THE RECOMMENDATION LAYER · AUGUST 10, 2026 · 9 MIN READ

How Do You Stop AI Content From Inventing Your Expertise?

Provenance rules, not proofreading: every claim traces to something you said or a source that exists, or it does not ship. Nothing publishes under 80 of 100.

By Jacquie Baker
TL;DR: WHAT TO REMEMBER

You stop AI visibility content from inventing expertise under your name by governing provenance, not prose: a standing rule that every claim, number, credential and case detail must trace to something you actually said or a source that actually exists, enforced by gates that refuse to publish anything that cannot show its receipts. Careful writers are not the fix. Systems that check are, because the failure this guards against is fast, fluent and disturbingly easy to miss.

You have probably had the moment. You open the draft the agency sent for AI visibility and stop on the third paragraph. A sharp claim. A tidy case detail. A sentence that sounds exactly like you, until you realize you have never said it to anyone. Somewhere between the brief and the deadline, a machine or a stranger decided what your expertise ought to be.

That is the moment publishing stops feeling like progress and starts feeling like a liability with your name on it. The brakes go on. They should. You just caught a system improvising you. But the conclusion most experts draw next, that publishing itself is the risk, doesn't follow. The risk was never the publishing. It is publishing without provenance.

Here is why invention is the default behavior of ungoverned content systems, the rule that actually stops it, and what the enforcement machinery looks like when it is real rather than promised.

The Third-Paragraph Moment

Start by taking the fear seriously, because it is better-founded than most experts know. The invented sentence in that draft is not an occasional glitch. Language models produce confident, specific, wrong material as a routine matter, and the best measurement of it comes from the news industry, which checks obsessively.

In October 2025, the European Broadcasting Union and the BBC published a study of more than 3,000 AI assistant answers about the news, across 22 public service media organizations, 18 countries and 14 languages. Forty-five percent of answers contained at least one significant issue. Twenty percent had major accuracy problems, including fabricated details. A third carried serious sourcing errors: missing, misleading or incorrect attribution.

Now notice what that means for you. Those systems were summarizing published, checkable journalism and still fabricated a fifth of the time. A content vendor generating articles about your unpublished expertise gives the machine far less to be right about. Where the record is thin, fluent systems do not go quiet. They fill.

The stakes compound because of where the words land. A wrong sentence in a chat window evaporates. A wrong sentence published under your byline becomes part of the permanent record that buyers and AI engines use to verify you, and part of what every machine reads the next time it checks your story. An invented credential under your name is not a marketing error. It is a public exhibit against you.

Why Plausible Fiction Is the Default

Understanding why this happens beats being angry at it. Generative systems are fluency machines: they produce the most plausible next sentence, and plausibility is judged against everything ever written, not against what is true of you. When a draft needs a differentiator, a statistic or a client anecdote to land, the most plausible version gets written whether or not it exists. Fluency is not provenance.

The ungoverned content workflow makes it worse by assigning verification to nobody. The writer assumes the tool was grounded in your materials. The account manager assumes the writer checked. You, the only person alive who could actually catch the invented case detail, are handed a stack of drafts and a Friday deadline, which is how the expert who bought a done-for-you service becomes the unpaid fact-checker of fiction about themselves. Everyone in that chain acted reasonably, and the system still shipped an invention.

Be precise about the villain here: it is not AI writing. The same failure shipped for decades out of human content mills, just slower and with fewer decimal places. The failure is any system where checking is a courtesy rather than a gate, and AI simply industrialized the output of that system. The debate about human versus AI writing is the wrong debate. The right debate is whether anything stands between plausible and published.

Which is exactly where the fix lives. Not better intentions, and not a better prompt. A gate.

The Rule That Fixes It

The rule is provenance, stated bluntly: no claim without a source, and for claims about you, the source is a record you control. If you did not say it, approve it, or publish it, and no verifiable external source holds it, it does not ship. Not softened, not "probably fine," not padded with a qualifier. Cut, or traced.

Making that rule workable takes two structures. First, capture: your expertise has to get onto the record in your own words before anyone writes anything, which is the job of a recorded interview about your cases, distinctions and opinions. Writers then develop what you actually said. The judgment in the articles is yours because it started as yours, the whole mechanic behind turning one conversation into a month of thought leadership.

Second, an approved-facts record: a maintained list of what may be claimed about you, your numbers, your credentials, your named clients, with everything else off-limits by default. The writer's question stops being "what would sound good here" and becomes "what am I allowed to say, and where is it written." Deny-by-default sounds bureaucratic until you watch it delete a fabricated statistic before anyone ever saw it.

The register changes too, and readers feel it. Governed copy doesn't reach for the biggest claim available; it reaches for the one it can prove, which is why it reads calmer than the category around it. Buyers who've been burned by content-speak read that calm as competence. They're right to.

Under a provenance rule, a thin draft is a signal, not a failure. A missing proof point means an interview gap to fill next month, never a blank to write around. The system is allowed to say less than it wishes it could. It is never allowed to say more than it knows. Less, proven, wins.

The Gates That Enforce the Rule

Rules without enforcement are wishes, so the question to put to any provider is what mechanically stops a bad article. In our operation the answer is three gates, in sequence, and your name does not ride until all three clear.

Every article is scored 0 to 100 through the Integrity Gate, grading whether it answers its question directly, carries proof that traces, cites real sources and reads as your thinking: nothing publishes under 80. Independent fact-checking models then attack the draft's claims against the live web, and an editorial pass reads it the way a skeptical buyer would. Scores are earned, never assumed; a piece that cannot prove a claim loses the claim or dies. No proof, no publish.

