Can a business control what ChatGPT says about it?

You cannot edit or delete ChatGPT's output, you cannot buy a placement that fixes a favorable answer, and you cannot force one canonical version that every user sees. The reason is mechanical, not a policy you can appeal. The model samples its wording with some built-in randomness, and answers can be personalized by per-user memory and context, so the same question returns different phrasing and sometimes different facts from one person or one request to the next. Even a "correct" answer is not a fixed thing you can lock; it is one draw from a distribution.

OpenAI does give you controls, but they answer a different question. Its account settings cover your own privacy, chat history, and memory, and its enterprise terms govern whether your data is used in training. None of those set what the model tells other people about your business. The platform that hosts the answer does not offer a dial for your brand's answer, which is exactly why the lever has to be the public sources the model reads, not a setting inside the product.

"What ChatGPT says about us is wrong" and "ChatGPT never mentions us" feel like one complaint, but they are two problems with two different fixes. Sorting which one you have is the first real step, because the work does not transfer between them.

Misrepresentation is when the model asserts something inaccurate, outdated, or unflattering about you. You fix it at the source it is reading: find the pages it pulls from, correct the wrong claim where you control it, and add credible sources that state the right facts so the accurate version is what most agreeing pages say. You are not deleting the bad answer; you are changing and out-weighting the inputs so the correct version becomes the likely one. When the fix involves publishing a clean, correct page the engine can lift from, that is answer engine optimization doing the work.

Omission is when the model simply does not name you, or names competitors instead. Nothing is wrong; you are just not in the answer. That is not a correction job at all - it is a citation-footprint problem, the same gap that decides whether any AI engine cites you, and it is covered in AI search visibility. Treating an omission like a factual error wastes effort on a page that was never inaccurate.

ChatGPT answers from two different places, and "can you change it" has a different answer depending on which one is speaking. The first is the trained base model: what it absorbed during training, months stale, with no live edit path. You cannot reach into that and rewrite a fact; it only refreshes on the next training cycle, which is months out and on no public schedule. The second is live web retrieval, where ChatGPT search reads current pages through a search index at answer time. That layer can change in days to weeks, because it reflects sources as they are now.

So the realistic lever is the live-retrieval layer and the public sources feeding it. Change what the trusted, current sources say and the retrieved answer can follow within days to weeks; the base-model layer lags by months. For the full duration picture and why it is two clocks rather than one date, see how long it takes to show up in AI search. The practical takeaway here is narrower: you are influencing inputs, not setting an output, and the speed you see depends on which source the engine used to answer.

If money or scale could buy a favorable answer, well-funded companies would already have one. In our Invisible 10 study we ran 600 model responses across the four largest AI engines (ChatGPT, Claude, Gemini, and Perplexity) against ten funded mid-market compliance vendors, and none of the ten was cited once. There was no placement to purchase and no budget that produced a mention. The answer is decided by what the engines read, which is why the work is on your sources and your footprint, not on a media buy or a request to the platform.

Because the output is a distribution and not a fact, a single check is meaningless on its own. Ask the same question twice and the wording shifts; ask it on two accounts and personalization can change the facts. One good answer is not proof you are fixed, and one bad answer is not proof you are broken - each is a single draw. The honest success measure is the share of answers that describe you correctly across many trials and several engines, watched over time, not one lucky or unlucky screenshot. That distinction, and the cadence that makes it reliable, is the subject of measuring AI citations.

You cannot control the answer, but you can see it clearly and move the odds. The first source-side fix is making sure the engines can read you at all; our AI crawler checklist is the concrete place to start. To know what the engines actually say about you today - and whether it is a misrepresentation or an omission - the Free Snapshot gives you a current read. For whether your changes are working, the SXO Audit runs 25 buyer prompts across six engines with three trials each and repeats over time, so you watch the share of correct answers trend rather than judge it from one check.

Try the Free Snapshot   See the SXO Audit

By the Web Cited Editorial Research Team. Last updated 1 June 2026.