← Research library

A marketing knowledge base is the best defense against fabricated claims

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

Content agents now draft a growing share of marketing copy, and the quality of that copy depends less on the model than on what the model can look up. When an agent invents a product feature or misstates a customer result, the usual cause is a gap rather than a reckless model.

The agent needed a fact, had no approved source for it, and produced something plausible instead. Plausible is the problem. A fabricated integration or an inflated number reads as fluently as a true one, and reviewers under time pressure can miss it.

The most effective fix is to make sure the right facts are always available, in a form agents can use. That is the job of a marketing knowledge base built for agents, and in our research we treat it as infrastructure rather than documentation.

Why agents fill gaps with plausible text

Language models are trained to continue text in ways that fit the context. Given a brief that says "mention our integrations" and no list of integrations, the model fills the slot with names that seem likely. It is doing what it was built to do.

Instructions such as "do not make things up" help at the margin, but they cannot supply a fact the agent does not have. An agent that must produce a paragraph about pricing tiers will produce one. The reliable fix is upstream: put the approved fact where the agent will find it, and make finding it part of the task.

This also changes how you read errors. When a reviewer catches a wrong claim, the first question is whether the right fact existed in a place the agent could reach. Often it did not.

What belongs in it

A useful knowledge base for marketing agents covers six kinds of material. Each answers a question a content agent will otherwise guess at:

  • Product facts: what each product and feature does, who it is for, integrations, limits, and availability.
  • Positioning: core messages, differentiators, and how you describe your category.
  • Approved proof: customer examples with their permission status, numbers you may cite with their sources, and certifications.
  • Voice and style: your brand voice specification, banned terms, and formatting rules.
  • Policies: legal requirements, claims that need review, and topics to avoid.
  • Glossary: definitions of terms your company uses, especially ones with specific internal meanings.

Approved proof deserves the most care. Customer names and results usually come with conditions attached, such as a signed release that covers one channel only. Record those conditions on the entry itself, so the agent can respect them without guessing. Voice belongs here too, and brand voice as a specification covers how to write it so agents can apply it.

Principles that make it trustworthy

A knowledge base earns trust through how it is maintained, not through how much it contains. Five principles do most of the work.

One source of truth per fact. If a product's capabilities are described in five places, they will drift apart. Pick one canonical entry and link to it from everywhere else.

Owners and review dates. Every section needs a named owner and a date by which it must be reviewed. Product facts go stale fastest, so give them the shortest review cycle.

Explicit status. Mark entries as current, deprecated, or upcoming. Agents should never describe an unreleased feature as available, and status is what lets them tell the difference.

Small, focused entries. Agents retrieve information more accurately from short, specific entries than from long documents covering many topics. One fact or one tight cluster of facts per entry is a good default.

Plain language. Write facts as direct statements. Marketing copy is a poor source of facts because it blends claims with persuasion, and an agent cannot tell which sentence is the claim.

A worked example: one feature entry

Consider a hypothetical analytics product with a new export feature. A useful entry is short. It states what the feature does in one sentence, lists the plans that include it, and names the three file formats it supports. It notes one known limit, a maximum export size.

The entry also carries metadata. Its status is current, its owner is the product marketing lead for analytics, and its review date is set for the next release cycle. A final line says that any comparison with competitors requires legal review.

When a content agent drafts a launch email, it retrieves this entry and cites it. A checker agent then confirms that every product claim in the draft maps to a cited entry. Any claim without a citation goes back to the writer agent or to a person.

Connecting it to agents

Agents should look facts up rather than rely on whatever their instructions happened to include. Give them a search or lookup tool scoped to the knowledge base, and instruct them to cite the entry they used for each factual claim. We discuss retrieval design in retrieval for marketing agents.

Citations change the review job. A reviewer no longer has to know every product fact. They check that each claim matches its cited entry, which is faster and easier to delegate. In a multi-agent content pipeline, a dedicated checker role can do most of this verification before a person sees the draft.

Some rules belong in code rather than in the knowledge base. A banned claim or a required disclaimer is better enforced by an automated check, as described in guardrails as code. The knowledge base supplies facts. Guardrails enforce limits.

What to measure

A knowledge base needs its own health measures, separate from the quality of any single draft. We recommend tracking four signals:

  • Citation coverage: the share of factual claims in agent drafts that cite an entry.
  • Unsupported claims: claims that reviewers flag as missing from the knowledge base.
  • Overdue entries: entries past their review date, grouped by owner.
  • Correction source: whether each factual correction traced to a missing entry, a stale entry, or a retrieval miss.

The last measure is the most useful for deciding where to invest. Missing entries call for authoring work. Stale entries call for an owner conversation. Retrieval misses call for changes to how entries are written or searched.

Start where errors happen

You do not need to document everything at once. Review the last month of edits and rejections on agent content, and sort the factual corrections by topic. The facts that were wrong or missing most often are where to start.

As the knowledge base grows, the hard part shifts from writing entries to keeping them true. The question we are studying next is how to detect drift automatically. One approach is to compare entries with release notes and pricing pages, so owners hear about a stale fact before an agent repeats it.

How to apply this: create canonical entries for your top five products or features, each with an owner, a status, and a review date. Point your content agents at those entries and require a citation for every product claim.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.