Fact-checking a company's own content mostly means finding the places where two of its own sources say different things, then asking the person who owns that topic to pick one. Your website might give one price and your sales deck another, while somebody quoted a third on a call. Software can find that disagreement in seconds, but it has no way to know which version is current. Whoever owns the topic at the company knows, and once they decide, we record their answer with their name and the date.
That's the whole method. Below we explain how we run it inside Propaganda, and how a founder could run it with a spreadsheet and a spare afternoon if they'd rather not pay us. (We'd rather they paid us. We'll live.)
A fact starts as a claim with a source
When we built the knowledge base, we decided a company fact would be the smallest thing a person could disagree with. "The paid plan costs $500 a month" is one. "We're affordable" is a mood. Every claim in the knowledge base carries its source and the quote it came from, so anyone reading it can see who said it and in what words. A claim with no source stays out, however true it sounds and however senior the person who said it.
This sounds fussy until you open a company wiki, where docs rot. We think a brand book is right only on the afternoon someone writes it, and after a pricing change, a renamed product and two new hires, it shows a company that no longer exists. On a wiki page, you can't tell who said what or when, and a stale sentence sits beside a fresh one in the same font. With a short claim and its quote attached, you at least know whom to ask.
We borrowed the instinct and made it stricter for a company's facts about itself, where the source is usually a call, a Slack thread or a page the company wrote and then forgot about.
One part is allowed to change what the company believes
The Guardian, which is in beta, is the only part of Propaganda allowed to change what the company believes. When Listen reads a call transcript or a Slack thread, it pulls out candidate facts. None of them goes into the knowledge base on its own. The Guardian looks at each candidate and does one of three things. The Guardian admits the candidate when the source holds up and no existing claim conflicts with it. It contests the candidate when it clashes with something the company already believes. And it routes the candidate to a person when only a person can settle it.
We built a single gate on purpose. If the Writer, the Strategist and the Checker could each update the company's beliefs, the knowledge base would end up holding whatever the last agent wrote into it, which is roughly how most sales decks get written. With one gate, there is one place to look when something goes wrong, and one log of every belief that changed and why.
Software finds the conflict and a person settles it
When two claims conflict, a model can flag the pair. Picture our pricing page listing a hosted blog in the free plan while a transcript has someone telling a prospect it doesn't. Any competent model will flag that pair. What the model can't see is which statement is current, since the answer lives in somebody's head or in a decision made in a meeting that nobody wrote down.
So we send the contradiction to the person who owns the topic. Pricing goes to whoever sets prices. We send product claims to whoever runs the product. That person reads both versions, each with its source and quote, and picks one, or writes a third that replaces both. Their answer becomes the claim, and the source on it is them, by name, with the date they decided.
We think this is the only honest arrangement for an AI tool that writes for a company. If we let Propaganda guess which version is true, it would guess with total confidence, and the confidence would end up in a published post under the customer's logo. Once we stop guessing, we have to send each question to the right person, and ownership becomes the hard part. Our view stops at facts about the company itself. Whether a market is growing or a competitor's feature works is a question for outside sources, and the knowledge base leaves those to them.
The plainest case of a person overruling software on our side came from our own plan. When we started Propaganda's blog, we answered the Strategist's questions and approved its plan. One target the Strategist questioned, we kept anyway. We made that call ourselves. Our strategy page now shows that we kept it.
Running it without us
You can run the method in a spreadsheet. Write each fact as one sentence. Next to it, paste the quote and the link or recording it came from, and drop any row that has neither. Sort by topic and read down each group for pairs of rows with different answers. Put an owner's name on every topic, send each conflict to that owner, and write their answer back in with their name and the date. Then hunt down the losing version everywhere it appears, on the site, in the deck and in old posts. That last step is the one people skip, and it is the one the rest of our product exists to do.
The Checker reads the draft first
Before a draft reaches the customer, the Checker reads it against its sources and the knowledge base and leaves its comments in the editor, where a human reviewer would leave them. A sentence quoting a price the knowledge base doesn't hold gets a comment. So does a statistic whose link points to a page with a different number. You open the draft with those notes already on it.
This is where AI writers tend to embarrass their owners. Fluent prose carrying plausible facts is worse than wrong, since a plausible error survives a skim. Given the choice, we'd rather a draft arrive with three awkward comments than arrive clean and false. Even so, the Checker has a limit of its own. Because the Checker works from the knowledge base, it will pass a wrong claim that got admitted, which is why the gate before it matters so much.
When a fact changes, old posts get flagged
Our own facts will expire too. Prices go up, features ship, a plan gets renamed. When a claim in the knowledge base changes, Keep it true, also in beta, flags every piece that used the old version and attaches the fix. Your post from March stops quoting last year's answer, and you don't have to remember the post existed.
Without that step, a company blog turns into an archaeological dig. Old pricing, retired features and positioning from two rebrands ago sit under the new material, all indexed and all findable by a buyer who searched for something specific. If you correct only new drafts, you fix the future and leave the past lying around in search results. All of this runs on the paid plans. Propaganda Light, the free one, gives you the editor and the blog with no knowledge base behind them.
Where the method stops
Sometimes nobody at the company knows which version is right. Suppose the website listed setup as a day, the sales deck listed a week, and nobody had ever timed it. Then the Guardian keeps the fact contested until someone decides, and the Writer leaves it out of every draft in the meantime. You publish the post without the number.
That is a limit, and an annoying one. You may publish a post that is thinner than you wanted. Some get delayed while the people who own the topic argue over something they should have known years ago. We'd still take the delay over a published guess. If you learn from your own knowledge base that you have no answer, you get to fix that in a meeting. A company that learns it from a customer gets to fix it in public.