A Knowledge Base to Keep AI-Written Content Factual

AI drafting has one failure that matters more than weak prose: confident wrong facts. A model asked to write about your pricing without your pricing will produce a number that sounds right. Publish it and you have a page that gets cited, and cited wrongly. The fix is boring and effective: give the writer a file of verified facts and make it write from that.

What a knowledge base is for

Think of it as the single source of truth your writers, human or machine, must use. It holds the claims that are expensive to get wrong: product names, plan names, limits, prices, feature availability, dates, legal wording. Anything that would embarrass you in a screenshot goes in.

Promptwatch's Knowledge Base is available from Professional up, with 100 items on Professional and 150 on Business. It sits next to content creation, so drafts can draw on it. Agency plans do not include the Knowledge Base, so check your plan before building a process around it.

What to put in

Product facts

Plan names, what each includes, and what it does not. The "does not" matters as much as the "does." Features gated to higher plans cause the most errors.

Company facts

Legal name, location, founding date, leadership and certifications. Use the wording you want repeated.

Claims you can support

Only include a claim if you can point to where it came from. Add the source and the date next to it. If a number comes from a customer case study, say so, and attribute it to the customer, not to your product in general.

Banned claims

A short list of things not to say. Retired customers, discontinued features, old pricing and anything legal has asked you to avoid. We keep this list for our own writing, and it has caught more errors than any style guide.

Voice rules

How you refer to the product, which terms to use, which to avoid. This keeps drafts consistent.

How to source it

Go to primary sources: your pricing page, your documentation, your own llms.txt, your release notes. Record the date you read each one. Facts go stale, and a date tells the next person whether to trust an entry or recheck it.

For a competitor, the rule is stricter. Include only what you can verify from their own pages or from a catalog you trust, and mark anything unverified as unknown. "Unknown" is a valid entry. A guess presented as fact is not.

Structure for usability

Short entries beat long essays. One fact per item, with source and date. Group by topic: pricing, features, company, claims, banned. This is also what makes retrieval work well, since a drafting tool can pull the relevant entries rather than wading through a manual.

Review cadence

Set an owner and a schedule. A reasonable pattern: price and plan entries checked monthly, feature entries each release, company facts quarterly. Mark each entry with a last-reviewed date and look for the oldest ones when you review.

How it connects to the publishing workflow

The sequence we recommend is:

1. Find the gap. Content gap analysis with coverage scores shows prompts where you lack a good answer.
2. Draft from the brief and the knowledge base. On-demand Content Creation takes a brief; on Business and Enterprise, the Content Agent plans, writes and publishes on a schedule.
3. Review. A review inbox gives a human a chance to catch errors before publishing. Hands-off mode with a quality check suits mature processes only.
4. Publish. Through WordPress, Webflow or Framer.
5. Measure. Add the page to the Page Tracker and read citations, crawls and clicks.

A knowledge base does not remove the need for review. It reduces how many errors reach the reviewer.

Testing whether it works

Take ten drafts produced with the knowledge base and ten without, if you can, and check each against your facts. Count wrong claims. If the count is not noticeably lower with the base, the entries are probably too vague or the retrieval is not finding them. Fix the entries.

Common mistakes

• Dumping marketing copy in. Marketing language is not a fact file.
• No dates. You will not know what to recheck.
• No banned list. Old claims creep back from older pages.
• Too many items at once. Start with the fifty facts that appear most often, then grow.
• Letting it rot. An outdated knowledge base is worse than none, because it carries authority.

Facts that matter for GEO

AI answers repeat what pages say. If your pricing page and your blog disagree, models may repeat either. A knowledge base fixes this at the source: one canonical statement, used everywhere. It also helps you spot old pages that contradict it, which then become refresh candidates in Content Refresh.

Recommendation

If you are going to use AI to draft at volume, put the facts first. Promptwatch offers a Knowledge Base on Professional and up, and connects it to content creation, review and publishing, with Page Tracker closing the loop on what happened afterward. Start with fifty sourced entries, assign an owner, and date each one. The first time a draft gets a price right because the file said so, you will see the point.