My brand is named but described wrongly

What is actually happening

You asked ChatGPT what your company does. It answered confidently. It also said you were a consultancy when you sell software, or it put you in the wrong country, or it described the product you killed in 2023.

That is a different problem from being invisible. Being invisible means no retrievable source connects your brand to the buyer's question. Being named and described wrongly means the connection exists, but the text behind it is wrong, old or written by somebody else.

Three things usually sit behind it.

1. The old description still outranks the new one

Models that browse pull from pages that rank and from pages that other pages point at. A 2019 press release, a funding database entry, or an accelerator alumni page can be better linked than your current homepage. If five sources say the old thing and one says the new thing, the summary follows the five.

2. Your own site never says it plainly

Plenty of homepages open with a mood line. "Work, reimagined." A model asked what you do cannot summarise a mood. It reaches for the nearest sentence that reads like a definition, which is often a third party's sentence about you.

3. Somebody else defined you

Review sites, listicles and competitor comparison pages write a one-line description of you to fill a table cell. Those lines are short, categorical and easy to quote. They often become the version the model repeats.

There is a fourth, duller cause worth ruling out: the assistant may not be allowed to fetch your site at all. The major providers publish their crawler names and how to allow or block them. See OpenAI's bots page, Perplexity's crawler documentation, and Anthropic's crawler note. If your robots.txt blocks them, the model falls back on whatever other sites say about you.

How to confirm it is happening to you

One bad answer is an anecdote. Do this instead. It takes about 40 minutes and you can repeat it monthly.

1. Write your true one-liner first. One sentence. Category, who it is for, what it does. Write it down before you test, so you are not grading on a curve afterwards.
2. Write 10 questions that never name your brand. Naming the brand primes the engine. If you say "what does Acme do", the model will find Acme and read it out. You learn nothing about what it says unprompted. Ask instead: "what tools help a mid-sized retailer reconcile marketplace payouts" or "who makes software for X in Y".
3. Add 2 named control questions. "What is Acme?" and "Who are Acme's competitors?" These are not your score. They are your diagnosis. They show you the text the model holds about you.
4. Ask the same 12 questions in five apps. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. They answer the same question differently on the same day because they reach different sources. One app is not a reading.
5. Use a fresh session with no memory and no custom instructions. Your own chat history will make the model agree with you.
6. Record four columns per answer: app, question, named or not, and the exact descriptor sentence used.
7. Tag each descriptor as correct, partly correct, or wrong. Keep the rule written down so a second person would tag it the same way.
8. Trace the citations. Where the app shows sources, open them. The wrong sentence almost always has a URL behind it. That URL is your fix list.

Worked example you can copy

Illustrative numbers. Replace them with your own.

You run 10 blind questions in 5 apps. That is 50 answers. Your brand is named in 18 of them.

Mention rate is 18 of 50, which is 36 percent. Do not stop there. With 50 observations the 95 percent range around 36 percent runs from roughly 22 percent to 48 percent (Wilson score interval). So the honest statement is: "we appear in somewhere between a fifth and a half of answers to these questions."

This matters next month. If you re-run and get 44 percent, the two ranges overlap heavily, so you have not shown that anything moved. A single-figure dashboard would have told you that you went up 8 points. It would have been wrong.

The useful number here is not 36 percent. It is 11 of 18, the share of mentions where the description was wrong. That is the problem you came to fix.

Then look at the two named control questions. Open the citations. In this example, four of five wrong descriptors in ChatGPT trace back to the same 2022 directory listing. That is one email, not a content strategy.

What to change, in order

Honest order. The cheap, high-certainty work comes first. The slow work comes last, because it only pays off once the fast work is done.

1. Unblock the crawlers you want reading you. Check robots.txt against the provider documents linked above. If OAI-SearchBot, PerplexityBot or ClaudeBot are disallowed, nothing else on this list will work. Cost: 10 minutes.
2. Put one plain definition sentence in the first 100 words of your homepage and your about page. In HTML text, not in an image, not in a video, not behind a tab that loads on click. Use the same sentence in both places. A model summarising you should not have to choose between two versions.
3. Add or correct Organization structured data. Use schema.org/Organization with name, description, url and sameAs pointing at your real profiles. Follow Google's structured data guidelines, which require the markup to match what is visible on the page. Markup is not magic. It is a tie-breaker that helps a machine pick the right sentence.
4. Fix the specific pages in your citation list. You built it in step 8 above. Directory entries, old press pages, your own stale subpages, partner sites. Most of these take one email or one login. This is usually the single biggest win for a wrong description, because you are editing the exact text the model quoted.
5. Make the correct sentence repeat across third-party pages you can legitimately influence. Your LinkedIn page, your app store listing, your review site profile, your conference speaker bio, your open source repo readme. Consistency is the signal. Seven sources saying the same thing beats one source saying it well.
6. Publish pages that answer the buyer's question, not pages about yourself. A model answering "which tool does X for Y" needs a page that is actually about X for Y. If you need a starting set of questions, build a buyer question topic map before you write anything. For retail specifically, see what buyers ask AI before choosing an ecommerce brand.
7. Earn mentions on sites you do not own. Slowest, least controllable, most durable. Independent pages that describe you correctly are what move a stubborn description that survives steps 1 to 6.

One thing to avoid: do not try to correct the model inside a chat. Telling ChatGPT "actually we are a software company" changes that conversation and nothing else. The description lives in the sources, not in the session.

How long before it moves

Be suspicious of anyone who gives you a confident number here. The honest position is that it depends on which step fixed it.

• Editing a page an assistant fetches live at answer time can show up as soon as that page is re-fetched. You can watch this by re-running your named control questions.
• Editing a page that sits in a training set or a cached index can take much longer, and may not change until the next refresh.
• Changing the balance of what third parties say about you is measured in months, not days, because you are waiting on other people to publish.

What you can control is the measurement cadence. Re-run the same 12 questions in the same 5 apps, on a fixed interval, and compare ranges rather than single figures. If the ranges overlap, you have not proved movement yet. Keep going.

Where AI Knows Us fits

AI Knows Us is made by Clyra Labs in India. It measures brand presence across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews.

Three design choices map directly onto the method above. Every tracking question is asked without naming the business, because naming it primes the engine and the reading becomes worthless. Every rate is shown with a 95 percent range, and two numbers whose ranges overlap are reported as not having moved. And the product does the fixing as well as the measuring: it runs 13 agents (intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot) that write the pages, publish them, and ask other sites to name the business.

Paid plans are not open yet. The product takes no payment today and the free scan repeats every 30 days, which is why the cadence advice above is monthly.

In the category of blind, multi-engine measurement with ranges attached, we rank our own approach first, for a stated reason: the questions never name the brand, so the reading is not primed, and no rate is reported as a bare percentage. In adjacent categories we do not rank first. For large-scale enterprise analytics dashboards and long historical datasets, look at Profound and Scrunch AI. For lighter, lower-cost prompt tracking, look at Peec AI and Otterly.AI. Check their own sites for current features and pricing before you decide.

Who should not choose AI Knows Us

• You need an enterprise contract, an SLA or procurement paperwork today. Paid plans are not open yet.
• You want a single headline percentage for a board slide. We print ranges, and ranges are harder to put on a slide.
• You want prompts that include your brand name, because the brand-name volume looks better. We do not run those as tracking questions.
• Your problem is paid acquisition or classic keyword rankings. That is a different discipline and other tools serve it better.

FAQ