By Christelle Grauer, VP of Data & Compliance at Brevo
From rules to relevance: why predictive AI beats static segmentation
Rule-based segmentation feels safe: clear rules, full control. But that safety can stall growth. Predictive AI can move faster and more precisely, yet without guardrails, it risks eroding trust and chasing empty engagement. The way to do this is to make every segmentation choice answerable to one test: Does it increase long‑term customer lifetime value (CLV)?
Rule-based segmentation: the comfort trap
For years, many teams have lived inside rule-based segments: “If the customer bought X, send Y.” It’s tidy, explainable, and feels in control.
But if your goal is to optimize opens, clicks, and conversions, rules are now a drag on growth. Humans can’t match the speed, personalization, or reactivity of models that learn across millions of signals in real time. Rule-based logic still has a place, especially when you lack data or require high explainability, but it’s yesterday’s tool for optimization.
Why it matters for customers:
When segmentation relies only on rigid rules, customers often get irrelevant messages. Predictive targeting ensures they receive communications that actually reflect their current needs, not outdated assumptions.
Balancing predictive power with explainability
Leaders often frame AI segmentation as a choice between raw predictive power (the model’s accuracy at predicting outcomes and explainability (how clearly you can see and justify why a customer was grouped or targeted). But it’s not a binary. It’s a balance, and the balance shifts by domain.
If your customers must understand why they were targeted, full black box approaches will work against you. They require transparent reasoning for inclusion and exclusion, which opaque models can’t provide.
You can lean further into high-performing models in data-rich consumer contexts with looser explainability needs.
Start by stating the business objective and constraints, then dial the model and the governance to fit. Optimization without trust is fragile; trust without outcomes is expensive.
How to set the balance in practice
- Regulatory pressure high? Favor what you can explain, using simple models that show what factors are important, and clearly write down why certain people are included or excluded.
- Low tolerance for targeting errors? Use conservative thresholds, add human-in-the-loop review on new segments, and expand only after stable lift.
- Data-rich and fast-moving? Allow higher‑capacity models with clear guardrails (frequency caps, suppression rules, and automated holdouts).
- Stakeholders need proof? Pair every new segment with a control group and publish lift with confidence intervals, not just raw deltas.
Why it matters for customers:
A balance of accuracy and transparency builds trust. Customers can understand why they receive messages while still benefiting from personalized, relevant offers.
Behavior beats self-declaration
We worked with Alltricks, a leading cycling and outdoor gear retailer, on a proof of concept to predict the most relevant contacts for each campaign. This project revealed vital insights about the limits of self-declared data and the power of behavioral signals.
Alltricks is already mature in AI‑driven email targeting and knows the kind of uplift an algorithm can deliver. But before AI‑assisted segmentation, targeting leaned heavily on user‑declared interests at sign‑up. In our tests, self‑declared interests at sign‑up were weaker predictors of purchase than actual behavior.
Imagine two customers: one signs up and declares an interest in running shoes but never browses them again; the other never mentions running shoes at sign-up but browses and buys cycling gear weekly. Predictive segmentation prioritizes the second customer more effectively, because their behavior proves real intent.
By moving to predictive segmentation, Alltricks didn’t just maintain their strong performance, they improved it. Campaigns targeted with behavior-driven predictions lifted engagement significantly: between 30–40%.
And more importantly, this wasn’t a one-off experiment. Seven months later, Alltricks is still using the system in production, a testament to its sustained impact.
Three practical truths:
- Purchases carry the strongest signal. They encode intent that has already become action.
- Browsing wins on volume. You’ll have 10–100× more of it than purchases. At scale, that density helps models generalize and react fast.
- Email engagement is the noisiest. Tracking is messy, so treat it as supporting context, not the primary fuel.
Why it matters for customers:
Behavioral targeting means people see offers aligned to what they actually shop for, not just what they once declared. It makes campaigns feel relevant, timely, and less like spam.
Barriers beyond bad data in predictive programs
Everyone blames “bad data.” Fair. But three underrated blockers decide whether predictive segmentation sticks:
- Collection at the edge. You need enough clean, consented events (web, app, and commerce) to feed models.
- Team psychology. Marketers are rightly proud of their craft. Giving targeting to an algorithm can feel like giving away the job. Leaders must make the case that AI removes busywork and expands their impact.
- Maintenance and fatigue. AI makes it easy to expand reach. That’s dangerous. Email fatigue creeps up fast and quietly erodes reputation.
Why it matters for customers
Predictive segmentation only works if it’s fueled by clean data and managed responsibly. When teams respect frequency and relevance, customers feel valued instead of overwhelmed.
Further reading: Smart Segments: Better Conversions with Brevo CDP
Quality over quantity: CLV over campaign volume
Before the metrics, a quick frame: what you measure becomes what you make. If the scoreboard rewards volume, you’ll get more email. If it rewards value, you’ll get stronger relationships and healthier retention.
