A founder types her own company name into ChatGPT and gets back a confident paragraph containing a former co-founder who left three years ago, a funding round that never happened, and a product description lifted from an outdated directory listing. Nothing in that answer links anywhere. Nothing in it is fixable with a press release. And the person who asked, a prospective partner doing quick due diligence, never sees her website at all.
That gap is the whole reason people go looking for Cory Maki AI reputation guidance in the first place. My name is Cory Maki, I’m an AI search strategist based in Taichung, Taiwan, and I spend most of my working hours on exactly this problem: what large language models say about a person or a brand, where those impressions come from, and how to correct them without resorting to anything manipulative.
What AI reputation management actually is
AI reputation management is the practice of shaping and correcting how generative AI systems — ChatGPT, Perplexity, Google AI Overviews, Claude, Gemini — describe you when someone asks. AI Overviews are the AI-written summaries Google now places above traditional results. An AI citation is when one of these systems pulls from, and often links to, a specific source while composing its answer.
The work breaks into four parts:
- Auditing what AI systems currently say about you across models and prompt variations.
- Correcting hallucinations, outdated claims and conflated identities at the source, not at the output.
- Building citable sources of truth that models can retrieve, parse and trust.
- Monitoring answers over time, because model updates and fresh crawls change things without warning.
Notice what isn’t on that list: fake reviews, review gating, sockpuppet accounts, or anything that tries to trick a model. Those tactics were bad practice in online reputation management and they’re worse now, because LLMs cross-reference. A suspiciously uniform wall of five-star reviews tends to produce hedged, skeptical summaries rather than glowing ones.
Why it’s now distinct from classic ORM
Over a decade in reputation and search taught me a specific playbook: suppress the bad result, build the good ones, earn the links, own page one. That playbook still matters. It is no longer sufficient, and the reasons are structural.
1. The unit of visibility changed
Classic ORM optimizes for positions. Ten blue links, and your job is to occupy as many favorable ones as possible. AI search optimizes for claims. A model doesn’t rank your bio page against a bad news article — it reads both, reconciles them, and produces one paragraph. You can rank first and still be described badly. In my work with clients, this is the single hardest idea to convey: in AI search, citations beat rankings.
2. Suppression doesn’t work the way it used to
Pushing a negative result from position three to position twelve used to solve most of the problem, because few people scrolled. A model has no page two. If a critical article exists anywhere in its retrievable index, it can surface inside the answer regardless of rank. The counter-move isn’t burial. It’s providing better-structured, more authoritative, more recent material on the same claim so the model has something clearer to reconcile against.
3. Errors persist and propagate
Traditional search has a self-correcting quality — wrong pages lose links and drift down. LLM hallucinations behave differently. An error gets repeated, quoted, syndicated to a content farm, and then re-ingested as apparent corroboration. I’ve seen a single stale directory entry become the model’s default description of a company for months. Correcting it means finding and fixing the upstream source, not arguing with the chatbot.
4. The surfaces that matter shifted
Models lean heavily on structured, consensus-bearing sources: Wikipedia and Wikidata, Crunchbase, LinkedIn, credible trade publications, documentation, and community discussion. Reddit in particular carries unusual weight in AI answers, which is why I wrote a book about the intersection of Reddit authority and AI search. If your reputation strategy never touches the places where answers are actually formed, you’re optimizing an empty room.
How LLMs form an impression of you
Simplified, three mechanisms are at play. Pretraining bakes in whatever was widely written about you before a model’s cutoff — slow to change, and the reason old job titles linger. Retrieval pulls live sources at query time; this is what AI Overviews and Perplexity lean on, and it’s the fastest lever you have. Synthesis is the model reconciling conflicting inputs into a single confident-sounding sentence, which is where most hallucinations are born.
A concrete example from the kind of work I do with law firms and public figures: a practitioner’s bio says one thing, an old association profile says another, a syndicated press mention garbles both, and the model splits the difference into a credential that was never real. Nobody lied. The information ecosystem was simply ambiguous, and the model resolved ambiguity by guessing. Remove the ambiguity — consistent naming, consistent titles, one canonical, well-structured bio that other sources point to — and the guess disappears.
Notes from the Cory Maki AI reputation playbook
The framework I built for this is the ARC Method — Authority, Relevance, Citability — and it’s the backbone of how I approach the ARC Method and AI-citation frameworks. Applied to reputation rather than pure Generative Engine Optimization (GEO), it looks like this:
- Authority — does a credible, independent source corroborate the claim? A model weighs a trade publication differently than your own about page. Earned coverage and verified profiles do real work here.
- Relevance — does your material actually answer the question being asked? People don’t prompt “tell me about [name].” They ask whether you’re legitimate, what you specialize in, whether anyone has complained. Your content should address those questions directly.
- Citability — can a model extract a clean, quotable statement without interpretation? Clarity and structure make content citable. Short declarative sentences, explicit dates, named entities, headings that match real questions, and schema markup all raise the odds of being the sentence the model reuses.
The way I think about this: you are not persuading an algorithm. You are making the truth easier to find and harder to misread than the alternatives.
What to do about managing your AI reputation this quarter
- Run a baseline audit. Ask five models the same eight questions a skeptical prospect would ask. Save the raw answers with the date. You cannot manage what you haven’t measured.
- Catalogue every factual error and trace it upstream. Most hallucinations have a findable source. Fix or update that source; where a model offers feedback or correction channels, use them.
- Publish one canonical fact page. Name, role, location, affiliations, publications, dates. Plain language, marked up with schema, linked from everywhere you control. This becomes the anchor other sources drift toward.
- Tighten consistency across LinkedIn, Crunchbase, author bios, speaker pages and directories. Inconsistency is what models resolve by inventing.
- Show up where the answers are formed. Answer real questions in the communities and publications your buyers actually consult. This is where SEO and search strategy and reputation stop being separate disciplines.
- Re-run the audit monthly. One thing that consistently works is treating monitoring as a recurring system rather than a one-off project — the same logic behind AI visibility and fulfillment automation for SaaS. Systems and automation scale quality; heroic manual effort doesn’t.
The durable principle
What I’m seeing across AI search is that the mechanics will keep moving. Retrieval methods change, citation behavior changes, new assistants arrive. What doesn’t change is the underlying trade: models reward sources that are clear, corroborated and consistent, and they punish ambiguity by filling it with invention.
Reputation is earned, not bought — that was true when the goal was page one, and it’s more true now that a single synthesized paragraph stands between you and a decision. If you want the long version of how I got here and what else I’ll be covering, start with the introduction to this site. Everything after it is built on the same premise: be the most citable version of the truth, and let the models do the rest.