How do you stop an AI from inventing facts about your business? On a real project, I measured the problem, then reduced it: invented claims went from almost one FAQ answer in five to fewer than one in a hundred. Here’s how to reduce AI hallucinations in content, based on what I went through. It’s one of the safeguards of AI content where every claim has a source.
In June 2026, I was testing my FAQ generator on a real case: a residential services franchise network. The answers were well written, relevant, on brand. Then I checked them one by one, fact by fact.
The result: 18% of the answers contained at least one invented claim about the company. Almost one in five. These weren’t debatable opinions, but precise facts: conditions, figures, commitments the company had never published.
That number taught me one thing: without measurement, you see nothing. On reading, these FAQs looked flawless. Only by checking every claim against a source did the problem appear.
The most instructive episode came from a comparison. An older version of my writer wrote better than the new one: more fluid, warmer, more engaging answers. I was tempted to go back.
Before doing so, I checked the facts. That elegant version was inventing. It announced a warranty condition that contradicted what the company publishes. It put forward price ranges out of nowhere. It named a competitor for no reason at all.
The lesson changed the way I work: quality of style and reliability of facts are two separate things. A text can excel at one and be dangerous on the other. So you have to measure them separately, and never let the first make you forget the second.
Why does an AI invent? Often because we ask it to answer without giving it the answer. It then fills the gap with whatever seems plausible, drawn from its general memory. A “typical” warranty for the industry becomes your company’s warranty.
The first change is therefore to give the AI, at writing time, the company’s verified facts: what it actually publishes on its own website. The AI no longer has to guess. It answers from what exists.
This change greatly reduces inventions, but doesn’t eliminate them. Even well informed, an AI can still slip in a real but unsourced fact, or round off a figure. That’s why a second change is essential.
The second change is an automated check, once the text is written. Every figure and every fact stated about the company is compared with the verified facts. If it isn’t there, it’s flagged and removed, or the answer is redone.
It’s a direct application of a rule I keep everywhere in the pipeline: an honest gap beats an invention. If no source supports a claim, we don’t make it. The answer can stay useful without that detail, by pointing to what can be verified.
With both changes combined, the rate of invented claims went from about 18% to about 0.5%. In other words, from almost one answer in five to fewer than one in a hundred. It isn’t zero, which is why a human always reviews before publication. But the gap is large enough to change the nature of the review work.
Along the way, I made a discovery that surprised me. As I improved the writing instructions, style flaws disappeared quickly: fewer defensive phrasings, more answers that truly addressed the reader’s concern. But unverified claims about the company only dropped by half.
The reason was simple: the writer kept stating real facts, but drawn from its general memory rather than from a source. Sometimes accurate, sometimes outdated, never verifiable. No instruction could fix that, because the problem didn’t come from the writing. It came from the data.
That’s what led me to build a complete file of the company’s facts, taken from its own website, and to make the automated check mandatory. A better pen doesn’t replace a better source.
If you entrust your content to an agency or a tool that uses AI, three questions will quickly tell you where you stand:
These three questions also apply to your own team, if it produces content with ChatGPT or another assistant.
An FAQ or a service page commits your business. A poorly described warranty can create a dispute; an invented price can create an expectation you can’t meet. By measuring inventions and bringing them below 1%, AI goes from being a risk to being a reliable production tool.
This work also rests on a broader distinction: knowing how to tell what the market asks for from what you offer. The company’s verified facts come only from the company itself.
Reducing hallucinations is just one piece of the method. To see how it fits into the whole, from listening to the market to monthly measurement, discover the AI SEO pipeline.