“Sounds complicated” is the most common objection we hear before a first audit. The assumption is fair: “predictive model” and “BigQuery” sound like a dedicated engineering sprint. Deployment usually needs no sprint, and the precise requirements below show why.

What’s required

Five things, all listed on the Foresight product page:

  • Google Analytics 4, active on the site with a collection history. A fresh install won’t do, because the model learns from past navigation data.
  • Google Cloud Platform access, meaning a GCP project that can receive the BigQuery export. If you don’t have one, setup takes a day and involves no procurement process.
  • Google Tag Manager, installed or installable. The preload tag lives there. If you don’t use GTM yet, this is where that conversation starts.
  • Enough traffic for the model to see patterns repeat. No hard minimum exists. A low-traffic site with repetitive navigation can work, and a high-traffic site with unpredictable paths may need more history than a smaller, more predictable one.
  • Admin access to the above, for whoever does the initial setup. You need it once.

That’s the full list. You install no new SDK, rebuild no mobile app, and add no consent banner beyond what your GA4 setup requires, because Foresight trains on data you already collect.

Where deployments get stuck

The model and the GTM tag rarely cause delays. The first prerequisite does: the GA4-to-BigQuery export doesn’t exist yet, or it exists but someone configured it for a different purpose and it lacks fields the model needs, or admin access to the GCP project sits in an internal approval process unrelated to the technical work.

Any project in a larger organization meets the same friction when it needs access, so plan for it. Once access exists, the technical build moves fast.

Timeline

If the GA4-to-BigQuery export already exists with a few months of history, expect model building and GTM integration to take about two weeks, followed by a reporting setup that is mostly configuration.

If the export doesn’t exist yet, add about a week upfront. The model needs real navigation data to train on, and you can’t compress “collect a week of traffic” into less time.

What “lightweight” leaves out

A lightweight deployment still needs you. Someone grants GCP and GTM access, someone signs off on the CNIL/consent posture if your organization has that question open (see our piece on privacy), and someone reads the reporting dashboard once it goes live, so it doesn’t join the pile of unopened monthly reports. The technical lift stays small. Getting the right people to say yes drives the timeline.

An audit tells you where your own setup lands: fast track or one-week wait.