How do I do next-best-offer ranking in a SaaS app?

Next-best-offer (NBO) ranking in a SaaS app usually means: given a user/account and the current context, choose the best action or offer to show next—upgrade prompt, feature recommendation, discount, onboarding step, add-on, renewal save offer, etc.

A good implementation is usually a mix of rules + machine learning + experimentation.

1) Define the decision clearly

Start by specifying:

• What are the candidate offers?
  Examples: upgrade plan, trial extension, add seat, invite teammate, enable integration, book demo.
• What is the objective?
  Examples: conversion, retention, expansion revenue, activation, reduced churn.
• What is the constraint?
  Example: don’t show more than 1 offer/day, exclude offers already seen, avoid discounting premium accounts.

In SaaS, NBO is often not “highest click probability,” but:

• expected revenue
• expected retention lift
• expected activation lift
• or a weighted blend of these

2) Build the candidate generation layer

Don’t rank every possible offer from scratch. First generate a small eligible set.

Typical filters:

• user segment: role, industry, plan tier
• product usage: features used/not used, frequency
• lifecycle stage: new user, active, dormant, churn-risk
• eligibility rules: account size, region, contract type
• suppression rules: already purchased, recently shown, legal restrictions

Example:

• If user is on Free and uses collaboration heavily, candidate offers might be:
  • upgrade to Pro
  • invite teammate
  • enable shared workspace

3) Choose a scoring approach

There are a few common options.

A. Rule-based ranking

Simple starting point:

• assign score by business priority
• use if you have limited data

Example:

• churn-risk save offer: 90
• upgrade CTA: 70
• feature education: 40

Good for:

• cold start
• low traffic
• strict business control

B. Propensity models

Train a model to predict:

• probability of click
• probability of conversion
• probability of retention
• expected spend

Then rank by something like:

expected value = P(action) × value(action) − cost(action)

Example:

• upgrade offer expected value = P(upgrade) × incremental MRR
• discount offer expected value = P(convert) × margin impact

C. Contextual bandits

Great for NBO because they balance:

• exploitation: show what seems best
• exploration: try alternatives to learn

This is useful when:

• you have many offer types
• user preferences vary a lot
• the environment changes frequently

D. Full recommendation models

If you have lots of data, you can use:

• learning-to-rank
• two-tower retrieval + ranking
• gradient boosted trees / deep models

But for many SaaS teams, propensity model + business rules + experimentation is the sweet spot.

4) Define features

Common features for SaaS NBO:

User/account features

• plan type
• seat count
• ARR / MRR
• company size
• industry
• region
• lifecycle stage
• tenure

Behavioral features

• logins last 7/30 days
• feature adoption
• usage frequency
• recent actions
• abandoned flows
• support tickets
• NPS or health score

Offer features

• offer type
• discount amount
• product area
• expected value
• complexity
• urgency

Context features

• channel: in-app, email, sales call, popup
• device
• time since last prompt
• page/location in app
• campaign/source

5) Decide the ranking formula

A practical formula is:

score(offer, user, context) = uplift × business_value × eligibility_weight − friction

Where:

• uplift = predicted probability improvement from showing the offer
• business_value = revenue, retention, or activation value
• friction = annoyance, support burden, discount cost, or overload

If you only have conversion prediction:

score = P(convert | user, offer, context) × value_of_conversion − cost_of_offer

6) Add constraints and guardrails

This is very important in SaaS.

Examples:

• Don’t show upgrade offers to users who just paid.
• Cap frequency: max 2 offers/week.
• Suppress if user is in a support incident.
• Avoid conflicting offers.
• Respect customer success / sales ownership.
• Exclude enterprise accounts from self-serve discounting.
• Don’t upsell before activation basics are complete.

Guardrails prevent the system from optimizing the wrong thing.

7) Measure with experiments

You need A/B testing or multi-armed bandits to validate. Track:

• CTR
• conversion rate
• incremental revenue
• retention/churn
• activation milestones
• complaint rate / hide rate / unsubscribe rate

The key metric should be incremental lift, not just click-through.

8) Recommended architecture

A simple production architecture:

1. Event pipeline collects product usage events
2. Feature store aggregates account/user metrics
3. Eligibility service filters allowed offers
4. Ranking service scores candidates
5. Decision engine applies rules, constraints, exploration
6. Experiment layer assigns variants
7. Logging records impressions, clicks, conversions, and outcome labels

9) Practical MVP path

If you’re starting from scratch:

Phase 1: Rules

• define segments
• define eligible offers
• hand-rank by business rules

Phase 2: Propensity model

• predict conversion for each offer
• rank by expected value
• add frequency caps

Phase 3: Bandits / uplift

• explore alternatives
• optimize incremental impact, not raw response

Phase 4: Personalization at scale

• separate models by lifecycle stage or offer family
• automate candidate generation
• optimize across channels

10) Common mistakes

• optimizing clicks instead of revenue or retention
• ignoring suppression/frequency rules
• showing offers without eligibility filtering
• training on biased impression data without counterfactual thinking
• not separating “who gets shown” from “what they see”
• using one model for all offer types when behaviors differ

11) Example

Suppose you have these offers:

• Upgrade to Pro
• Invite teammate
• Connect Slack
• Book onboarding call

For a user on Free plan with heavy team usage:

• Upgrade to Pro: high revenue value
• Invite teammate: high activation value
• Connect Slack: moderate engagement value
• Book onboarding call: high for new accounts only

Your ranker might compute:

• Upgrade: 0.22 × $120 = $26.40
• Invite teammate: 0.35 × $30 = $10.50
• Slack: 0.40 × $15 = $6.00
• Onboarding call: not eligible

So the system shows “Upgrade to Pro” first, unless rules say “invite teammate” is preferred during onboarding.

---

If you want, I can also give you:

1. a concrete ML design for NBO ranking,
2. a SQL/schema + event tracking plan, or
3. a Python/pseudocode example for scoring and ranking offers.