Skip to content
Standard Deviation

Brand-grounded AI

Generic AI produces generic output. Yours is generic until you ground it.

Most marketing AI products today are pass-throughs to a foundation model with a system prompt tacked on. They generate competent, brand-anonymous content because they have no model of the brand they’re working for. StandardOS is built on a different premise: every AI call injects the brand kit before generation, and refuses to generate when the brand kit is missing.

What “the brand kit” is

In StandardOS, the brand kit (formally: the StandardGraph) is a structured representation of the brand:

  • Voice profile — tone words, do/don’t language patterns, exemplar copy from the brand’s existing presence
  • Business context — what the company does, who it serves, what makes it different
  • Customer segments — demographic + lifestyle + psychographic descriptions of the people the brand targets
  • Competitive set — direct + indirect competitors, positioning gaps, AI-search visibility share-of-voice
  • Source materials — the raw documents (brand bibles, prior campaigns, research) the AI has read and can cite

The brand kit lives in StandardGraph and is the substrate every other module reads from.

What “grounded” means at runtime

Every AI call in StandardOS — whether it’s drafting a campaign strategy, pre-testing a creative concept, or generating a media plan — does the same thing:

  1. Load the brand kit from the database for the active brand.
  2. Inject the relevant fields into the model call’s context (system prompt + examples + retrieval).
  3. Make the call.
  4. Return the output.

If step 1 returns nothing (no brand kit on file), the call is rejected before it hits the model. There’s no silent fallback to un-grounded generation. The user sees an explicit error: “This module requires a complete brand profile in StandardGraph. Set it up first.”

That refusal is deliberate. The cost of generic output is reputational — clients can tell, audiences can tell, and the brand drifts a little further from itself with every shipped asset.

A concrete example

Without grounding, a generic ad-copy generator asked to write three headlines for “a yacht charter business in St. Barths” might produce:

  1. Experience luxury on the open seas.
  2. Your perfect getaway awaits.
  3. Sail in style. Book today.

Indistinguishable from a thousand competitors. Could be a yacht business, a cruise line, a rental car company.

With grounding — same prompt, but with the Ocean Pro SBH brand kit loaded (water-taxi positioning, French-speaking captain, EUR pricing, English/French bilingual audience, day/evening/late/charter trip types) — the same model produces:

  1. St. Barths to St. Martin in 35 minutes. Book your seat.
  2. Day trips, late returns, private charters. EN/FR.
  3. Avec Jean au gouvernail depuis 12 ans. Avec carte bancaire depuis 2026.

The model didn’t get smarter. The model got informed.

Why this is hard to fake

A lot of “AI marketing” products bolt grounding on as a feature, optional. The user can technically use the product without setting up a brand profile. The path of least resistance — skip the setup, get a result — produces brand-anonymous output that the user then attributes to “AI being generic.”

Refusing to generate without grounding is a product-design choice with a real ergonomic cost. Some users will bounce. The trade-off is intentional: a refused generation is recoverable (set up the profile, retry). A shipped-then-noticed-as-generic asset is not.

How clients see this in practice

When you start with Standard Strategy, the first session is brand-profile setup — voice, segments, competitive set. Roughly 30 minutes of human work, plus an auto-extraction pass on your existing website and Wikipedia / About-page sources. After that, every subsequent module call (strategy briefs, creative generations, content-calendar drafts) reads from that profile automatically. You don’t re-explain the brand to the AI on every prompt; you tell it once, and it remembers.

Where to go next