What is in the Brand Brain?
Four kinds of knowledge. Facts: what the brand sells, what its materials are, what its shipping and returns policies say, what its guarantees are. Personas: who the brand writes for. Voice samples: passages that demonstrate how the brand sounds. Word rules: the words it uses and the ones it never uses.
All four are extracted from sources the brand supplies, which can be documents, files or URLs, and each entry keeps a link back to the source it came from. Nothing enters the knowledge base without provenance.
- Facts: materials, policies, guarantees, sizing, shipping and returns terms, each with its source.
- Personas: who the brand writes for, in the brand's own words.
- Voice samples: passages that demonstrate how the brand sounds when it is at its best.
- Word rules: the words the brand uses, and the ones it never uses.
- The approval record for each entry, and the schedule on which it is re-checked.
Why each entry is approved by a person
Extraction is a first draft. A model reading a policy page can misread a condition, and a fact that is ninety percent right is a liability when it is published on a storefront. Each extracted entry waits for a person to approve, edit or reject it, and the approval is recorded.
The effect is that the knowledge base is small, correct and owned. It is not everything the brand has ever written; it is the set of statements the brand has agreed can be repeated.
Why are facts re-checked on a schedule?
Facts go stale. A returns window changes, a material is substituted, a guarantee is updated. The engine re-checks each fact against its source on a schedule and flags the ones whose source no longer supports them, so the entry can be updated or retired before another article repeats it.
This is the honest-data promise applied to knowledge rather than numbers: a statement whose foundation has moved is marked, not silently carried forward.
How grounding works when an article is written
When the engine writes a long-form article, it draws its claims from approved Brand Brain entries and cites them internally, so a reviewer can see which fact each sentence rests on. A claim with no approved fact behind it does not get written; the draft flags the gap instead.
That is why grounded generation is publishable. The reviewer is checking a draft whose factual claims already trace to approved statements, rather than fact-checking free text from scratch.
What the Brand Brain is not
It is not a training set; the brand's documents are not used to train a model. It is not a chat memory of past conversations. And it is not a place for customer data; it holds brand knowledge, not personal information.
It also does not publish. Approved facts make drafting safe; publishing still goes through Review & Send, and on the storefront that means a person approves the article before it appears.
Which roles use it on ftes.ai?
The Visibility Officer FTE and the Content Officer FTE write from it and are judged on cited reach and published output that holds up. The Social Officer FTE uses the voice samples and word rules for posts. The Order Support FTE reads policy facts when drafting replies, so support and marketing quote the same policy.
A brand starting out should load the policies, the size guides, the materials and three good pieces of existing copy. That is usually enough for the first grounded article, and the base grows from there.
How big should a Brand Brain be?
Smaller than a brand expects. Thirty to fifty approved facts, three or four personas, a handful of voice samples and a clear word list will ground most of what a brand publishes in its first months. A large base of unapproved extractions is worse than a small approved one.
The discipline is to add a fact when a draft needs one and flags the gap. That keeps the base made of statements the brand actually uses, and it keeps the re-check schedule manageable, because every entry in it is worth re-checking.