Brand books were written for people. A designer can look at a logo page, infer what “premium” means, remember how the product opens and quietly avoid a visual claim that would never pass legal review. A multimodal generation system needs those missing decisions expressed as evidence and rules.
That distinction matters now. In September 2026, Google said Gemini Omni in Asset Studio could interpret a URL, marketing brief, product shots, video clips and other uploaded references. Google also recommended supplying four or five real product views, realistic backgrounds and other concrete examples. The operational question is therefore no longer simply whether AI can make a video. It is whether a brand’s current library can prove enough of the product and context to make the output reviewable.
What Google has actually confirmed
Google’s Creative Drop #1 describes a workflow in which an advertiser selects sources and an aspect ratio, writes a brief, chooses among three narrative directions and then edits the storyboard scene by scene. Brand-guideline import can convey fonts, hex values, logo rules and visual style. The workflow supports vertical, horizontal and square video, and Google says the capability is rolling out globally inside Google Ads.
Those are product facts, not proof that a PDF can prevent a false product detail or a risky claim. Google’s help page for auto-generated video ads makes another dependency explicit: generated videos rely on the status of the original assets. If an original asset is edited or deleted, the existing generated video stops serving and an updated version is created. Source assets have become live production inputs, not an archive that can be forgotten after upload.
The practical implication is narrower, but more useful, than “AI replaces production.” Better references can reduce ambiguity. They cannot transfer product accountability, usage rights or market approval to the model.
Choose the production path from the evidence gap
Before commissioning a new shoot, classify the gap.
| Production path | Use it when | Buy this deliverable |
|---|---|---|
| Reuse the existing library | Priority SKUs already have truthful multi-angle coverage; materials, scale, ports and packaging are visible; approved contexts exist | Version cleanup, metadata and prohibited-use rules |
| Run a reference capture | Missing evidence is limited to a few angles, hand interactions, reflective materials, opening mechanics or lighting conditions | A compact capture that fills named gaps and produces reusable reference modules |
| Keep full or hybrid production | The idea depends on performance, complex physical action, regulated claims, a hero launch or a low-tolerance brand moment | Directed live action and post-production, with AI used for previs, controlled extensions or selected replacements |
This turns “Should we reshoot?” into a coverage decision. A missing rear connector should not trigger a full campaign production. A safety demonstration, human performance or tightly regulated product claim should not be downgraded to loosely supervised generation.
The ONCE ANCHOR reference-pack framework
An AI-ready brand reference pack needs more than identity assets. ONCE uses six layers—ANCHOR—to connect what the model receives with what people must approve.
A — Authentic product evidence
Supply four or five useful views for each priority SKU: front, rear, side, top and a critical detail where relevant. Record the true material, color, relative scale, packaging and accessories. These are not merely attractive photographs. They are the evidence reviewers need when checking geometry, texture and function.
N — Non-negotiables
Name the details that cannot drift: port count, label position, hinge direction, screen ratio, prohibited actions, claim language and forbidden product-context pairings. “Keep the product consistent” is too vague to be an acceptance criterion.
C — Context library
Separate approved environments, lighting, people or hands, actions, props and negative examples. The same product can imply different things in a kitchen, a factory, outdoors or on a retail shelf. State which combinations are valid and which would mislead.
H — Hierarchy of identity
Keep the conventional logo, typography, color, tone, layout hierarchy and safe-zone rules. Add approved examples that show how those rules are executed. Adobe’s September 2026 guidance on training custom models with approved brand imagery makes a useful vendor-side distinction: guidelines describe intent, while campaign imagery reveals composition, lighting, subject treatment and hierarchy. That observation supports the need for examples, although it does not establish a universal industry standard.
O — Ownership and origins
For every source, record the rights owner, talent or location release, permitted markets and channels, expiry date, AI-use restrictions, source file and version. A visually accurate output can still be unusable if nobody can establish whether its reference material was cleared for generation in a particular market.
R — Review and replacement
Assign an owner for narrative choice, product truth, claim review, local regulation and final release. Keep the source-to-output dependency, and define which derivatives must be reviewed again when a source changes. Because Asset Studio separates narrative selection from scene-by-scene refinement, approve the story direction and the final execution as distinct gates.
Measure cost per verified task, not per generation
One hundred generations are not one hundred usable assets. A more defensible cost model is:
Reference-system cost = audit + targeted capture + asset structuring + rule encoding + scene QA + rights and version maintenance
Divide that total by verified generation-task coverage. Define one task as a combination such as SKU × approved scene family × market or language × aspect-ratio family. Count it only when the reference pack can support that combination and the output passes product, brand, rights and localization review.
This model does not assume that AI saves money. It exposes whether a capture unlocks repeatable work, which derivatives are invalidated when a source changes and whether a growing asset library is actually creating controlled production capacity.
A procurement brief that can survive the next model update
List the required SKUs, markets, aspect ratios and scene families before discussing shot count. Mark every existing asset as covered, capture required or prohibited for generation. Keep the targeted shoot limited to facts the current library cannot prove. Reserve full production for performance, difficult physical behavior and hero content where failure carries a higher cost.
For teams already building modular commerce video, pair this upstream reference system with ONCE’s Amazon REC modular asset method. ANCHOR governs what a generation system is allowed to infer; the REC method governs how approved modules can be assembled downstream. In both cases, scale comes from traceable, replaceable and reviewable components—not simply a larger pile of finished videos.
Google has made the input surface richer. The brand’s next decision is to make its evidence equally structured. Start by auditing one priority SKU against the six ANCHOR layers. The result will tell you whether the next purchase should be metadata work, a focused reference capture or a conventional production.
Sources and limits
Variable product facts were checked against Google’s Creative Drop #1 (September 2026), Google Ads Help on auto-generated video, and Google Ads Help on Demand Gen generative tools, accessed October 11, 2026. Adobe’s August and September 2026 materials are used only as vendor evidence of a similar direction in brand governance. ANCHOR, the three production paths and the coverage-based cost model are ONCE’s analytical framework, not a Google or Adobe promise and not an unverified efficiency benchmark.
