← Blog
Posts

We check every company fact before any draft uses it

We write down the facts that embarrass a company, compare them, hand disagreements to a person and flag old posts when a fact changes.

Ghaith Ayadi · 11 Nov 2026 · 5 min read

The search results for fact-checking AI content fall into two piles. One pile teaches you to sniff out machine-written prose by its tics and favourite words. In the other pile, you click every citation a chatbot gave you, confirm the page exists, and confirm the page contains what the bot attributed to it. Both piles were written for facts about the world: a population figure, a court ruling, the year a law passed.

The facts that embarrass a company are about the company. The price in a post from two years ago. A feature listed as shipped in a post while it still sits on the roadmap. The discount one sales rep offered on one call, which a draft later presented as policy. ChatGPT cannot check those, and Google will not contradict them, since the only record lives in the company's calls, its Slack and its own pages. We found one way to check them, and it is dull. Write each fact down once. Compare each new one with the ones already there before a draft may use it. And give every disagreement to a named person.

A fact goes in as one short claim

We built our knowledge base around a single rule. We write each company fact down as one short sentence, with the place it came from and the exact words of the source. The place is a call transcript, a Slack thread or a page on the company's site. We copy the words as a quote, exactly the way they were said or written.

We chose short claims over documents after watching documents rot. Your brand book was probably right on the day somebody wrote it, and then a price moved, a plan got renamed, and nobody opened the file again. A one-line claim with a quote attached can be compared, contradicted and replaced on its own. You never rewrite a whole document to fix one sentence.

One part decides what the company believes

Before a new claim counts, it gets compared with the claims already there. We gave that job to one part of Propaganda, the Guardian, and to nothing else. The parts that read call transcripts and Slack can propose a claim. The Writer can use a claim. Only the Guardian can change what the company believes.

With a candidate in front of it, the Guardian does one of three things. We set it to admit a claim that matches what is written down or adds something new. It contests the claim when it collides with one that is already there. And when the comparison comes back unclear, it sends the claim to a person.

We kept that door narrow on purpose. We think a knowledge base with five doors ends up holding whatever came through the last one, and a draft built on it inherits the mess with total confidence.

The owner of the topic settles it

When two claims conflict, we do not let the newer one win by default. A rep's improvised answer on a call will always be newer than the pricing page, and a later date is no evidence of who was right. So we send each contradiction to the person who owns its topic, and the owner of pricing settles a pricing fight while the owner of the product settles a fight about what ships.

That person picks the claim that stands, and the other one is retired with its quote still attached. If nobody owns the topic, you have learned something more useful than a wrong price.

Our own pages disagreed about what we sell

The best contradiction we can show came from our own material, which is the most embarrassing place to find one. We answered it the way we tell customers to. The claim that stood was the boring one, which is usually how these go.

A draft gets read against its claims first

Before you see a draft, the Checker reads it against the claims and outside sources it used. The Checker leaves a comment in the editor on any company fact with nothing behind it. So does a sentence that conflicts with a claim in the knowledge base.

Nothing gets published without your approval, and the Checker does not change that. You still read the draft. The difference is that the argument about whether a sentence is true happened before the draft reached you, and the comment shows you which claim the sentence leaned on.

Old posts get flagged when a fact changes

Published posts keep talking long after anyone at the company has stopped reading them. When a claim changes, Keep it true flags every piece that used it and attaches the fix. Raise a price and the post from last year that quoted the old one shows up in the editor with the new sentence ready.

We built it for the post that ranks. Search keeps sending strangers to a page for years, and a stale fact on a page that gets traffic does more damage than a stale fact on a page nobody visits.

Researchers split fact-checking into the same steps

None of this was our invention. In a 2024 survey of claim detection, Rrubaa Panchendrarajan and Arkaitz Zubiaga of Queen Mary University of London laid out automated fact-checking as a series of steps: identify the claims to check, prioritise the important ones, gather the evidence, then give a verdict by checking the claim against that evidence.

In that model, a company's knowledge base plays the evidence store. And we think the choice of evidence store matters more than the vendors of these tools let on. When Dominik Stammbach, Boya Zhang and Elliott Ash studied the choice of knowledge base in automated claim checking in 2021, they found no universally best knowledge base.

For a claim about your price, the closest knowledge base on earth is your own record of what you charge. Wikipedia has never heard of your discount.

Checking stops where nobody ever disagreed

Our process catches a fact that conflicts with a claim already written down. It cannot flag a written-down fact as wrong if nobody ever said otherwise. If the only claim about a feature came from one confident call, and no page, thread or colleague contradicted it, the Guardian has nothing to compare it against.

We keep opinions outside the process entirely. "We think" has no source to check against, and neither does "best", "fastest" or "the only platform that". Those words stay out of the knowledge base, and whether they belong in a post is your call to make in the editor, with your name on it.