The Large Brand Model

Every technology that makes something abundant moves value somewhere else. Printing made text abundant and moved value to editing. Recorded music made performance abundant and moved value to curation and rights. Cloud computing made servers abundant and moved value to orchestration. The pattern is old. Its latest instance is unfolding in commercial creative right now.

AI made creative abundant. Abundance doesn't raise the value of making things. It raises the value of controlling them.

The abundance is measurable. Half of video advertisers already use generative AI to build video ad creative, and the IAB projects that 40 percent of all video ads will be generative AI creative by 2026.¹ Meta reported more than eight million advertisers using at least one of its generative AI creative tools in the first quarter of 2026.⁵ EMARKETER forecasts that AI-powered ad spend in the United States will reach $57 billion in 2026, up 63 percent in a year, while spend without AI grows at roughly 5 percent.⁶ Midjourney priced its first video model at about eight times the cost of an image job, which it described as more than 25 times cheaper than what the market had previously shipped.⁷ The marginal cost of a frame is collapsing toward zero.

When the marginal cost of a frame approaches zero, the frame stops being the scarce thing. The scarce thing becomes the set of judgments that surround it. A campaign that could once be produced twenty cuts at a time can now be produced twenty thousand variants at a time, and every one of those variants is a decision somebody is accountable for.

The consequences are already visible in market. The 2024 Coca-Cola holiday spot, produced entirely with AI, hid Santa's face because the models could not hold him steady.⁸ A year later, McDonald's Netherlands withdrew an AI-generated Christmas advert after public criticism, and a critic noted that the truck at the center of Coca-Cola's own spot seemed different in every frame.⁹ Google withdrew its "Dear Sydney" Gemini spot from Olympics rotation after backlash, noting that the ad had tested well before airing.¹⁰ Those are the failures visible from the outside. The ones that matter more are invisible: a tone slightly off, a color a shade wrong, a claim the legal team never saw, repeated ten thousand times across a media plan.

Four questions carry the new scarcity. They are the questions a chief marketing officer is actually accountable for, and none of them is answered by generating more.

A Large Brand Model exists to answer these four questions continuously.

Does it hold the brand? Consistency is the thing that generation models are worst at and that brands are most dependent on. The IPA and System1's study of more than 4,000 ads from 56 brands found that consistent brands generate 27 percent more very large brand effects and 28 percent more very large business effects than inconsistent ones.¹¹¹² Holding the brand is a body of knowledge: voice, visual grammar, story world, and the meaning behind them, applied to every frame.

Who approved it? Abundant creative means abundant decisions. When a team produced twenty cuts a quarter, approval was a meeting. When a system can produce twenty thousand variants a week, approval is an architecture. Somebody, or something, must hold the record of what was cleared, by whom, under what conditions, and what was refused. Only 33 percent of marketers have adopted formal AI governance tools, and among the 70 percent who have experienced an AI incident, 40 percent had to pause or pull ads.³

Can we repeat it? A campaign that works once is an anecdote. A campaign that can be reproduced, extended, and adapted without starting over is an asset. Reproduction requires memory, and memory requires a place to live. Prompts are not memory. A production partner's hard drive is not memory. Neither survives a change in vendor, model, or staff.

Did it work? Marketing leaders report that demonstrating financial impact is their single greatest challenge, with 63 percent facing increased pressure from the CFO and half facing more from the board.¹³ Creative is the largest single driver of advertising's sales effect, responsible for 49 percent of incremental sales in NCSolutions' analysis of nearly 450 campaigns,¹⁴ and the same research finds that marketers estimate that contribution at 19 percent.¹⁵ The most valuable lever in the system is the one that is least measured.

These are governance questions. They are not production questions. The organizations that win the next decade of brand building will be the ones that answer them systematically, at the speed the models now demand. Answering them requires a new kind of model, and the model belongs to the brand.

