Ask ChatGPT about yourself and you’ll get one of three answers: something accurate, something outdated, or something that never happened. The third one is the reason people start paying attention to their AI reputation in the first place. A prospect runs a quick query before a sales call, and the model confidently attaches you to a company you left years ago, a claim you never made, or a controversy that belongs to someone with a similar name. Nobody wrote that. The model assembled it.
That’s the uncomfortable part. There’s no page to take down, no reviewer to appeal to, no editor to email. The Cory Maki AI reputation approach I use with clients starts from accepting that reality: you can’t delete a hallucination, you can only make the truth easier to find, easier to parse, and more frequently repeated than the error.
Why AI gets it wrong in the first place
Large language models are not databases. They don’t store a record of you and retrieve it on request. They predict the most statistically plausible next piece of text given everything they’ve absorbed. When the evidence about you is thin, contradictory, or scattered, the model fills the gap with what usually follows for people who look like you on paper. That’s a hallucination — not a lie, not malice, just confident pattern completion where facts should be.
In my work with clients, most bad AI answers trace back to one of five causes:
- Thin evidence. If there are only two or three places on the open web that describe what you do, the model has almost nothing to anchor on and will improvise the rest.
- Stale evidence. Old bios, outdated titles, abandoned profiles and archived press pages age badly. Training data is a snapshot; retrieval often favors pages with long-standing link equity, which skews old.
- Entity collision. Shared names are brutal. If someone with your name is more written-about than you, the model blends the two into a single composite person.
- Contradiction. When your LinkedIn, your site, a podcast bio and a conference page all say slightly different things, the model picks one — or averages them into something that’s true nowhere.
- Low-quality source dominance. Scraper sites, AI-spun directory pages and content farms republish errors at scale. Volume can outweigh accuracy in a retrieval system that’s optimizing for relevance, not truth.
Understanding the mechanism matters, because it tells you what the fix looks like. If a model is guessing, the cure isn’t outrage. It’s evidence. This is the same reasoning behind how LLMs form an opinion about you in the first place — impressions get built from repeated, corroborated signals, not from a single authoritative page you control.
Why stale beats false, and both beat nothing
Outdated information is the more common problem, and the one people underestimate. An invented award is obvious and rare. A three-year-old job title repeated across every AI assistant is subtle, plausible, and quietly expensive. It shapes how a prospect frames the first conversation. It shapes what a journalist assumes before they call. It shapes how a hiring committee reads your name.
What I’m seeing across AI search is that answer engines are increasingly grounded — Google AI Overviews, Perplexity and ChatGPT’s browsing modes pull live sources and cite them. That’s genuinely good news, because it means retrievable, current, well-structured content can beat a stale training-data impression. But grounding only helps if there’s something clean to ground in. If your best source of truth is a bio you last updated when you had a different role, the retrieval layer will faithfully surface the wrong thing.
Over a decade in reputation and search has taught me that the instinct to suppress is almost always weaker than the instinct to publish. Suppression is slow, adversarial and frequently impossible in an AI context. Publishing a better, clearer, more citable source of truth is something you can start this afternoon.
How the correction actually works
Here’s a concrete example of the pattern, generalized from client work. A founder finds that three assistants describe her as the CTO of a company she advised but never worked at. The origin is a single conference speaker page that mislabeled her, which then got syndicated to a dozen event aggregators and two scraper sites. Twelve sources now say the same wrong thing. From the model’s point of view, that’s consensus.
The fix isn’t to argue with the model. It’s to change the consensus:
- Correct the origin source first — the conference page — so the error stops propagating.
- Publish a canonical, plain-language bio on an owned domain that states the correct role, dates and affiliations in unambiguous sentences a model can lift verbatim.
- Update every high-authority profile that AI systems routinely retrieve, so they corroborate each other instead of contradicting.
- Earn one or two new mentions in places that get cited — editorial coverage, a podcast with a real show-notes page, a Q&A on a community platform — stating the correct facts in context.
- Re-check the answers over the following weeks, across multiple assistants, using varied phrasings.
That’s not a trick. It’s online reputation management adapted to a retrieval-driven world, and it’s the operating principle behind the ARC Method I built for earning AI citations: make the answer easy to find, easy to verify, and repeated often enough that it becomes the path of least resistance. You can go deeper on the ARC Method and AI-citation frameworks if you want the full structure.
What to do about your AI reputation this month
A practical sequence, in order:
- Audit before you assume. Ask four or five assistants the same set of questions: who are you, what do you do, what are you known for, what should someone know before working with you. Vary the phrasing. Save the outputs and the cited sources. That source list is your map.
- Classify each error. Hallucinated, outdated, or entity confusion. Each has a different fix. Hallucinations need new corroborating evidence. Outdated claims need source updates. Entity confusion needs disambiguation — consistent naming, location, role and organization stated together so the model can tell you apart.
- Build one canonical page. One URL, on a domain you own, that answers the obvious questions in direct declarative sentences. No throat-clearing. No metaphors. Structure it with clear headings and short paragraphs, because clarity and structure are what make content citable. Models quote what’s easy to quote.
- Fix the origin, not just the copies. Chasing syndicated duplicates without correcting the source is whack-a-mole.
- Show up where answers are formed. Community platforms, credible publications and Q&A surfaces are disproportionately represented in AI citations. A well-reasoned answer in a place models actually retrieve from does more than ten more pages on your own site. That’s a core part of why Reddit authority matters in AI search.
- Monitor on a schedule. Answers drift. Models update. Set a recurring check — monthly is reasonable for most people, weekly during an active correction. Systems and automation are how you keep quality consistent without it eating your calendar, which is the same logic behind AI visibility and fulfillment automation work generally.
What this process is not: fake reviews, manufactured praise, scraped citation networks or anything designed to launder a claim that isn’t true. Those tactics fail on their own terms in AI search, because retrieval systems increasingly weight source credibility — and they fail ethically regardless. Reputation is earned, not bought. If the accurate version of your story isn’t compelling, no amount of GEO fixes that. If it is compelling and simply hard to find, that’s a solvable problem.
The durable principle
Anyone working in AI reputation management for more than a few months learns the same lesson: you don’t control the output, you control the input. The way I think about this is that every assistant is running an open-book exam about you, and your only real move is to improve the book.
Models will keep changing. Retrieval will get better, hallucination rates will fall, and some of today’s specific failures will quietly resolve themselves. What won’t change is the underlying dynamic — citations beat rankings, corroboration beats assertion, and the clearest, most verifiable, most frequently referenced version of the truth wins by default. That’s true in classic SEO and search strategy and it’s even more true in Generative Engine Optimization.
One thing that consistently works: stop trying to win the argument with the machine, and start making the correct answer the easiest one to give.
Photo by Quinten de Graaf on Unsplash