Predictions use cases

These examples show how businesses use predictions to run smarter campaigns. Each use case shows how a single prediction model drives real decisions — on budget, timing, incentives, and channel.

Churn-driven retention with tiered intervention (all verticals)

Not every at-risk customer needs the same treatment. Sending a deep discount to a customer who was about to leave can recover the relationship. But sending that same discount to someone who would have returned anyway is wasted margin.

Churn prediction lets you separate these groups and respond to each one with the right level of effort.

How it works

1. Set up a churn prediction with the activity event for your business. Use purchase for retail, session_start for media and subscription, subscription_renewed for SaaS, or app_open for mobile-first. Any custom event with static filters also works.
2. Run a daily scenario that checks for customers crossing churn probability thresholds.
3. Apply a tiered treatment based on the probability score.

Live example

This is the most deployed prediction use case in the Bloomreach customer base. The key advantage over rule-based approaches is timing. Triggering re-engagement at the precise moment a customer crosses the churn threshold beats any fixed inactivity window.

Recommended channels: Email, SMS, push

KPIs: Retention rate, reactivation rate, revenue recovered per campaign

Channel-specific optimal send time (all verticals)

Customers don't behave the same way on every channel. A customer's best time to read an email might be morning. Their best time to act on a push notification might be evening. Treating these as the same wastes the send time optimization.

How it works

1. In any scenario, add the Wait node, select optimal send time, and select the matching channel:
   • Email for email campaigns
   • SMS/MMS/RCS for text-based campaigns
   • Push for mobile push notifications
   • All channels only if a customer has limited per-channel history
2. Choose a send strategy:
   • Send within a time window: Set a maximum wait — for example, 12 hours. Optimal send time picks the best time within that window.
   • Send before a local-time deadline: Set a cutoff in the customer's local timezone — for example, 20:00. Optimal send time sends at the best time before that deadline. Enable Use customer profile time zone for global audiences.
3. Optionally, use the optimal send time prediction template. This generates a per-channel send-time attribute on each customer profile. You can use it in segments, reports, or custom scenario logic.

Live example

Our customer tested channel-specific optimal send time for their weekly push blasts across 3 weeks and 6 campaigns:

• 10% uplift in unique conversions (100% statistical significance)
• 13% uplift in revenue per visitor (100% statistical significance)

Previously, optimal send time used aggregate data across all channels. A customer's best email time and best push time were treated as one. Now the model trains per channel, so each channel reflects its own behavioral patterns.

Recommended channels: Email, SMS, push

KPIs: Conversion rate, revenue per visitor, click-through rate

Purchase probability-driven ad spend optimization (retail, fashion, ecommerce)

Retargeting ads work by nudging high-intent customers to complete a purchase. The problem is that customers with very high purchase probability don't need nudging — they'll buy anyway. Every ad served to them wastes the budget.

The customers who benefit most from a retargeting ad are in the middle. They're interested but not yet decided.

How it works

1. Set up a purchase prediction with a window that fits your business cycle. Use 7 days for fast-moving retail, 30–90 days for fashion, or 90–365 days for high-ticket items.
2. Segment customers by purchase probability: high, medium, and low.
3. Route each segment to a different paid media strategy.

Live example

American Eagle Outfitters built affinity segments for purchase likelihood within a specific category and ran targeted campaigns. The result: +33% lift in category revenue, equal to $11,400 per month or $135,000 annualized.

Recommended channels: Ad networks (Meta, Google), email

KPIs: Return on ad spend, incremental conversions, cost per acquisition

Budget-conscious discount strategy (retail, fashion, travel)

Blanket discount campaigns don't protect margins. If 30% of your high-probability customers would have bought without an incentive, every discount you send them is pure margin loss.

Purchase prediction helps you target discounts at the customers who need a reason to convert — and spare the ones who don't.

How it works

1. Score all customers on purchase probability.
2. In any discount or promotion campaign, add a branch in your scenario based on the prediction score.

Variant — first-to-second purchase acceleration: Apply the same logic to first-time buyers. Run a purchase prediction for the next 30 days on customers with exactly one purchase. Tier the treatment the same way. Align the prediction window with your repurchase cycle: 7–14 days for consumables, 30–90 days for fashion, or 90+ days for high-ticket items.

Live example

Our customer tested different voucher amounts across purchase probability segments. The medium-probability segment showed the highest incremental lift from a medium incentive (23.4% conversion rate versus 10.7% in the control group). This confirmed that high-probability customers convert without discounts.

Solmar Villas applied the same approach to their promotional strategy — estimated at £756,000 in additional revenue by redirecting discounts to where they actually move the needle.

Recommended channels: Email, SMS

KPIs: Revenue per email, margin saved per campaign, conversion rate by segment

Prediction-powered cross-sell (retail, transport, multi-category businesses)

Winning a second category from an existing customer is cheaper than acquiring a new one. Most cross-sell campaigns miss this opportunity. They push category recommendations at customers who already buy there — or have no interest in it.

Purchase prediction, filtered to a specific category, finds the customers who are ready to try something new.

How it works

1. Create a purchase prediction with the event filter narrowed to one category — for example, purchase where category = 'accessories' or purchase where product_line = 'e-bikes'.
2. Target customers with a high cross-category purchase probability who haven't yet bought from that category.
3. Send a targeted campaign with category-specific product recommendations.

Live example

Our customer created a purchase prediction for e-bike rentals. They targeted existing motorbike-only customers. Customers with high prediction scores were 4x more likely to rent an e-bike than those with low or medium scores. Cooltra piloted the approach in 2 towns and converted approximately 100 customers. They're now scaling it across their full customer base.

Recommended channels: Email, push

KPIs: Cross-category conversion rate, category revenue, customer lifetime value

Real-time discount targeting (ecommerce)

Showing the same discount banner to every visitor wastes margin and clutters the experience. A customer who was already going to buy doesn't need a discount — and a customer with no purchase intent won't be moved by one either. In-session prediction scores customers based on their current session behavior — pages viewed, items browsed, time on site — and lets you act on that signal before the session ends.

How it works

1. Set up an in-session prediction model and select, for example, purchase conversion goal to define the event that represents a purchase in your project.
2. Segment customers into three tiers based on their real-time score.
3. Set a different weblayer treatment for each tier.

Recommended channels: Weblayers

KPIs: Session conversion rate, discount redemption rate, revenue per session

A/B test every treatment

The strategies above are starting points based on customer results and Bloomreach Consulting experience. They're not universal rules. Every business has different customer dynamics. What works in one vertical may not work in another.

Run every prediction use case with an A/B test to confirm the treatment delivers incremental value. Bloomreach has a built-in framework for this — see Evaluate predictions.