A Knowledge Base for AI-Written Articles: What to Put In It
AI-drafted content is only as accurate as the facts it is given. How to build a knowledge base of verified facts, voice rules and sources so drafts need less fixing.
The most common complaint about AI-generated articles is not tone. It is facts. Prices that are wrong, features that do not exist, a customer quoted who never said it. Models fill gaps with plausible text, and a plausible falsehood is worse than an honest gap.
The fix is not a better prompt. It is giving the writer something true to work from. That is what a knowledge base is for.
What a knowledge base does in a content workflow
A knowledge base is a set of items the writing system can draw on: product facts, positioning, approved claims, source documents, style rules. Instead of asking a model to "write about our pricing", you point it at the item that states your pricing and tell it to use only that.
Promptwatch offers a Knowledge Base on the Professional plan and above, with 100 items on Professional and 150 on Business. It sits alongside author profiles, a rich text editor with AI images, content creation from a brief and Content Refresh.
Why accuracy is a visibility issue
It is tempting to treat accuracy as a quality concern separate from AI visibility. It is not. Engines compare sources. A page that contradicts what other credible sources say about your product may be treated as less reliable. And a wrong claim that gets quoted in an answer becomes a reputation problem you then have to chase.
There is also a plain ethical point. Readers rely on what you publish.
What belongs in it
Think in categories, and keep each item short and single-purpose.
Product facts. Features, plans, limits, integrations, supported platforms. One item per area. Include the date you last checked each.
Approved claims. Statements that legal or leadership have cleared, such as certifications or awards. Include exact wording and the evidence.
Prohibited claims. Things the writer must never say. Superlatives you cannot prove, comparisons to named competitors without verified data, unreleased features.
Customer proof. Case studies, with who said what and where it was published. Attribute them to the named customer and do not generalize them into promises.
Audience notes. Who you write for, what they already know, what they worry about.
Voice and style. Sentence length, terms to use and avoid, formatting rules. Include examples of good and bad paragraphs.
Source documents. Product documentation, pricing pages, policy documents, research you have rights to use.
How to write items
Write plainly. State the fact, give the source, give the date. A good item reads like a note from a careful colleague.
Bad: "Our platform offers powerful analytics."
Better: "Visitor analytics reports AI-referred traffic and conversions. Source: product docs, checked on a given date."
Add a short "do not say" line where a fact is easy to overstate. For example, if a feature exists only on a higher plan, say which plan and say not to imply otherwise.
Competitor facts are the risk
Competitor claims cause the most damage. Writers, human or machine, love a comparison. If your knowledge base contains competitor details, each needs a source and a date, and anything you cannot source should be left out. When in doubt, the instruction should be: do not state it.
This is the core discipline of review content. A comparison that is wrong about a rival's price will be quoted and shared, and corrections rarely travel as far as the error.
Keep it current
A stale knowledge base is a liability. Set a review cadence. Product facts monthly, pricing whenever the pricing page changes, case studies when the customer asks. Give each item an owner. When a fact changes, update the item before the next article, and consider whether published pages need a refresh.
Content Refresh helps with the second part. It produces diff-style suggestions for existing pages, so after a fact changes you can find and update affected articles without rewriting them.
Connect it to the briefs
A knowledge base does not replace a brief. The brief says what to write and for whom, drawn from the data: which prompts, which gaps, which sources. The knowledge base says what is true. Together they reduce the editing load, which is the point.
Content gap analysis and query fan-outs feed the briefs. The Action Board lists the gaps in priority order. The Content Agent, on Business and Enterprise, can plan, write and publish on a schedule. Even then, start in review mode and read the first batches closely before moving to hands-off with a quality check.
What it will not fix
A knowledge base does not make a weak idea strong or add experience you do not have. It does not guarantee citations, either. Engines choose sources for reasons nobody outside them fully controls. It simply stops you from publishing things that are untrue.
It also needs people. Someone has to write the items and keep them honest.
A starter set of ten
- Company description in two sentences.
- Product overview with plans and limits.
- Three approved customer stories.
- Integrations list.
- Pricing, with date.
- Style guide with examples.
- Prohibited claims list.
- Audience profile for each persona.
- Glossary of terms you use.
- Contact and policy details.
Build those first. Add more only when a draft shows a gap.
The recommendation
If you plan to scale content with AI, build the knowledge base before you scale. Promptwatch includes one from the Professional plan with the content gap, refresh and publishing tools around it, so the facts, the briefs and the pages live in one workflow. Take a day to write the ten items, then try a draft. See the platform at promptwatch.com.