What It Costs to Build a Self-Optimizing Email Program with Smart Banners, and What It Returns by Level 5

When a retailer evaluates Smart Banners, the conversation almost always starts with a test. Four weeks of live longitudinal testing, real transactions, deterministic attribution, measurable lift in click-to-conversion. The test works. It works reliably enough that most teams stop there, and that is exactly where the economics break.

A self-optimizing email program is not a software purchase. It is a capital allocation decision spread across five email personalization phases, each building infrastructure the next phase depends on. Smart Banners are the entry point, but treating Phase 1 as the destination caps your return at the moment infrastructure value has barely begun to accrue. With marketing budgets flat at 7.7% of company revenue for the second consecutive year, and paid media absorbing 30.6% of the total marketing spend (the only budget category that grew share over five years), every dollar needs a thesis behind it. Here is the thesis for email.

This post lays out what each phase costs, what you are actually buying at each stage, and what stopping early leaves on the table. It is written from the CEO seat because the shared economic buyer above the email team and the performance marketing team is the person who needs to see this math.

Why Self-Optimizing Email Is a Capital Allocation Decision, Not a Tool Purchase

The standard industry framing for email maturity measures operational adoption: deliverability practices, list hygiene, automation footprint. Salesforce’s Underperformer/Performer/Leader taxonomy and Litmus’s five-level framework can both crown a retailer a “Leader” while 95% of their send volume still has zero content-level personalization. That framing does not help a CMO decide where to put budget.

Maturity, in terms that matter for capital allocation, is what the program earns per subscriber per year, measured by RPM and click-to-conversion at the block level. Zembula’s email maturity model anchors each phase to economic output because that is the only way the investment thesis holds. Phase 1 buys observability. Phase 2 buys hero-level brand control. Phase 3 buys image-based product recommendations. Phase 4 buys category curation at scale. Phase 5 buys a self-optimizing system that compounds. None of those are software line items. They are infrastructure layers.

Meanwhile, the paid channel that absorbs most of the marketing budget is getting more expensive and less measurable. Average ecommerce ROAS fell to 2.87 in 2025, declining across 13 of 14 industries. Meta CPMs climbed roughly 20% year over year. iOS ATT means only 40 to 60% of conversions are even visible to ad platforms. Email, by contrast, runs on first-party identity, owned audience, and deterministic measurement. The question is not whether email deserves more budget. The question is how much infrastructure you need to build before the return compounds.

Phase 1: Smart Banners and Block-Level Observability (The 4-Week Test)

Phase 1 is the wedge. You deploy Smart Banners across your broadcast email program, creating conditional content blocks that render personalized messages at open time: abandoned cart reminders, loyalty point balances, back-in-stock alerts, price drop notifications. The content appears at the top (Banner) and bottom of every email, covering 100% of broadcast volume without touching your existing template design.

The test is four weeks of live longitudinal testing looking at actual transactions. No projections, no modeled conversions — real purchase data tied to personalized content impressions. The low-end result that justifies continuing with Smart Banners is a 3.75x ROAS. That alone clears the bar most paid channels cannot reach. But the average Zembula customer achieves 15x ROAS, and the top end we have seen — with J.Crew — is 42x ROAS. The range is wide because the return scales with how much infrastructure you build on top of the initial test.

Smart Banners consistently produce click-to-conversion rates that outperform the overall email baseline by a wide margin. Across the Zembula platform, the average CTC for personalized Smart Banner content runs around 13.6%, compared to the roughly 2.5% baseline for the entire email. That is a 5.4x lift.

But the most valuable output of Phase 1 is not the lift itself. It is the observability layer. For the first time, you have block-level RPM and CTC attribution with 7-day click windows. You can see which content types generate revenue, which subscriber segments respond to which messages, and where the untapped opportunity sits. Phase 1 gives you the measurement infrastructure that makes every subsequent phase a data-informed decision rather than a guess. That attribution data, tracked at the block level rather than the campaign level, is what separates email-as-performance-channel from email-as-batch-and-blast.

Phase 2: Triggered Hero Personalization, Where Brand Risk Meets Real Lift

Phase 2 moves personalization into the hero image of your triggered flows: welcome series, cart abandonment, post-purchase, browse abandonment. These are the highest-intent moments in the subscriber lifecycle, and they are where brand teams get nervous. The hero is the most visible real estate in the email. Handing it to a dynamic content engine feels like a risk.

It is a risk worth quantifying. Triggered emails already convert at multiples of broadcast sends. Adding personalized hero content to those flows — product images dynamically composed based on the subscriber’s actual browsing or cart behavior — compounds that conversion rate further. The infrastructure you are buying in Phase 2 is brand-safe templating for your highest-value moments, plus the operational muscle memory of collaborating between the CRM team, the brand/creative team, and the personalization platform.

That three-team collaboration pattern becomes the operating model at Phase 5. Phase 2 is where it starts.

