“Predictive preloading” can sound like a black box: you hand over your data and hope. You should be able to point at each stage and know what happens to your data and why. This is the sequence from the Foresight product page, with the reasoning behind each step.

Step 1: Data collection

Before we build anything, we check whether GA4 exports to BigQuery and whether the export holds enough history to be useful. “Enough” depends on your traffic volume and the variety of your navigation paths. As a rule of thumb, the model has to see a pattern repeat often enough to trust it.

If the export doesn’t exist, we set it up and wait about a week while GA4 fills BigQuery with real sessions. No shortcut exists. A model trained on three days of data learns whatever was unusual about those three days.

Step 2: ML model building

Here we turn the navigation data into a predictive model. For a given visitor context (entry page, traffic source, device, and the pages already visited this session), the model estimates the probability of each possible next page.

We train the model on your own navigation patterns. Next-page behavior varies by site: a media property’s looks nothing like an e-commerce category page’s. A model that has never seen your structure guesses. A model trained on your data matches patterns against precedent.

Step 3: Google Tag Manager integration

A prediction helps only if something acts on it. We deploy a preload tag through GTM, with no new SDK and no site rebuild. The tag checks the model’s top prediction for each visitor and preloads that content when confidence is high enough to justify the bandwidth.

GTM already runs on most of these sites, and the tag reads the prediction and acts on it without touching your application code. We designed around that constraint because the approach only makes sense if shipping it takes no engineering sprint.

Step 4: Reporting

The last step gives you a dashboard in your own Looker Studio that tracks average LCP, conversion rate, and the percentage of sessions with a successful preload. It updates continuously and is not a one-time report. We also build the before/after comparison from the business case piece here, with your own numbers instead of a typical-deployment estimate.

What this means for you

None of the four steps touches your site’s core codebase, and none needs a consent flow beyond what your GA4 setup already handles. Step 1 is the heaviest lift when the BigQuery export doesn’t exist yet. The later steps move fast because they use data you already collect.

An audit shows where your own setup would start in this sequence, usually in the first conversation.