The brand knowledge base your agents actually cite
Generic AI copy is a retrieval problem, not a model problem. What to put in a brand knowledge base, and how to tell whether the agents are really using it.
Every team that tries AI copy hits the same wall: the output is fluent, on-topic, and unmistakably generic. The instinct is to blame the model and go shopping for a better one. Usually the model is fine — it simply has nothing specific to say about your brand, because nobody gave it anything specific.
What actually belongs in a brand KB
A brand knowledge base is not a folder of PDFs. It is the small set of facts an agent must be able to quote before it writes anything:
- Positioning and the one sentence you would defend in a room — plus what you explicitly are not.
- Claims you are allowed to make, and the evidence behind each. This is what keeps a confident model from inventing a statistic.
- Voice rules with examples of both good and bad, because "friendly but professional" means nothing without a sample.
- Offers, pricing rules, and the constraints that are non-negotiable (regulated wording, prohibited comparisons, regional differences).
- Recent performance context: which angles worked, which fell flat, and the audiences behind those results.
A crawl of your own site gets you a usable first version of most of this in minutes; the value comes from correcting it afterwards. The corrections are the part competitors cannot copy.
Citations are the only real test
The way to know whether retrieval is working is to require citations and then read them. If an agent claims your onboarding takes ten minutes, the claim should point at the source that says so. If it points nowhere, the sentence is decoration and should be treated as a draft, not a fact.
This is also the fastest way to find gaps in the knowledge base itself. Claims that repeatedly cite nothing are usually claims your documentation never made — write them down, and the next discussion is measurably better grounded.
Keeping it alive
Knowledge bases rot. Prices change, a claim gets legal review, a campaign changes the positioning. Treat the KB like a small product: an owner, a review cadence tied to your campaign calendar, and a habit of correcting the source the moment an agent quotes something stale rather than fixing the output and moving on.