Most predictive-analytics content is written for readers with millions of monthly events. Nepali SMEs typically have hundreds — and the entire vendor conversation changes at that scale. This post covers what actually works here: GA4’s built-in predictions where thresholds allow, rules-based scoring everywhere else, churn signals sized to real lists, and an honest account of when enterprise tooling is overkill. For the measurement foundations, see Google Analytics in Nepal; for the CRM these scores live in, the CRM comparison.

1. The volume prerequisite nobody advertises

Statistical prediction needs examples to learn from. Practical floors:

• GA4 predictive metrics (purchase probability, churn probability, predicted revenue) activate only after your property clears Google’s volume thresholds sustained over 28 days. Below that they simply do not appear — no setting unlocks them early.
• Custom models need thousands of labeled outcomes per segment. A clinic with 40 monthly bookings cannot train churn prediction; it can barely A/B test.
• Rules-based scoring has no minimum volume and full transparency — which is why it wins below the floors above.

Start every engagement by checking which tier your volume supports. Selling models below their data floor is the industry’s most common analytics fraud, usually unintentional.

2. GA4 predictive metrics (use what’s free first)

Where thresholds clear: purchase-probability and churn-probability audiences exported to Google Ads for bid adjustments and remarketing, predicted revenue for budget planning. Documented Google functionality — setup is property settings plus patience, not consulting mystique. Review quarterly whether thresholds still clear; seasonal businesses cross in and out.

3. Rules-based lead scoring that beats black boxes

For the sub-threshold majority, transparent points:

• Fit (who they are): district serviceability, budget band stated, decision timeline — from form fields and chat answers.
• Behavior (what they did): pricing page visited, WhatsApp engaged, return visit within 7 days, configurator/quiz completed.
• Decay: subtract on inactivity; a 90-day-cold “hot lead” is a cold lead with seniority.
• Publish the rubric to sales. Scores nobody understands get ignored; transparent scores get followed and improved.

4. Churn signals sized to Nepali lists

Without model infrastructure, watch: login/usage gaps vs personal baseline (not averages), support-ticket sentiment shift, payment failures and plan downgrades, NPS detractors with sudden silence. Trigger human outreach, not automated win-back journeys, below a few thousand accounts — the personal call outperforms the drip at this scale, and it generates the labeled outcomes future models will need.

5. Evaluating enterprise AI/CDP pitches honestly

Ask vendors: minimum monthly volumes for their models to function (in writing), what happens below threshold (rules fallback or garbage scores?), data residency and export terms, total cost including implementation and the analyst to run it, and two reference customers at your volume — not their enterprise logos. Most Nepali SMEs fail at least two of these gates; the right answer is usually rules + CRM discipline now, platform later. See n8n vs Zapier TCO for the same buy-vs-build discipline applied to automation.

What to do this week

1. Check whether your GA4 property clears predictive thresholds; if yes, activate and export one audience.
2. Write the 10-line rules rubric with sales and start scoring manually — one week of manual scoring teaches more than any demo.
3. Price one vendor quote against the rules-based alternative fully loaded; keep the comparison on file.

Related reading: Google Analytics in Nepal, CRM comparison, COD attribution, marketing analytics, and hire a digital marketing expert.