Case Studies
35+ Projects Delivered
Research
Technical Articles · 100+ Posts
White Papers · Strategic Briefings
Research · 3 Papers
Pillar 1 solved seeing. The platform now captures behavioral signals across the full customer journey — every search query, every collection browse, every return-policy check, every hesitation pattern.
But data without intelligence is just storage cost.
The challenge facing every ESP in 2026 is the gap between knowing what a customer did and knowing what to do about it. Omnisend's segmentation engine currently supports standard filters: purchase history, email engagement, browse events. Agencies create 5–10 segments per client. Those segments describe demographics and transaction history. They do not describe intent.
Pillar 2 closes that gap. It takes the behavioral signals from Pillar 1 and converts them into three outputs: who to target (micro-segmentation), what to offer (promotions engine), and what to say (campaign ideation). The combined output is a complete campaign brief — segment, offer, angle, timing — ready for execution in Pillar 3.
These are not independent features. They are an interlocking system where each component makes the others more effective. And surrounding them are two layers that make the system accessible and self-improving: MCP integration as the interface through which specialists interact with all of it, and a Content Hub where accumulated intelligence compounds over time.
Component 02
Every email marketing specialist has had this experience. They open their abandoned cart segment in Omnisend — 2,000 contacts. And they know, intuitively, that these are not 2,000 versions of the same person. There's the person who abandoned at the shipping cost screen. There's the person who checked the return policy three times and left. There's the comparison shopper who viewed eight similar products over four sessions. There's the impulse browser who added something at midnight and forgot about it by morning.
These are fundamentally different people with fundamentally different hesitations. The specialist knows this. They've known it for years.
But the platform gives them 3–5 filter dropdowns and calls it segmentation. "Purchased in last 90 days." "Opened email in last 30 days." "Located in US." So the specialist sends the same abandoned cart email to all 2,000 people — "Hey, you left something behind! Here's 10% off!" — and watches the 2% conversion rate and wonders why it isn't higher.
Omnisend's segment builder — much better.
Klaviyo's segment builder — slightly worse.
It isn't higher because 2,000 different hesitations received one generic response.
The generic abandoned cart email every brand sends — same message, 2,000 people, zero differentiation.
The problem is not that specialists lack segmentation instincts. The problem is that the platform cannot express what the specialist already knows. Micro-segmentation closes that gap — it gives the platform the same resolution the human already has.
Segments today are also static. A customer enters when they meet criteria and stays until they don't. There is no understanding of trajectory — why they entered, how their behavior is shifting, whether they're warming or cooling. The segment is a snapshot, not a story.
The answer is not more filters. The answer is a fundamentally different model.
Broad segments (5–10 per brand) leave substantial revenue on the table because every campaign is a compromise. True 1:1 personalization sounds ideal but is operationally impossible — no agency can create thousands of unique campaigns, no content pipeline can produce them, and statistical sample sizes become meaningless.
Micro-segmentation operates between these extremes: 50–200+ segments per brand, defined by behavioral signal clusters rather than demographic checkboxes.
Our working micro-segmentation engine — behavioral clustering from live Shopify data. Explore it live at microsegments.ai →
A micro-segment is not "women aged 25–34 who purchased recently." A micro-segment is "customers who viewed 3+ eco-friendly products, checked the return policy at least once, arrived from a sustainability-focused ad, and have not yet purchased."
That segment contains 47 people. You know exactly what objection they have (risk/returns), what they care about (sustainability), and where they are in the decision process (deep research, no commitment).
The campaign for those 47 people writes itself: sustainability credentials of the specific products they viewed, free returns emphasis, social proof from other eco-conscious buyers. That campaign will dramatically outperform "Hey, you left something in your cart. Here's 10% off."
Instead of the specialist manually building segments through Omnisend's filter builder — spending 2–3 hours per client per month maintaining and updating them — the system surfaces micro-segments automatically: "New segment detected: 'Return-Policy Researchers' — 284 contacts who viewed products, checked return policy 2+ times, but did not purchase. Average cart value: $127. Recommended approach: objection-removal campaign emphasizing satisfaction guarantee." The specialist reviews, approves, and the campaign ideation engine immediately suggests an angle.
The intelligence moves from the specialist's head into the platform.
Segmented campaigns generate more revenue than non-segmented sends — some claim up to a 760% jump. That benchmark is based on current broad segmentation, 20–25 segments with basic filters.
Micro-segmentation pushes this further. Conservative estimates based on comparable personalization studies: 20–30% improvement in click-through rates and 15–25% improvement in conversion rates on top of the existing segmentation lift.
That is the number that goes in the agency's client report. That is the proof that solves the ROI problem.
The engineering approach starts with what already exists. Pillar 1's enriched behavioral data gives us the raw material — every product view, search query, cart action, return policy view, checkout step, and DOM interaction mapped to individual contact profiles. That's the foundation. No new data collection required.
From there, we build signal interpretation rules. This is largely prompt engineering and domain expertise, not novel ML. Raw events get translated into behavioral indicators using contextual logic. A product_removed_from_cart is not just a removal — combined with checkout_shipping_info_submitted it indicates price sensitivity at the shipping cost stage. The same removal combined with return_policy_viewed indicates risk aversion instead. A repeated collection_viewed for the same category paired with search_submitted for specific product attributes indicates a customer who knows what they want but hasn't found the right match. Same events, different meaning depending on context.