Categories become real when they are defined tightly enough for others to use. "Customer data platform" was named in a 2013 blog post by the analyst David Raab and became an industry within three years.¹⁶ Scott Brinker's first marketing technology landscape catalogued roughly 150 companies in 2011; the 2025 edition catalogued 15,384.¹⁷¹⁸ In each case the name preceded the market and shaped it. What follows is offered in that spirit: a definition precise enough to be adopted, argued with, and built against.

Definition

A Large Brand Model (LBM) is a persistent, governed intelligence built around a single brand. It encodes what the brand looks like, sounds like, and stands for; how its stories are told in motion; what has been approved, revised, and refused, and by whom; and what has performed in market. It is not itself a generation model. It sits above generation models, briefing, constraining, selecting, and grading them, and it retains everything it learns when those models change. Its output is a decision rather than an asset: this is on brand, this is approved, this is repeatable, this worked. Where a large language model is trained on the world, a Large Brand Model is trained on one brand.

What it is not

An LBM is not a brand guidelines document, which describes rules but cannot enforce them. It is not a digital asset management system, which stores outputs but holds no judgment about them. It is not a fine-tuned generation model, for reasons Section 4 makes plain. It is not a prompt library, which is memory without state. And it is not a single vendor's proprietary model, because a brand's intelligence that can be held hostage by one vendor's roadmap is not the brand's intelligence at all.

Where it sits in the stack

An LBM does not replace every system in the marketing stack. It connects and operationalizes the brand-specific intelligence those systems already contain.

Each of those prior tools solves a fragment of the problem. The LBM is the category that unifies the fragments into one system with one owner: the brand.

A general model is a remarkable instrument for a general problem. It has read more than any person and can imitate almost any style on request. That breadth is precisely why it cannot serve as a brand's intelligence. It knows everything, and therefore knows nothing in particular. It has the average of every brand, which is the same as having none. The Stanford group that coined the term "foundation model" in 2021 chose the word to underscore what they called the models' "critically central yet incomplete character."¹⁹ A foundation is meant to be built on. The question for a brand is what gets built on top, and who owns it.

Prompting is not memory

The first instinct of every marketing team encountering a general model is to describe the brand to it. Paste the guidelines. Attach the tone-of-voice document. Add the logo. This works for a single session and fails as a system, for three reasons.

A prompt is a request, and a request is not a rule. A model asked to stay on brand will usually try, will not always succeed, and has no mechanism to know when it has failed. The knowledge-retrieval literature has been explicit about this since 2020: a model's parametric memory has a limited ability "to access and precisely manipulate knowledge," which is why retrieval systems were developed to supplement it.²⁰ A prompt has no history. It does not know that a tagline was retired in March, that legal rejected a claim last quarter, or that the spot with the blue grade outperformed the spot with the warm grade. And a prompt does not scale across people. Two hundred marketers, each prompting differently, produce two hundred slightly different brands. Salesforce's 2026 State of Marketing found that 75 percent of marketers have adopted AI and 84 percent admit to running generic campaigns.²¹ Generic is the natural output of a general model prompted by a crowd. Prompts alone are not a durable system of brand memory.

Fine-tuning is not a sufficient foundation for persistent brand intelligence

The more sophisticated instinct is to fine-tune: take a generation model, train it on the brand's archive, and produce a bespoke instance with the brand in its weights. This is technically possible and strategically insufficient, for one reason above all others. Generation models turn over every few months, and a fine-tune dies with its base model.

