Ask most email tools how they stay on-brand and you get one of two answers: drag-and-drop templates, or a "brand kit", a logo, two hex codes, and a font. Both are caricatures of what a brand's visual identity actually is. Templates make every brand's email look like every other brand's email, and a three-color brand kit captures almost none of what makes a brand recognizable at a glance.
How it is done today, and why it is weak
Visual identity does not live in a logo and two colors. It lives in specifics: how many sections an email runs, how banners and product grids interleave, how dense or sparse the copy sits, the feel of the imagery, the rhythm of the layout.
Templates freeze those specifics into a generic mold. Brand kits reduce them to a few tokens. And AI tools that take a text description of a brand's style ("clean, minimal, premium") lose almost all the information, because every brand describes itself that way and the words do not carry the visual detail. You cannot describe a brand into existence; you have to show it.
Why we think this is worth getting right
We treat on-brand as product quality, not a finishing touch, because a campaign that does not look like the brand quietly erodes the trust between the brand and its customer. If you can get an agent to generate creative the brand would have made itself, autonomously, you remove the last thing that forces a human back into the loop for every send. That is what makes finished, approvable creative possible at all, and it is a much harder bar than "generate something plausible."
How LTV.ai approaches it
The design work is done by an agentic orchestrator that sees the brand's visual context directly, as images, not as descriptions. Its structural authority is the brand's own reference emails.
It draws on an indexed library of the brand's past creative that grows with every campaign, and, when an idea references a competitor, it can look at that competitor's actual creative. Before it designs, it is handed guidance from the evaluation stack, the design guardrails mined from past rejections and the levers that have performed for this brand, so "what this brand rejected before" and "what works here" constrain the design from the first stroke rather than being corrected after.
Proposed design rules do not touch the models on their own. They sit as pending suggestions on the brand's Brand DNA page, each with the evidence behind it, how often the objection recurred, and do nothing until a human approves them on the dashboard. Only then are they injected into the design prompts.
These rules read like a creative director's notes, because that is what they are: keep layouts image-led with no more than two short text blocks, never put clearance items in the hero, use lifestyle photography rather than studio cutouts, no countdown timers or red urgency banners, one banner then a product grid rather than stacked banners. Approved, a rule becomes a hard constraint the design agent sees on every future draft. Dismissed, it stays gone unless new evidence accumulates.
It also learns from how you edit
Rejections are not the only teacher. The system also watches how marketers actually work in the editor: which AI choices they change and which they leave untouched, not only on the proactively generated campaigns but whenever they edit an AI-suggested design.
Swap the hero image every time, and that is a signal about imagery taste. Consistently trim the copy, and copy density is off. Ship the AI's layout unchanged ten times, and that layout is confirmed brand language. Every edit, and every deliberate non-edit, flows back as feedback, and over time it is this accumulated behavior, not a questionnaire, that fills in the Brand DNA page.
How it stays honest and compounds
The competitor references, synthesized identity, and approved design rules all live in Brand DNA (see the umbrella). Visual identity here is not one asset or one model, it is an accumulation: curated references, a growing creative library, brand and competitor context, approved guardrails, and a performance playbook, each making the next design better. Users do not train a model directly; they curate references and approve or dismiss rules, and the learned identity follows from that, showing up as output that looks progressively more like the brand made it.
Frequently asked questions
Do I configure a design model? No. You curate references and approve suggestions; the learned identity does the rest.
Whose visual style does it use? Yours, learned from your own creative, not a template.
What if a rule is wrong, or my brand evolves? Every rule is visible on your Brand DNA page and can be dismissed or revoked at any time. The system only re-proposes a rule if fresh evidence for it accumulates.
Part of the machine learning behind LTV.ai.
See it on your store: book a demo.