How to Control What ChatGPT Says About Your Brand

Turn browsing on and ask a model "what is [your brand]?" and it will fetch a handful of live pages, then stitch them into a confident paragraph. When that paragraph is wrong, stale, or still selling the product you pivoted away from two years ago, the first instinct is to hunt for the button that fixes it. There isn't one. Knowing how to control what ChatGPT says about your brand starts with dropping the idea of editing the model and getting exact about editing what the model can retrieve.

I run this check constantly on sites built with AI, and the pattern holds: the answer is almost never invented from nothing. It is assembled from pages you can find, and often from pages you own.

Start here: there is no edit button

There is no edit button because the public model is shared, and the only copy of ChatGPT you can change is your own. Custom instructions and memory steer the answers inside your account, and nobody else's session sees them. So the fix people reach for first, opening ChatGPT and correcting it in the chat, changes nothing for the buyer across town.

Two different answers come out of a model, and they fail for different reasons. A reply drawn from memory alone reflects training data, which runs years behind and updates only when the model is retrained. A reply with browsing on reflects what the model fetches at that moment. The browsing answer is the one you can move inside a month, because you control part of what it fetches.

A retrievable source is any page a browsing model can open and quote when it answers a question about you: your website, a G2 or Trustpilot profile, a Crunchbase entry, a Wikipedia article, a Reddit thread. OpenAI does not publish a form where a brand submits a correction, so the correction has to reach the model the long way, through those sources. (If that ever changes, it will be the exception. Today the sources are the lever.)

The Mirror Loop: how to actually move the answer

The Mirror Loop is a five-step cycle: capture the answer word for word, diff it against your own site, trace each wrong claim to its source, fix that source, then ask again and compare. You repeat it until the answer reads the way the truth reads.

An AI mirror is a check that asks a language model what your product is, saves the reply verbatim, and holds it against your live site so you can see exactly where the two disagree. Run that by hand or let a tool keep the history. The steps are the same.

1. Ask the buyer's question and save the answer. Not "is my brand great." Ask "what is [brand]?", "is [brand] reputable?", and "what are the best tools for [your category]?" Copy each answer into a doc with the date. That saved text is your baseline.
2. Diff it against your site. Read the answer next to your homepage and about page. Mark every claim that is wrong, out of date, or missing. Wrong category. Old pricing. A feature you killed. A founder who left.
3. Trace each wrong claim to a source. Where could the model have read that? A stale category usually traces to your own served HTML. A reputation line traces to reviews or a Reddit thread. A funding or founded-date error traces to Crunchbase or a reference page.
4. Fix the source, starting with the one you own. Your served HTML is the fastest to change and the most authoritative page about you, so it comes first. Then the third-party profiles, in order of how much the model seems to quote them.
5. Ask again and compare. Same prompts, a week or two later, once pages are recrawled. Put the new answer beside the baseline. If a claim moved, you found its source. If it didn't, you fixed the wrong page, so go back to step three.

Tabkeel, the exam I build, runs this loop as a feature: it asks a model what your site is, keeps every answer with a timestamp, and flags where the reply and your page part ways, so the before and after sit side by side instead of living in your memory. You can run the loop with nothing but a notepad, though. The method is what matters.

Where ChatGPT gets what it says about you

ChatGPT builds a brand answer from a small set of sources, and knowing which one feeds which claim tells you where to aim. Across 30 million citations tracked from August 2024 to June 2025, Wikipedia was the single most-cited source in ChatGPT's answers (Profound). A Similarweb count for early 2026 put Wikipedia at 13.15% and Reddit at 11.97% of ChatGPT's US citations, the two biggest domains it pulls from.

Map the wrong claim to the page that most likely feeds it before you touch anything.

The top two rows are yours to fix today. The rest wait on other people's pages changing, which is why founders who only edit their own site still watch the reputation lines sit stuck.

A wrong answer, and the source behind it

Here is the shape it takes, from our own mirror history. Early on, a browsing model described Tabkeel as "an SEO audit tool," which is true enough to pass and wrong enough to lose the buyer who needs the AI-readability and billing-integrity checks that make it different. Nobody wrote that sentence. The model read our served page, found a loose headline, and filed us on the nearest shelf.

The source was the served HTML: the first plain sentence a crawler met described the category loosely. We rewrote the served copy to state the category in one literal line, high on the page, before any JavaScript ran. On the next pass the mirror answer named the exam and the fronts. One source, traced and fixed, and the answer followed. That is the loop working on the page you have the most control over, the one covered in making your site readable to AI.

What trips founders up

Most failed attempts to change an AI answer repeat the same few moves. Skip these.

• Pasting the identical marketing sentence on twenty pages. Duplication reads as manufactured, and repeating one claim everywhere can push it into low-value territory rather than make it stick. State the core facts consistently, in your own words each place, not as copy-paste boilerplate.
• Correcting the model in your own chat and assuming it took. It changed your session. Everyone else still gets the old answer.
• Fixating on one bad Reddit thread you can't delete. You rarely remove it. You outweigh it, with more and better signal on the sources the model quotes next to it.
• Leaving the homepage rendered only in JavaScript. Most AI crawlers receive an empty body, so the one page that is unambiguously about you tells the model nothing, and it falls back to guessing from third parties.

Your first pass this week

Run the loop once, end to end, on the single question that matters most to your pipeline. The category question, usually.

1. Open a browsing model and ask what your product is. Save the answer with today's date.
2. Mark every claim in it that is wrong or stale.
3. For the top one, open your own page and check whether a crawler, not a browser, would even read the right sentence. View source, not the rendered view.
4. Rewrite that served sentence to state plainly what you are and who it's for.
5. Ask the same question again in a week and compare.

If you'd rather not keep the baselines in a doc, point the exam at your URL and the AI mirror records each answer, diffs it against your site, and keeps the run history for you, alongside the read check that tells you whether a crawler received your category at all. Start at the Tabkeel exam, or isolate just the read with the AI readability tool. For the wider question of being cited rather than only correct, the gate-by-gate version is in how to get cited by ChatGPT and AI search, and the recall side, why a model may not know you by name at all, is in why ChatGPT doesn't know your startup.

Frequently asked questions