we know that user behaviour can easily and quickly change within a single page; the micro-behaviours on page that reveal intent shifts.

We can achieve this, because our intent model is purpose-built for real-time ecommerce, comprising of over 50bn events culminated from behaviours of over 145 retailers.