Consider the record of the past 24 months. OpenAI released Sora as a standalone product in December 2024,²² released Sora 2 with synchronized audio in September 2025,²³ and discontinued the Sora web and app experiences on April 26, 2026, with the API scheduled to follow in September.²⁴ Disney had committed a billion dollars to the platform and reportedly learned of the shutdown less than an hour before the public did.²⁵ Google shipped Veo 2 in December 2024,²⁶ Veo 3 with native audio in May 2025,²⁷ and Veo 3.1 in October 2025.²⁸ Runway moved from Gen-4 in March 2025²⁹ to Gen-4.5 in December.³⁰ Kuaishou's Kling completed more than 20 iterations in its first year³¹ and shipped a 3.0 model in February 2026.³² On the image side, OpenAI replaced its own default image generator within nine months,³³³⁴ and Google's Gemini 2.5 Flash Image, released in August 2025, was superseded by a Pro version in November.³⁵³⁶

Andreessen Horowitz's 2026 review of generative media summarized the cadence plainly: model releases in 2025 came every four to six weeks, with no reason to expect the pace to slow, and enterprise production deployments now use a median of 14 different models.⁴ The platform vendors document the consequence. OpenAI's deprecation policy states that inference on fine-tuned models continues only until the base model is deprecated,³⁷ and Microsoft's model lifecycle policy specifies that once a model is retired, it is no longer available for fine-tuning and deployments return errors.³⁸

A brand that fine-tunes a generation model is building its most valuable intelligence on a foundation with a published expiry date. Every retirement forces a rebuild. Every rebuild resets the accumulated learning. The lesson is not that generation models are bad. They are extraordinary and getting better every quarter. The lesson is that the brand's intelligence must live somewhere that does not depend on any one of them.

An LBM is composed of four layers. Each answers one of the four questions from Section 2, and each is a distinct body of knowledge that a general model cannot hold.

The brand knowledge layer holds identity. Voice, tone, and vocabulary. Visual grammar: palette, typography, composition, the rules for how the logo lives in a frame. The story world: characters, settings, recurring motifs, the narrative territory the brand owns and the territory it avoids. Distinctive brand assets, in the sense established by the Ehrenberg-Bass Institute,³⁹ are encoded here as constraints, along with the meaning behind them. Kantar's analysis finds that visual assets account for roughly 30 percent of a brand's perceived difference, with the balance coming from positioning and experience,⁴⁰ which is why this layer must hold meaning and history and cannot be reduced to a style sheet. This layer answers: does it hold the brand?

The cinematic grammar layer holds craft. It encodes how the brand moves: shot language, pacing, lighting, sound design, transitions, the rhythm of a cut, the way a product is revealed, the way a face is lit. It knows the difference between a scene that belongs to the brand and a scene that merely contains its logo. This is the layer that separates a system built for commercial creative from a system built for text, and it answers the hardest part of "does it hold the brand," the part that no guideline document has ever been able to write down.

The governance and approval layer holds the record. Every asset carries its lineage: which model generated it, which brief produced it, who reviewed it, what was changed, what was rejected and on what grounds, which legal and regulatory constraints applied, and whether disclosure was required. The layer enforces rules at generation time rather than catching violations after publication, and it makes every decision auditable. It answers: who approved this, and can we repeat it?

The performance feedback layer holds the results. It connects each asset to what happened after it ran, and it writes what it learns back into the other three layers. A creative choice that lifts results becomes a preference. A choice that fails becomes a caution. This is the layer that turns a brand's archive into a compounding asset. It answers: did it work?

Model-agnostic by design

Model-agnostic by design. The phrase describes an architectural choice, and the choice follows directly from the churn documented in Section 4. If the generation layer turns over every few months, anything built to last must treat generation models as interchangeable components. An LBM routes a brief to whichever model best serves it today, and to a different model tomorrow, without losing a single thing it knows about the brand.

This is what makes an LBM stronger as models proliferate. Every new model is a new capability the brand can draw on: better motion, longer clips, native audio, finer control over a character's face. Every new model is also a new source of variance the brand must manage: a different house style, different failure modes, different tendencies to drift. A system that holds the brand independently of the models converts the first into an asset and neutralizes the second. The brand gains every improvement in the field and pays for none of the instability. It also means the brand, its agencies, and its production partners can adopt whatever generation tools they prefer without fracturing the brand's control. The LBM is the constant. Everything else is a plug-in.

