A few years ago, a law professor discovered that ChatGPT had accused him of misconduct he never committed. Asked to name legal scholars involved in harassment cases, the model listed him, described a complaint that never existed, and cited a newspaper article that was never written. Every detail was fluent, specific, and false. He was left correcting a record that no human had ever actually created.
I open with that case because it is the professional’s version of the AI-search nightmare, and it happens in smaller, quieter ways far more often than the headline example suggests. An engine invents a bar admission you do not hold, or drops the one you do. It attributes a case result to your firm that belongs to someone else, or hands your win to a competitor. It assigns a founding year, a partner, or a specialty that is simply wrong. For a practice that runs on trust, a confident fabrication is not a technical curiosity. It is a reputation problem a prospect may act on before you ever hear about it.
Why AI invents facts about a firm
The mechanism is worth understanding, because it tells you what actually fixes it. A language model does not look up facts; it predicts plausible text.
When it has strong, consistent information about a firm, the plausible completion and the truth are more likely to line up. When that information is thin, contradictory, or ambiguous, the model fills the gap with something that fits the pattern of similar firms and may state it with the same confidence as a fact it genuinely knows.
Three conditions make a firm especially exposed:
- Entity ambiguity: A shared firm name or common surname can leave the model unsure which “Anderson” or “Delta” you are, so it may borrow attributes from another entity with a stronger signal.
- Data gaps: Limited third-party coverage, no knowledge-graph footprint, or sparse structured markup gives the model too little reliable information to work from.
- Contradictory information: A rebrand, old office address, or outdated profile can leave conflicting signals that the model cannot reliably resolve.
A Stanford study of legal AI tools found hallucinations in roughly one in six queries, even in a domain with structured source material.
It also helps to separate two different failures that often get grouped under “hallucination”:
1. The model does not have reliable facts
It fills the gap using patterns from similar firms.
The fix: Publish clear, authoritative facts and structure them so the model can identify and retrieve them.
2. The facts exist, but the model assigns them incorrectly
It may attach a competitor’s result, credential, or attribute to your firm.
The fix: Use clearer content and more explicit markup that binds the right facts to the right entity.
Most firms face some combination of both. That is why the fixes work best together rather than as a single switch.
The firms most likely to be misrepresented
Risk scales inversely with how much the internet already says about you. A national brand with thousands of pages, press mentions, and a Wikipedia entry has its identity anchored from every direction; a model has little room to invent. A boutique or solo practice may have a few dozen pages of meaningful third-party content, and when a model meets a question about a low-data firm, the ratio of guesswork to genuine recall shifts sharply toward guesswork.
The result is a specific and unfair pattern. A lesser-known family law firm can inherit the described specialties of a larger competitor simply because the model associates those specialties with the category and has more data on the bigger name.
The firms that most need to be represented accurately, the smaller practices competing on reputation, are the ones the machine is most likely to get wrong. Closing that gap is much of the work we do at meanderingSapien, and it starts with giving the model more of your own facts to hold.
What an uncorrected fabrication actually costs
A hallucination rarely stays put. It can enter a loop:
- A model states something false.
- A person or aggregator republishes it.
- The claim gets indexed.
- A later system finds the republished version and treats it as corroboration.
Over time, the fabrication starts to look like a confirmed fact. Academic work on AI-generated misinformation shows that false claims can move through the content ecosystem within days.
For a professional firm, the damage often appears in the queries that matter most:
- Comparison and “who is best” queries: A false claim about your specialties or results can influence whether you make a shortlist.
- Credential and compliance queries: A wrong admission, qualification, or fabricated disciplinary note can cause direct reputational harm.
- Background checks: A prospect, referral source, or journalist may take an AI-generated summary at face value without verifying it.
Every day a false claim remains uncorrected gives it more time to spread and harden. That is why I treat monitoring as an ongoing process rather than a one-time check.
How to give the machine facts it cannot get wrong
The fixes that work share one logic: reduce the model’s need to guess by giving it high-confidence, machine-readable facts. Four moves do most of the work.
