Someone types your name into ChatGPT or Perplexity and gets a confident paragraph back. Two sentences are right. One is three years out of date. And one is simply invented — a company you never worked for, a claim you never made, a controversy that belongs to someone with a similar name. There’s no complaint button, no support ticket, no obvious human to email. That gap between what’s true and what the model repeats is the heart of what I think of as Cory Maki AI reputation work: not arguing with a chatbot, but repairing the record the chatbot is reading from.
The good news is that this is fixable more often than people expect. The bad news is that it’s fixable slowly, through source correction and patience, not through a trick or a takedown. Over a decade in reputation and search has taught me that the mechanics of a fix are boring, repeatable, and mostly about evidence.
Why AI gets people wrong in the first place
Large language models don’t store a file about you. They generate the most statistically plausible answer given their training data, plus whatever they retrieve live from the web at the moment of the question. That produces three distinct failure modes, and the fix is different for each:
- Hallucination. The model invents a detail to complete a pattern — a plausible-sounding job title, a made-up book, a fabricated affiliation. There’s no source to correct because there was never a source.
- Stale truth. The information was accurate in 2021 and nothing newer or more authoritative has replaced it. This is the most common problem I see, and the most solvable.
- Entity confusion. You’ve been merged with someone who shares your name, your company name, or your industry. The model isn’t wrong so much as it’s collapsing two people into one.
Before you do anything, diagnose which one you’re facing. Ask the same question across ChatGPT, Perplexity, Google AI Overviews (the AI-generated summaries at the top of search results), and Gemini. Ask it a few different ways: who is this person, what are they known for, what controversies exist, who do they work for. Then ask the follow-up that matters most — where did you get that? Perplexity and AI Overviews will show you citations. Those citations are your worklist. I’ve written more about the diagnostic side in when AI gets your reputation wrong, and about the underlying mechanics in how LLMs form an opinion about you.
Correct the sources, not the chatbot
The single biggest mistake I see is people spending an afternoon arguing with a model in a chat window. You can correct it inside a conversation and it will politely agree with you — and then produce the same wrong answer for the next stranger who asks. The conversation isn’t the record. The web is.
The way I think about this: an AI answer is a summary of the sources the system trusts most. If you want the summary to change, change what’s available to summarize. That’s the entire premise of AI reputation management as a discipline — it’s closer to publishing and evidence-building than to crisis PR.
A concrete example of how a bad claim survives
Here’s a pattern I’ve seen repeatedly in client work. A founder leaves a company. A press release from the launch year, a conference bio page, and a syndicated news roundup all describe them as CEO of that company. The founder updates LinkedIn and their personal site — two pages — and considers it handled.
Two years later, an AI assistant still says they’re CEO. Why? Because the three stale pages are on domains with more authority and more inbound links than the two corrected ones, they’re mirrored across syndication networks, and none of the corrected pages state the change explicitly. The personal site says Founder, New Company. It never says previously CEO of Old Company through 2023. The model has no bridge between the two facts, so it defaults to the version with more corroboration.
The fix wasn’t a takedown. It was publishing a clear, dated, factual statement of the current role on an owned page, getting the conference bio updated, and earning two or three fresh mentions that stated the current role in plain language. Within a few months, the AI answers followed. Not instantly, and not because anyone gamed anything — because the weight of evidence shifted.
Seven steps for managing your AI reputation when the answer is wrong
- Document the error. Screenshot the answer, note the platform, the date, the exact prompt, and any citations shown. You need a baseline to measure against later.
- Trace the citations. Open every source the model names. Often you’ll find one origin page and a dozen copies. Fix the origin first.
- Correct at the source where you can. Contact the publisher with a factual correction request. Editors update things when you’re specific, polite, and provide evidence. This is ordinary online reputation management — nothing exotic.
- Publish the canonical version yourself. One owned page that states the facts plainly: current role, past roles with dates, what you do, what you don’t do. Structured, scannable, unambiguous. Clarity and structure are what make a page citable.
- Disambiguate deliberately. If you’re being confused with someone else, say so in your own content — name your city, your field, your company, your book. Schema markup, consistent naming, and linked profiles all help systems tell you apart.
- Earn fresh corroboration. One page rarely outweighs five. Interviews, guest articles, podcast appearances, community answers — these are the mentions that shift the balance. This is where the ARC Method and AI-citation frameworks come in: the goal is to be the source worth citing, not to shout louder.
- Use official feedback channels. Most AI products have a thumbs-down or report mechanism, and Google has feedback options on AI Overviews. It’s slow and unglamorous, but for outright hallucinations and defamatory output, it’s a legitimate path.
What not to do
There’s a growing market for shortcuts here, and most of it will make things worse. Don’t buy reviews. Don’t spin up fake profiles or sockpuppet accounts to seed a narrative. Don’t mass-publish thin, near-duplicate pages hoping volume wins — AI systems are increasingly good at spotting low-quality repetition, and a pattern of manufactured content becomes its own reputation problem. And be careful with aggressive removal campaigns against legitimate journalism; the attempt often generates more coverage than the original item.
Reputation is earned, not bought. That’s not a moral posture, it’s a practical one — earned sources hold up over time, and manufactured ones tend to collapse in a way that’s very hard to clean up afterward.
Monitoring: how you know the fix took
Corrections don’t announce themselves. Retrieval-based answers can update within days; answers baked into model weights wait for the next training cycle. So you check on a schedule.
One thing that consistently works is a simple recurring audit: the same five to ten prompts, run across the same set of assistants, logged monthly with the citations each one surfaces. You’re watching two things — whether the wrong claim persists, and whether your corrected sources start appearing in the citation list. Citations beat rankings in AI search; when your own pages show up as sources, the answer usually follows. If you’re running this for a team or product at scale, it’s worth automating, which is the approach I take to AI visibility and fulfillment automation.
What I’m seeing across AI search is that the people who get corrections to stick are the ones who treat this as maintenance rather than an emergency. They keep an accurate, current, well-structured owned page. They show up where answers are formed — the forums, publications and communities their industry actually reads. They check quarterly instead of panicking annually.
The durable principle
You can’t edit an AI’s mind. You can only change the evidence it reads. Every correction that lasts works the same way: find the source, fix or outweigh it, publish something clearer and more citable than what was there before, then verify. It’s slower than people want and more reliable than anything that promises otherwise.
The version of you that AI describes is a reflection of the record you’ve left behind. Make that record accurate, structured and current, and the reflection tends to take care of itself.