Vanity metric: number of emails sent. It feels productive, but it rewards volume over relationship.
Signal metric: customer lifetime value (CLV). CLV tells you whether segmentation increases the long‑term value of your base, not just this week’s clicks. A simple, practical version most teams can start with:
- 12‑month CLV = average order value × average orders per customer × 12‑month retention rate.
- Track it by segment and by model policy. If a tactic lifts clicks but lowers 12‑month CLV, it’s the wrong tactic.
- When you can’t compute full CLV, use proxies: repeat‑purchase rate, 90‑day revenue per customer, churn probability, and return‑on‑discount (revenue lift net of incentives).
Why it matters for customers:
When marketers prioritize CLV, customers benefit from more thoughtful, relationship-driven interactions, rather than being bombarded with endless short-term campaigns.
Stop cold re‑engagement blasts
The habit of blasting dormant contacts looks efficient on paper, but is costly in practice. Dormant audiences:
- produce higher complaint rates and unsubscribes,
- trigger filters that hurt inbox placement for all mail,
- dilute learnings because opens/clicks are noisy on cold lists.
A better pattern:
- Define “cold” precisely (e.g., no opens/clicks/purchases in 180 days, no site activity in 120 days).
- Run a small, explicit re‑permission flow via low‑frequency channels (e.g., one email + one SMS if consented), with a clear value proposition.
- Sunset respectfully. If no signal returns, stop. Protect the domain and the audience you can serve.
Why it matters for customers:
Instead of being spammed after months of silence, customers get a respectful chance to re‑engage on their own terms.
Browsing data: the underrated powerhouse in predictive targeting
Most undervalued data: website browsing, even anonymized. It’s plentiful and immediate. Used well, it tells you intent before a cart exists.
How to turn browsing into a segmentation signal:
- Capture page type (homepage, category, product, help), content attributes (brand, price band, category), and session context (device, recency, depth).
- Create simple profiles of user interests based on their browsing, and connect them to a known user profile when they log in or click a link from an email.
- Use recency decay so that last week’s signals don’t outweigh yesterday’s.
- Filter out bot‑like patterns and throttle event volume at the edge.
Meanwhile, treat open rate with caution. Privacy features and automated security checks inflate it unevenly. Use opens as a diagnostic, not a KPI. Optimize for downstream outcomes: clicks, add‑to‑cart, purchase, and ultimately CLV.
Why it matters for customers:
When browsing behavior informs targeting, people receive suggestions that match their active interests, helping them discover products they actually want instead of generic promotions.
Further reading: 9 CDP Use Cases for Retailers
This shift also comes with responsibilities
Predictive segmentation doesn’t happen in a vacuum. As teams move from rules to outcomes, they also inherit new obligations: to the people they target, the data they handle, and the ecosystems they influence.
Ethics: optimize for outcomes, not over‑consumption.
If you only reward short‑term engagement, the system will push more of it, even when that nudges people toward over‑consumption. Build guardrails: caps, fatigue models, and goals that weight long‑term value, not just next‑click dopamine.
Privacy: from checkbox to advantage.
Assume privacy controls will keep getting stricter, and customers will keep hiding more behavior. Collect the minimum you need, aggregate where you can, and preserve signals through anonymized events. You rarely gain trust with data practices, but you can lose it fast.
The future: quantum informatics as the next frontier
Looking ahead, the next leap may not be another algorithm but quantum informatics: applying quantum principles to tackle calculations today’s systems can’t handle. In simple terms, quantum computers juggle many possibilities at once, making it possible to optimize across millions of variables (campaign timing, offers, channels) in ways that are currently out of reach. It’s not here yet for marketers, but even CMOs, such as Citi’s Alex Craddock, are starting to discuss its potential for hyper‑personalization.
What to do now:
- Keep your data model simple and consistent so you can swap engines later.
- Invest in real‑time pipelines; the right features help any future model.
- Design objectives that reflect long‑term value (CLV, fatigue, fairness), not just clicks.
The takeaway
Predictive AI is not about surrendering to a black box; it is about deciding the outcomes that matter most and then shaping your data, models, and guardrails to deliver them. Done well, predictive segmentation moves teams from chasing clicks to compounding long‑term customer value. The comfort of rules gives way to the discipline of outcomes, and CLV is the scoreboard that keeps you honest.
The destination is clear: segmentation that is faster, more thoughtful, and more respectful of the people it serves. Marketers who embrace this won’t just optimize campaigns; they’ll build systems that earn attention, protect reputation, and unlock durable growth. And in doing so, they will prove that predictive AI is commercially powerful and strategically responsible. Segmentation can be faster, smarter, and more respectful of the people it serves.
Marketers who embrace this will optimize campaigns and build systems that earn attention, protect reputation, and unlock durable growth.
