Ask ChatGPT, Perplexity or Google AI Overviews about a person and you rarely get a research report. You get an impression — a confident paragraph assembled from whatever the model absorbed during training and whatever it can pull from the live web at the moment you ask. Sometimes that impression is accurate. Often it’s three years out of date, blended with someone who shares your name, or padded with a detail nobody can source. Figuring out where those impressions come from is the part of the work most people skip, and it’s where almost every Cory Maki AI reputation conversation with a founder or a law firm partner starts.
Understanding the mechanism matters more than any tactic, because the mechanism tells you which levers actually move and which ones are theater.
Why an AI’s impression of you now matters more than page one
For fifteen years, reputation work meant managing a results page: ten blue links, a knowledge panel, some review stars. You could see the battlefield. A user still made the final judgment by clicking.
AI search changed the shape of that judgment. When someone asks an assistant whether your firm is legitimate, or what your SaaS company does, or whether a founder has a history worth worrying about, the model collapses dozens of sources into a single answer — and most users never click through to verify it. The summary is the reputation for that user, in that moment.
That’s why AI reputation management isn’t just SEO with a new label. In traditional search you compete for position. In AI search you compete to be the source the model trusts enough to paraphrase. Citations beat rankings — a page that never cracks the top three but gets quoted in an AI Overview does more for your reputation than a number-two ranking nobody reads.
How the impression actually gets built
There are three layers to how a model forms and repeats a view of you, and they behave differently.
1. Training data: the slow, sticky layer
Models are trained on enormous crawls of public text — websites, news archives, forums, wikis, book and paper corpora. If your old job title, your former company, or a five-year-old press release appeared frequently enough across that corpus, it’s baked into the model’s baseline sense of who you are. This layer is slow to change and you cannot edit it directly. What you can do is change what the next crawl sees, and change what the model can retrieve right now.
2. Retrieval: the fast layer
Most modern assistants don’t rely on training data alone. They search, fetch a handful of pages, and ground the answer in what they just read. This is the layer you can influence in weeks rather than years. If a clear, current, well-structured page about you exists and is easy to parse, it has a real chance of being the thing the model reads before it answers.
3. Consensus and repetition: the layer nobody plans for
Here’s the part that surprises people. Language models are pattern machines, and agreement across independent sources reads as truth. One claim on your own website is an assertion. The same claim echoed on a profile page, a podcast show-notes page, a directory listing, a conference bio, and a forum thread becomes a fact the model will repeat without hedging.
This cuts both ways. A wrong detail that got syndicated across twelve aggregator sites in 2021 has the same structural advantage as a true one. That’s the mechanism behind most LLM hallucinations about real people: not invention from nothing, but confident amplification of a stale consensus.
A concrete example
A common pattern in my work with clients looks like this. A founder exits one company, starts another, and updates the new company’s site. Six months later an AI assistant still describes them by the old venture — because the old venture generated years of press, bio blurbs and profile pages, while the new one has a homepage and a LinkedIn update. The model isn’t wrong about its sources. It’s reading a corpus where the old story outweighs the new one by fifty to one.
Nothing gets fixed by complaining to the model. It gets fixed by rebalancing the evidence.
Managing your AI reputation without manipulating anyone
Over a decade in reputation and search — startups, law firms, public figures, and now as Head of Fulfillment at Reputation Pros — the ethical line has stayed simple for me: you can correct the record, you can make true information easier to find and easier to parse, and you cannot fabricate consensus. Fake reviews, sock-puppet forum accounts and spun bio pages don’t just fail the sniff test; they create exactly the kind of low-quality, contradictory signal that makes models hedge and hallucinate more.
Reputation is earned, not bought. What follows is how you earn it in a machine-readable way.
Build one canonical source of truth
Pick the page that should be the definitive answer about you or your company — usually an about page or a dedicated bio — and make it genuinely complete: current role, prior roles, location, what you actually do, what you’ve published, where you’ve been featured. Write it in plain declarative sentences. Models paraphrase clean prose far more reliably than they paraphrase marketing poetry.
Make it citable, not just correct
Clarity and structure are what make content citable. Practically, that means:
- Front-load the answer. The first sentence of each section should stand alone if lifted out of context.
- Use descriptive headings that mirror how people actually ask questions.
- Keep entity details together — name, role, organization, location in proximity, so the model doesn’t have to guess which facts belong to which person.
- Add structured data (Person, Organization, Article schema) so machines get an unambiguous version alongside the prose.
- Date things. “As of 2025” resolves conflicts between your page and an older one.
Distribute corroboration
One page can’t outvote a stale consensus by itself. The ARC Method I developed for earning AI citations exists for this reason: consistent, corroborated, retrievable evidence across independent surfaces. In practice that means keeping profiles, author bios, directory entries and interview show notes aligned on the same facts, and earning genuine mentions on sources that models already lean on. Community platforms matter more here than most people expect — Reddit threads are disproportionately influential in AI answers, which is why showing up honestly where the answers are formed beats broadcasting into a vacuum.
Monitor the answers, not just the rankings
You can’t manage what you don’t measure, and rank trackers won’t tell you what an assistant said about you this morning. Query the major assistants directly on a schedule with the questions your buyers actually ask — is this company legitimate, who runs it, what are the complaints — and log what comes back, including which URLs get cited. What I’m seeing across AI search is that citation sets shift constantly, so a monthly snapshot beats a one-time audit. Systems and automation scale quality here; a simple recurring check catches a drifting narrative long before a prospect does.
Correct at the source
When a wrong claim traces back to a specific article or listing, go there. Request a correction from the publisher, update the outdated profile, retire the page you control that still says the old thing. This is unglamorous online reputation management work, and it’s the highest-leverage thing you can do, because retrieval-based answers are only as good as the documents they retrieve. Pair it with the broader Generative Engine Optimization fundamentals and you’re improving both what the model finds today and what the next training crawl absorbs.
The durable principle
Language models don’t have opinions about you. They have a weighted average of what the internet has already said, filtered through whatever they can read in the half-second before answering. That’s not a threat and it’s not magic — it’s a supply chain, and supply chains can be improved.
The way I think about this: stop trying to persuade the machine and start improving the evidence it reads. Publish the clearest true account of who you are, make it structurally easy to quote, get it corroborated in places that independently matter, and check the answers regularly. Do that consistently and the impression corrects itself — not because you gamed anything, but because the record finally reflects the work.