The hardest case: motion

Video is where brands are made and where they are most easily broken. It carries voice, image, motion, music, performance, and narrative in a single object, and it is the form in which the largest share of brand-building investment is concentrated; the IPA's foundational analysis of long-term effectiveness recommended that roughly 60 percent of budgets be devoted to brand-building work,⁴¹ which has long been dominated by film. It is also the hardest thing for generation models to do consistently. A still image must be right once. A ten-second spot must be right 240 times in a row, and the character must be the same person in frame 240 as in frame one, in the same light, in the same world. Kantar's research on audience response to AI-generated advertising found that ads with seamless generative AI integration reached top-tier branded cut-through in more than 40 percent of cases, while ads with obvious AI elements performed noticeably worse.⁴² Audiences do not penalize AI. They penalize bad craft.

That asymmetry is why motion is the right case to design an LBM around. A system that can hold a character, a palette, a lighting scheme, and a story world steady across hundreds of frames can hold them steady in one. Stills, key art, social cutdowns, and display units become byproducts of the same governed intelligence. The reverse is not true. A system built for static graphics has no concept of pacing, performance, or continuity, and cannot be extended into motion without starting over.

Every campaign returns performance data that sharpens the next brief. That sentence is the mechanism by which a Large Brand Model compounds, and an example makes it concrete.

A brand approves three campaign concepts. One uses humor, one uses nostalgia, one uses product demonstration. The humor-led work performs substantially better with a particular audience. The LBM records more than "Campaign A won." It records which creative decisions contributed to the result: the opening, the pacing, the character, the claim. The next brief begins with that learning already applied. The nostalgia concept is not discarded; it is filed against the audience where it held, so the next time that audience is in the plan, the brief starts there.

Performance data is not defensible in itself. Media platforms have it. Measurement vendors have it. What makes it valuable inside an LBM is where the data lands: inside the same system that wrote the brief, governed the generation, and recorded the approval. Only there can a result be attributed to a creative decision, and only there can the attribution change the next decision automatically. A generation model receives a prompt and returns an asset. It never sees the brief's history, the approval, the media plan, or the result. It sits at one point in the chain and has no visibility into the rest.

The evidence that this loop pays is strong and predates AI. Kantar's matched analysis of roughly 450 ads across its creative database and the WARC ROI database found that the most creative and effective ads generate more than four times as much profit.⁴³ The IPA and System1 named the cumulative effect of consistent creative "compound creativity," and found that the most consistent brands' creative quality scores rise measurably each year.¹¹ Peer-reviewed work adds a caveat worth respecting: Becker and Gijsenberg, writing in the International Journal of Research in Marketing, found that consistency in advertising content drives long-term cumulative sales, with the largest gains for smaller brands and diminishing returns for the largest,⁴⁴ which argues for an intelligence that knows when to hold the line and when to vary, and that learns the difference from results.

The flywheel has one more property worth naming. It gets more valuable as the brand's own archive grows, and the archive belongs to the brand. A brand that has run a hundred governed campaigns through its LBM has a hundred campaigns' worth of attributed learning that no competitor and no vendor can reproduce. Today that asset is scattered across decks and drives and the memories of people who have since changed jobs.

Governed, repeatable, measurable. Those three words describe what changes when a brand operates through a Large Brand Model, and they map directly onto the pressures CMOs already report. The most recent CMO Survey from Duke University finds that AI now runs a rapidly growing share of marketing activity, that more than 70 percent of marketers prioritize immediate results over long-term gains, and that no marketing technology capability scores above five on a seven-point performance scale.⁴⁵ Technology has outrun the operating model. The LBM is an operating model.

What changes in the workflow

The brief changes first. Today a brief is a document that a team interprets. Under an LBM, the brief is an input to a system that already knows the brand's constraints and history, and the first round of candidates arrives already inside the brand's grammar. Review shifts from catching errors to making choices.