Then you review, and the difference from the third-paragraph moment is the point. You are reading gated, provenance-traced work as the final authority on your own voice, a skim and a nod, not hunting a stack of drafts for fiction with your name on it. Your corrections feed the record, so the system gets more accurately you with every batch.

Ask any vendor three questions before they publish a word: where does a claim about me have to come from, what number does a draft have to clear, and who checks the checker. Vague answers to those are the third-paragraph moment scheduled in advance.

What Governed Publishing Buys You

The reward for the discipline is not just avoided embarrassment. It is that the record starts matching the room. What buyers read in advance is what they get in the first meeting, so calls start from accurate trust instead of inflated claims that need quiet walking back. A record you would stand behind sentence by sentence is also one you stop having to police.

The machines reward the same honesty, measurably. In the Semrush study of 11,882 prompts, published in January 2026, strong E-E-A-T signals, the machine-readable face of demonstrated, checkable expertise, correlated with a 30.64 percent lift in AI citation likelihood, and Google's own content guidance asks for demonstrated first-hand expertise rather than impressive-sounding summary. Verifiable and modest outperforms invented and grand. It's a rare market where the incentives and the ethics point the same direction, and it happens to be the one your name lives in.

Start by finding out what the record currently claims about you, including anything wrong that is already out there. The free Recommendation Check scores your visibility across 109 checkpoints and captures what five AI engines actually say when buyers ask who to trust in your field. If something false is circulating under your name, better to meet it in a report than on a sales call. Corrections are cheaper before the buyer has read the mistake.

Publishing under your name should feel like progress again, and it does, the moment nothing reaches the record that you would not say in the room. That is the entire promise of governed publishing, and the only version of AI visibility worth signing your name to.

Want to Learn More?

Probably Genius was built by Jacquie Baker and Christopher Shaw, who spent more than 40 combined years translating what makes an expert the best in the room for the audiences that decide who gets chosen. Their machine writes from what you said, proves what it publishes, and scores every piece before your name touches it. You're probably a genius at what you do. We make sure AI gets the memo.

How do I stop AI content from making up facts about my business?
Govern provenance instead of relying on careful proofreading. Require that every claim, number, credential and case detail in content under your name traces to something you actually said, a document you approved, or a verifiable external source, with everything else off-limits by default. Structurally that means capturing your expertise in recorded interviews before writing starts, maintaining an approved-facts list of what may be claimed about you, and enforcing a quality gate that scores every piece and refuses to publish below a stated floor. Systems that check beat people who promise to be careful.
Does AI-generated content really invent things that often?
Often enough to plan for it. In a 2025 study by the European Broadcasting Union and the BBC covering more than 3,000 AI assistant answers across 18 countries, 45 percent contained at least one significant issue and 20 percent had major accuracy problems, including fabricated details, even though the systems were summarizing published, checkable journalism. Content generated about an expert's largely unpublished judgment gives a fluent system less to be right about, so ungoverned drafts fill the gaps with plausible invention. The failure is systemic, which is why the fix has to be systemic too.
What should I ask a content vendor before they publish under my name?
Three questions expose most of the risk. First, where does a claim about me have to come from: is there a captured record of my actual statements and an approved-facts list, or does the writer improvise? Second, what does a draft have to clear before it ships: is there a scored quality gate with a numeric floor, independent fact-checking, and an editorial read, or just a proofread? Third, who checks the checker: does my review come last, on gated work, or am I the only line of defense? Vague answers predict invented expertise.
Will strict fact-governance make my content boring or thin?
It changes what thin means. A governed system is allowed to say less than it wishes it could, never more than it knows, so a missing proof point becomes an interview question for next month rather than an invented statistic today. The evidence suggests honesty also performs better: in the Semrush study of 11,882 prompts, strong E-E-A-T signals correlated with 30 percent more AI citations. Specific and verifiable beats grand and invented, with buyers and machines alike.

CITATIONS

  1. "AI's systemic distortion of news is consistent across languages and territories" (European Broadcasting Union and BBC, October 2025). A study of more than 3,000 AI assistant answers about the news, run by 22 public service media organizations across 18 countries and 14 languages: 45 percent of answers contained at least one significant issue, 31 percent had serious sourcing problems, and 20 percent had major accuracy problems including fabricated details. The scale measurement of how routinely fluent systems invent. ebu.ch
  2. "How We Built a Content Optimization Tool for AI Search [Study]" (Semrush, published January 2026). Analysis of 11,882 prompts across ChatGPT Search, Google AI Mode and Perplexity: citation correlated most strongly with clear summarization (+32.83 percent), strong E-E-A-T signals (+30.64 percent) and question-and-answer formatting (+25.45 percent). Evidence that verifiable, demonstrated expertise is what the machines select for. semrush.com
  3. "Creating Helpful, Reliable, People-First Content" (Google Search Central). Google's guidance asks whether content demonstrates first-hand expertise and a depth of knowledge, and warns against material that mainly summarizes what others have to say. The systems are explicitly hunting for real, demonstrated expertise, which only provenance-governed publishing can supply at scale under an expert's name. developers.google.com
WRITTEN BYJacquie ("Jax") Baker

Founder of Probably Genius, an AI visibility firm helping professional service brands become the named answer in AI search. Nearly two decades across technology, digital strategy and branding, including 1,000+ digital projects through her previous agency, now focused on making experts visible, verifiable and recommendable to AI. Let's talk →

Related reading

Read the method.
Then see it run on you.

The 109-point diagnostic is this library, applied to your business. Free, about an hour to present, no obligation.

See where you stand →

A real diagnostic, not a sales call in disguise.