Phase 3: Product Grids and the Shift From HTML Cards to Image Composition

Most email product recommendations ship as HTML cards: text-based layouts pulling product name, price, and image from a feed. They work, but they are constrained by email client rendering inconsistencies, limited design control, and zero ability to personalize the visual presentation per subscriber.

Phase 3 replaces HTML cards with image-based product grids composed at render time. The product image, price, badge (“New,” “Sale,” “Low Stock”), and layout are composited into a single image asset that renders identically across every email client. This is the shift from templated data insertion to actual image composition, and it unlocks design quality that HTML-in-email simply cannot match.

The infrastructure you are buying here is the template library. Every product grid template you build in Phase 3 becomes a reusable asset. By the time you reach Phase 5, the template library is the single most valuable piece of email infrastructure your team owns, because every new campaign can pull from it without starting from scratch. Templates are an appreciating asset: each one you build reduces the marginal cost of every future send.

Phase 4: Category Smart Banners and Audience-Specific Curation at Scale

Phase 4 extends the AI email optimization model to category-level curation. Instead of a single Smart Banner message, subscribers see category-specific content based on their purchase history, browsing behavior, and affinity signals. A subscriber who shops activewear sees activewear-focused banners. Someone browsing home goods gets home goods.

McKinsey’s research puts the conversion lift from personalization at scale at 15%, with a 10 to 20% reduction in marketing costs across retail verticals, which Zembula achieves with the Smart Banner alone. Applying more personalization to the 95% of broadcast volume that currently has limited content-level personalization, and the math gets material quickly.

Phase 5: Broadcast Hero and the Three-Team Operating Model

Phase 5 is where the program becomes self-optimizing — the end state of autonomous email marketing. The broadcast hero image, the single most valuable piece of real estate in every email, is now dynamically composed from subject line coordination, title text, product imagery, and promotional context. The Campaign Decision Engine selects content at open time based on subscriber-level signals. The CRM team, brand team, and performance team operate as a coordinated unit, with shared attribution data from Zembula’s block-level measurement serving as the common language.

This is the phase where Litmus’s finding that 87% of businesses use AI in email workflows but only 6% qualify as AI high performers stops being a statistic and becomes a competitive moat. The gap between adopting AI tools and actually achieving AI-driven email personalization performance is the template library, the attribution data, and the operating model built across Phases 1 through 4. You cannot skip to Phase 5. The infrastructure is the product.

The Compounding Return: What Level 5 Earns vs. Stopping at Smart Banners Phase 1

Here is the capital allocation argument in plain terms. The 4-week test produces a measurable, attributable lift. Smart Banners work. The low-end result that justifies the investment is a 3.75x ROAS — already stronger than the 2.87 average that paid media delivers. That is enough to greenlight the test on its own economics. It is also exactly why most programs plateau there.

But the 3.75x floor is the return for teams that stop at Phase 1. The average Zembula customer reaches 15x ROAS. The top end — J.Crew — has hit 42x ROAS. The difference is not a better product. It is more infrastructure. Each phase adds a compounding layer: the template library reduces production cost, the block-level attribution data improves content selection, and the three-team operating model aligns Editorial, Data, and Performance Marketing around shared economic metrics instead of siloed campaign reports.

Consider the alternative allocation. That same budget directed at paid media buys you a 2.87 ROAS that is declining year over year, measured through attribution windows that miss 40 to 60% of conversions, on audience you rent rather than own. The return-on-spend comparison is not close. At 3.75x you are already ahead of paid. At 15x you are in a different category. At 42x you have built a compounding asset that no ad platform can replicate — but only if you commit to the full infrastructure build.

Per-subscriber email revenue has declined roughly 35% in real terms since 2016, even as send frequency grew 63% (Klaviyo benchmark data). Volume cannot fix a per-subscriber yield problem. Infrastructure can. That is the investment thesis.

The 2025 email performance benchmark report breaks down the specific CTC and RPM metrics across each maturity phase if you want to model this against your own program’s numbers.

Key takeaways

• Smart Banners are the entry point, not the destination. A 4-week live test looking at real transactions sets the floor at 3.75x ROAS. The average Zembula customer hits 15x. J.Crew reached 42x. The gap is infrastructure, not luck.
• Each phase buys infrastructure, not features. Observability (Phase 1), brand-safe hero personalization (Phase 2), image-based product grids (Phase 3), category curation at scale (Phase 4), and self-optimizing broadcast (Phase 5) are compounding layers.
• The template library is the appreciating asset. Every template built across Phases 1 through 5 reduces the marginal cost of every future send and increases per-subscriber yield.
• Paid media is getting more expensive and less measurable. Average ROAS fell to 2.87. CPMs rose 20%. Attribution windows miss up to 60% of conversions. Email runs on first-party identity, owned audience, and deterministic measurement.
• Marketing budgets are flat. Allocation is the only lever. At 7.7% of revenue for two consecutive years, CMOs are not getting more budget. They are choosing where the existing budget compounds.
• Maturity is measured in economics, not stack adoption. RPM and CTC at the block level, per subscriber per year. If your email maturity model does not anchor to revenue output, it is measuring the wrong thing.