Building these interpretation rules is where our domain expertise in e-commerce behavioral analysis is most critical — and where most AI implementations fail, because they treat events as flat signals rather than contextual indicators.
We then apply clustering algorithms to group contacts exhibiting similar behavioral patterns. These are well-proven techniques from recommendation systems — the algorithmic foundation has existed for over a decade. The innovation is not in the clustering. It is in the signal interpretation layer above it, and in applying the output to email marketing specifically.
This is not theoretical. We have a working micro-segmentation engine producing behavioral clusters from live Shopify data. It identifies intent-based groupings that standard ESP segmentation cannot. The POC exists. The implementation for Omnisend integrates with Pillar 1's data layer, scales across the merchant base, and connects directly with the campaign ideation and promotions engines.
Klaviyo's strongest competitive asset is predictive analytics — CLV prediction, churn risk scores, predicted next order date. These capabilities are genuinely best-in-class.
But prediction and intent are fundamentally different things.
Prediction looks backward. It analyzes historical purchase patterns across millions of customers and says: "This customer will probably buy again in 14 days." It tells you when to send.
Intent looks at the present. It analyzes what this specific customer is doing right now and says: "This customer checked the return policy twice, compared four yoga mats, arrived from a sustainability ad. She's hesitating because of a specific objection." It tells you what to say.
Klaviyo tells you a customer will churn. Micro-segmentation tells you why they're about to churn and what message will prevent it. The first is a forecast. The second is an intervention.
The moat is not the algorithm. The moat is the data the algorithm generates over time.
Omnisend cannot out-predict Klaviyo — they have years of data advantage. But Omnisend can out-understand Klaviyo by capturing behavioral signals Klaviyo's architecture was not designed to ingest. If Omnisend starts now, in 18 months they will have 18 months of intent data that Klaviyo cannot replicate backward. In 36 months, the system has observed multiple full purchase cycles for most customers — it can predict intent shifts before they manifest in behavior. That dataset does not exist anywhere else.
Here is what this looks like in practice. A single customer session generates raw behavioral events. The micro-segmentation engine extracts intent signals from those events — not just what happened, but why. Those signals map directly to marketing vectors: the specific message, angle, and offer that addresses this customer’s actual hesitation.
Component 05
There is a pattern that every SaaS company needs to internalize: users have started interacting with their tools through AI assistants rather than the tool's own dashboard. Notion through Claude. Shopify through ChatGPT. Slack through Claude Code. GitHub through Cursor. MCP adoption has been rapid — the protocol is mature, well-documented, and integrated by dozens of major platforms.
Users who work this way do not go back. The cognitive load of context-switching disappears. The friction of navigating dashboards is replaced by natural language.
When a specialist using this workflow switches to Omnisend, they are forced into manual mode — separate dashboard, click-through menus, manual filter configuration. The most intelligent part of their stack becomes the most friction-heavy. This is not a future problem. It is happening right now, and the gap widens every month as more platforms integrate.
A specialist is in Claude planning next week's campaigns. They ask: "Pull up last month's performance for Client A's eco-conscious segment." Omnisend returns the data — inside Claude. The specialist sees open rates, revenue, placed orders. They ask: "How did guarantee-based offers compare to discounts for this segment?" The data comes back. They decide on an approach. They say: "Create a campaign for the eco-conscious segment. Sustainable sourcing angle. Satisfaction guarantee offer. Tuesday 10am EST." The agent builds the campaign, applies the segment, sets the schedule — confirming each step. The specialist approves without ever opening the Omnisend dashboard.
That's the cognition side — querying data through the assistant. And the action side — executing campaigns through the assistant. But there's a third function that changes everything.
Reasoning Capture. Every time a specialist plans a campaign through Claude connected to Omnisend's MCP, the platform doesn't just execute the request — it captures the reasoning chain. Which segments were considered, which angles were debated, what past performance was referenced, why one approach was chosen over another.
That reasoning is the fuel for the Campaign Ideation Engine. The more people interact through MCP, the smarter the platform gets. The more it learns about how real marketers think, the better its suggestions become. This is not a side effect. This is the strategic purpose of MCP integration.
MCP is an open, well-documented protocol. The engineering lift is moderate — primarily exposing Omnisend's internal APIs as MCP-compatible tools and handling authentication and permissions. The protocol itself is mature and well-adopted.
The real expertise is in knowing which Omnisend operations to expose for maximum specialist value:
The domain knowledge matters more than the engineering. We've worked extensively with MCP — we understand the protocol's capabilities, its auth model, and where implementation typically breaks down. The hard part is getting the tool definitions right so that the specialist's natural language maps cleanly to Omnisend's operations.
Klaviyo already has an MCP server. But Klaviyo's implementation is a read layer — AI assistants can pull data from Klaviyo. Query segments, retrieve campaign results, access contact information. Read-only.
Klaviyo's MCP server announcement — they're already moving on this. Read the announcement →
We're proposing bidirectional — read and write — with reasoning capture. Those are fundamentally different products. Klaviyo built MCP to keep pace with the ecosystem. Omnisend can build MCP to capture value from the ecosystem.
