Predictive marketing, forecasts not just reports.

What do predictive marketing models predict?

The aim is spend that follows real value. Experiments at eBay found that “returns from paid search are a fraction of conventional non-experimental estimates”, in Blake, Nosko and Tadelis (NBER), which is why models are validated on held-out data before they move budget.

• Customer lifetime value: the expected revenue from a new customer over a set period, so acquisition spend matches what customers are actually worth.
• Conversion probability: how likely a lead or visitor is to convert within a window, to prioritise follow-up, bidding and retargeting.
• Channel forecasts: expected performance under different budgets and seasons, such as Ramadan and the summer slowdown.

How are the models built and validated?

1. Data audit (week 1): conversion tracking, customer identifiers, CRM quality and how much history exists. Gaps get fixed first.
2. Features (weeks 2–3): raw data turned into signals, such as time to first purchase or depth of engagement.
3. Training and validation (weeks 4–6): several model types tested on held-out data against business metrics, not just accuracy.
4. Advisory mode (weeks 7–8): predictions inform people’s decisions before anything is automated.
5. Monitoring: drift is watched and models retrained on a schedule that matches how fast your market moves.

“Eight in ten respondents say AI has improved their own productivity.”

McKinsey, The state of AI (2026 survey)

Where predictions go

• Ad platforms: value-based audiences and conversion signals for Meta, Google and TikTok.
• CRM: scores on each record, triggering follow-up.
• Dashboards: for budget and retention planning.
• Decision lists: customers at risk, segments under-funded, campaigns to pause.

What can’t predictive models do?

Treat outputs as ranges. Only 37% of organisations in McKinsey’s 2026 State of AI survey attributed any EBIT impact to AI; value comes from acting on predictions, not producing them.

• Not a crystal ball: outputs are probabilities with ranges.
• Not a strategy: it predicts within your current approach.
• Not a fix for bad data: tracking comes first.
• Not unsupervised: people stay in charge, especially when markets shift.