1. Publish clear, answer-ready facts
State the facts about your firm plainly, including:
- Your category and practice areas
- Your locations
- Your admissions
- Your leadership
- Your founding year
- Services you do not offer
That last point matters. Explicit “negative clarity” closes the gaps where a model might otherwise infer services or capabilities that are not yours.
2. Make those facts machine-readable
Use schema markup, particularly Organization and Person types, to connect the right attributes to the right entity.
Plain-language content gives the model facts to recall. Structured markup reduces the chance that it misreads or misattributes them.
3. Anchor your identity in the knowledge graph
A well-formed Wikidata entry can give your firm a clearer, more defined identity across the web.
This is especially useful when your firm shares a name with another business or uses a common surname, because it helps distinguish which entity a fact belongs to.
4. Monitor and correct regularly
Run the same questions about your firm across ChatGPT, Perplexity, and Google’s AI every few weeks. Track:
- What each platform says about you
- Which competitors appear instead
- Which sources the answers rely on
- Any false claims that recur
When an error keeps appearing, publish a clear, extractable correction. That ongoing cycle of measure, find, and correct is the audit-and-maintenance work behind being recommended by AI.
What does not work
Two responses can feel productive without fixing the underlying problem.
1. Adding an AI disclaimer
A note saying “AI can make mistakes” on your site does not change what a model says about your firm or what gets copied elsewhere.
Once a false claim is stated confidently and republished, a disclaimer on your own site does little to stop it from spreading.
2. Waiting for the model to correct itself
Platform feedback tools can help, but corrections may not persist across model updates. Relying on the next version to get your firm right leaves your reputation dependent on a training cycle you do not control.
The stronger approach is to improve the source material AI systems learn from and retrieve, so the accurate version of your firm becomes the easiest one to surface.
A standing problem, not a one-time cleanup
The uncomfortable truth is that this never fully ends. Your firm’s representation across AI platforms drifts as content ages, competitors publish, and models retrain, so a fabrication corrected today can resurface in a slightly different form next quarter. The firms that stay accurately represented are not the ones that fixed it once. They are the ones that keep feeding the machines good facts and checking what comes back.
So I tell clients to treat this the way they treat their reputation offline: as something maintained, not installed. Give the machines a clear, corroborated, machine-readable account of who you are, then keep watching what they say and correcting what they get wrong.
In a market where your next client may meet your firm through an AI summary before they ever reach your site, the accuracy of that summary is worth guarding. If you want to see how AI currently represents your firm and where those gaps are coming from, book a call with meanderingSapien.
Frequently Asked Questions
What is the difference between an AI hallucination and outdated information about my firm?
Outdated information was once true, such as an old office address or former partner. A hallucination was never true. Outdated facts need current source content; hallucinations need stronger identity signals that reduce guessing.
Can I get ChatGPT or Google to remove a false claim about my practice?
Sometimes, but not reliably. Feedback tools may help, but corrections can disappear after model updates. Strengthening accurate source content is the more durable fix.
Why does AI get facts wrong about smaller firms more often?
Smaller firms often have less third-party coverage and fewer strong identity signals, giving models more room to fill gaps using patterns from better-documented firms.
Is a fabricated fact about my firm a compliance or liability problem?
It can be. False claims about credentials, disciplinary history, admissions, or results can create reputational or regulatory risk and should be corrected quickly.
Does schema markup stop hallucinations on its own?
No, schema helps bind the right facts to the right entity, but it works best alongside clear content and stronger knowledge-graph signals.
How often should I check what AI says about my firm?
Every few weeks is a sensible cadence. Run the same questions across major AI platforms and track recurring errors before they spread.
Raja Sampathi
Strategic consultant with 19+ years of enterprise experience, now focused on one problem: getting ChatGPT, Perplexity and Google AI to recommend professional-services firms by name. Every post draws on actual client work and live visibility audits not theory.