Approval changes second. Approval becomes an event that the system records, with lineage, rather than an email that the system loses. When a regulator, a partner, or a board member asks who approved something, the answer is a query rather than an investigation.

Measurement changes third. Performance data returns to the system that made the creative decisions, and the attribution is at the level of the decision rather than the campaign. That knowledge is retained and applied. It does not leave with the agency or the analyst.

What the CMO keeps control of

Everything that matters. Humans set the rules. Humans approve. Humans direct the creative ambition, an approach Coca-Cola's head of generative AI described when he said creative direction and thought leadership remain human-led, with AI applied to execution and production.⁴⁶ The LBM enforces what humans have decided and remembers what they have approved. It does not decide for them. The CMO also keeps ownership of the accumulated learning: the brand knowledge, the approval history, and the performance record persist when a generation model is retired, when an agency is changed, and when a team member leaves.

What the CMO stops paying for

Rework driven by inconsistency. Lucidpress's survey of more than 200 organizations found that 81 percent still deal with off-brand content and that consistent branding is associated with revenue increases of up to a third, a figure that should be read as self-reported but that points in a direction the IPA's independent data confirms.⁴⁷¹¹ Re-briefing every new vendor and every new model from scratch. Governance that lives in people's heads, which is paid for in incidents. Measurement that cannot reach back to the creative decision, which is paid for in budgets that cannot be defended. And the recurring cost of rebuilding brand intelligence every time the model underneath it is deprecated.

Generation is commoditizing. Enterprise creative operations are not.

When the ability to produce a beautiful frame becomes available to everyone at a price near zero, the ability to produce a beautiful frame is not where value accrues. Value accrues to everything that must be true around that frame for a large organization to put its name on it and pay to place it in front of the public. That "everything" is enterprise creative operations, and it has none of the properties of a commodity. It is specific to each brand. It accumulates over time. It is embedded in workflows, approvals, legal review, agency relationships, media plans, and measurement systems that took years to build and cannot be swapped out with an API key.

Two features of the market make this layer durable. The first is the adoption gap. BCG's 2026 survey of 300 CMOs found that 96 percent believe AI is driving end-to-end transformation of the marketing function, while 42 percent use generative AI only to assist humans with individual tasks and 8 percent run campaigns with multiple autonomous agents.⁴⁸ Gartner's 2026 CMO Spend Survey found that CMOs now allocate 15.3 percent of their marketing budgets to AI, and that only 30 percent report mature capabilities to scale it.² McKinsey's 2026 global survey found that 88 percent of organizations use AI regularly in at least one function while only 37 percent report any earnings impact.⁴⁹ The money is committed. The operational capability to spend it well is missing.

The second is the trust gap. Gartner's consumer research found that half of consumers prefer to give their business to brands that avoid generative AI in consumer-facing content, and 68 percent frequently wonder whether what they see is real.⁵⁰ Gallup found in 2026 that 49 percent of Americans view business use of AI in advertising negatively, against 19 percent who view it positively.⁵¹ The IAB found that 82 percent of advertising executives believe younger consumers feel positively about AI-generated ads, while 45 percent of those consumers actually do.⁵² Trust cannot be generated. It is earned through consistency, disclosure, and craft, all of which are governance functions. Yahoo and Publicis found that a noticed AI disclosure produced a 73 percent lift in ad trustworthiness and a 96 percent lift in overall trust for the company.⁵³ Governance is not the tax on AI creative. It is the source of its value.

Together, these two gaps describe a market in which the scarce capability is the ability to run AI creative inside the enterprise safely, consistently, and provably. That capability is operational. It is learned from inside the workflow. It cannot be downloaded, and it cannot be supplied by a vendor whose product ends at the moment of generation. Value in a commoditizing layer migrates to whoever controls the interface between the commodity and the customer. In commercial creative, that interface is the brand's own intelligence.