Build a Marketing Knowledge Base AI Can Keep Current

A folder full of documents is not automatically a reliable knowledge base. AI needs material that is current, attributable and organized around questions the team actually asks. Start with a small set of authoritative records and a clear update process.

Define what belongs

Include approved product facts, audience definitions, brand guidance, documented workflows and relevant research notes. Separate current policy from historical material and drafts.

Do not combine conflicting documents and hope retrieval will select the right one. Identify the authoritative source for each fact and preserve the relationship between old and current versions.

Give each record useful metadata

Access permissions should follow the sensitivity of the material. A marketing assistant does not need every internal document to answer routine product questions.

Worked example: two pricing documents

Suppose a hypothetical knowledge base contains last year's pricing presentation and the current approved price sheet. Both use similar language. A retrieved answer may select the older one unless the system and content clearly distinguish their status.

Mark the presentation as historical, identify the current source and test questions involving the changed plans. Do not delete legitimate history simply to make retrieval easier; control its use and labeling.

Test retrieval and answers separately

Prepare a small question set with known answers and required source records. Check whether the right material is retrieved, then whether the final answer accurately reflects it.

Include questions the knowledge base cannot answer. A useful system should admit missing evidence rather than invent a feature, customer quote or policy.

Maintain through events, not hope

Connect product releases, policy changes and approved positioning updates to review tasks. Record which documents and answer tests are affected. Schedule periodic checks for material without a clear event trigger.

Track stale-answer incidents and the effort required to resolve them. If the same conflict recurs, improve ownership or document structure rather than adding more prompt instructions.

The goal is a dependable source of answers the team can trace back to current evidence. More documents help only when they improve coverage without making authority and freshness harder to determine.

Why "just add all our docs" fails

The intuitive first version of a marketing knowledge base is a folder with everything in it and a retrieval layer on top. It answers simple questions well for a few weeks, then starts producing answers that are fluent, cited, and wrong. The mechanism is predictable: retrieval returns the passages most similar to the question, and the most similar passage to "what does the Business plan cost" is often in a two-year-old sales deck, because that deck talks about the Business plan at length and the current price sheet is a terse table.

Three properties of an unmanaged corpus produce this:

• Duplicates with drift. The same fact in five documents, each edited at a different time, so retrieval picks by similarity rather than by currency.
• Drafts and rejected versions indistinguishable from approved material.
• Context-free chunks. A pricing table split from the heading that says which year and region it applies to.

The fix is not a better model. It is curation: fewer documents, each with a status, an owner, and a date, and a retrieval layer that can be told to prefer current over historical.

A record template that makes authority visible

Every record in the base carries a short header that both people and the retrieval system can use. Keep it to what changes behavior:

title: Business plan pricing (current)
status: current            # current | historical | draft | withdrawn
owner: pricing-owner@
source: internal price sheet v14, approved 2026-08-30
effective_from: 2026-09-01
supersedes: Business plan pricing 2025
review_trigger: any pricing change; quarterly check
audience: sales, marketing, support

The supersedes field is the one teams skip and regret. It links versions so that historical material can stay in the base, labeled, without competing with the current record. When a system is asked about last year's price, it can find the historical record and say so; when asked about today's, the status field lets the retrieval layer filter or the prompt instruct the model to prefer current.

An evaluation set you run after every change

The question set described above is the knowledge base's test suite, and like any test suite it is only useful if it is run. A workable starting set covers four kinds of question, each with the record that should be retrieved and the answer that should result:

Score retrieval and answer separately, because they fail separately: the right record can be retrieved and misread, or the wrong record can be retrieved and faithfully summarized. Run the set after every batch of document changes and after any change to the retrieval configuration or model, and add each real stale-answer incident to it as a new case. Over time the set becomes a record of what has gone wrong before, which is the most reliable predictor of what will go wrong next.

Put this into practice

Copy the worksheet columns below into a spreadsheet and keep one row per item you check. The filled row is an illustrative example, not a reported customer result; replace it with your own verified records.

Use the following prompt only after supplying the records it requests:

Answer using only records marked current and authorized for this task. Cite record IDs and effective dates. Flag conflicts and missing evidence. Do not use historical or draft records as current policy unless explicitly requested.

Research context

A marketing knowledge base needs source ownership and maintenance as well as retrieval. The related Ahrefs starting points are How I Use My AI Marketing Assistant After 200+ Hours and What Is Content Engineering, and How Do You Do It?. This guide’s checklist, examples and proposed workflow are independently written; they are not results of a SEOVision experiment.

Continue with the next task

• Build an Evidence Ledger for AI-Assisted Articles
• Training, Retrieval and Tools: Diagnose an AI Answer’s Source
• SEO Prompt Engineering: Reusable Briefs, Guardrails, and QA
• Technical SEO Audit Checklist: A Practical 30-Minute Workflow

Sources

• How I Use My AI Marketing Assistant After 200+ Hours — Research starting point; not an endorsement of this original workflow
• What Is Content Engineering, and How Do You Do It? — Research starting point; not an endorsement of this original workflow