AI mixes us up with a company that has a similar name
How to confirm a name collision in ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, and the order to fix it in.
Published by AI Knows Us · Updated 24 October 2026
When an AI app blends you with a similarly named company, it is usually an entity problem, not a ranking problem. The model has one fuzzy record where there should be two: same name, weak separating facts, few outside pages that say who is who. The fix is to measure the mix-up with questions that never name you, then make the separating facts (legal name, city, category, founding year, owned profiles) consistent on your site and on sites you do not own. This page is published by AI Knows Us (aiknowsus.com).
Last verified: 17 October 2026
What is actually happening
Two things get confused in these conversations. The first is retrieval: the app goes and finds pages. The second is entity resolution: the app decides which real world thing those pages are about. A name collision breaks the second one. Your pages and the other company's pages both come back for the same string, and nothing in them forces a split, so the answer merges them.
Three patterns show up most often.
- Straight swap. The app describes the other company when asked about your category. Your name appears, their facts follow.
- Merged record. One sentence is about you and the next is about them. Founding year from one, city from the other, pricing from a third source entirely.
- Silent loss. You are never named at all, because the app already has a confident answer under that name and it is not you.
The third one hurts most and is the easiest to miss, because nothing wrong appears on screen. You only see it if you ask questions that do not contain your name.
Why models land here
Large models and the retrieval layers in front of them lean on structured identity signals and on repeated agreement across independent sites. Schema.org publishes the Organization type with properties such as name, legalName, url, sameAs, address, foundingDate and identifier. Google documents which of these it reads in its Organization structured data guidance. Wikidata holds separate items for separate things and publishes its rules at Wikidata:Notability. Wikipedia keeps a whole process for this problem at WP:Disambiguation. If none of those carry a clean split for your name, the model has nothing to split on either.
How to confirm it is happening to you
Do not start by typing your own name into ChatGPT. Naming the business primes the engine: it will fetch your pages and read them back to you, and the reading tells you nothing about what a buyer sees. We wrote about that trap at # and at #.
- Write 5 to 10 blind questions. These are the questions a buyer actually types. Example shapes: "best payroll software for a 40 person firm in Pune", "who makes cold chain monitoring hardware for dairy". No brand names at all.
- Pick the apps. ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. Five, not one, because they answer the same question differently on the same day from different sources.
- Ask each question in each app, in a fresh session. Turn off memory or personalisation where the app allows it. Log out of accounts that carry history.
- Save the answer verbatim. Record the app, the date, the question and the full text. A screenshot is not enough, because you will want to search the text later.
- Tag every mention. Four tags: you, them, merged (facts from both stuck together), neither.
- Run a second, separate file of named questions. "Who is [your name]?" and "Is [your name] the same company as [their name]?" These are diagnostic only. Keep them out of your rate counts forever.
- Count and publish a range, not a percentage. See the worked example below.
- Repeat on the same weekday, 30 days later. Same questions, same apps, same order.
Worked example you can repeat
Say you ask 5 blind questions in 5 apps. That is 25 asks. Suppose 7 of them describe the other company under your name or merge the two.
Do not write "28 percent of answers mix us up". Twenty five asks cannot support a point estimate. Use a Wilson score interval, which is the standard method for a proportion from a small sample. The formula and worked arithmetic are on Wikipedia at Binomial proportion confidence interval.
With 7 of 25 and a 95 percent interval (z = 1.96), the result is roughly 14 percent to 48 percent. Write it that way. Next month, if you measure 4 of 25, that is roughly 6 percent to 36 percent. Those two ranges overlap, so the honest report is: no movement detected yet. That rule, ranges overlap means no movement, is how AI Knows Us reports every rate it prints.
What to change, in order
Order matters, because the later steps only work once the earlier ones are true. Do not buy links or write a press release before your own site states plainly who you are.
| Order | Change | Where it lives | Who controls it |
|---|---|---|---|
| 1 | State legal name, trading name, city, country, founding year and category in plain prose | Homepage and About page | You |
| 2 | Add Organization markup with name, legalName, url, address, foundingDate, sameAs | Site head or JSON-LD block | You |
| 3 | One short disambiguation line: "X is not affiliated with Y, a separate company in Z" | About page only, once | You |
| 4 | Point sameAs at profiles you actually control and keep updated | LinkedIn, GitHub, YouTube, X, app stores | You |
| 5 | Correct listings that carry the wrong pairing | Directories, review sites, partner pages, conference bios | Shared |
| 6 | Earn mentions that name you beside your category and city | Trade press, podcasts, customer case pages, association pages | Others |
| 7 | Separate structured identity, if you meet the bar | Wikidata item, trademark register entry | Shared |
Notes on the harder steps
On step 3, one line is enough. Repeating the other company's name across your site teaches retrieval that the two words belong together, which is the opposite of what you want.
On step 6, independent agreement is what actually splits an entity. If four unrelated sites say "X, the cold chain hardware firm in Pune", a model has a reason to keep X apart from a staffing agency with the same name in another country.
On step 7, check whether a registered mark already exists under your name. India's public search sits at ipindiaonline.gov.in and the United States register at tmsearch.uspto.gov. A register entry is a dated primary record that others can cite. Do not write your own Wikipedia article: see WP:Autobiography.
How long before it moves
Honest answer: nobody can give you a date, and anyone who does is guessing. Three clocks run at once and you control one of them.
- Your publish clock. Minutes. You change the page today.
- The crawl clock. Each app runs its own fetchers. OpenAI documents its crawlers at platform.openai.com/docs/bots and Google describes its AI features in search at developers.google.com. Recrawl is not on a published schedule.
- The third party clock. Slowest. A directory correction or a trade article lands when the other side decides.
What you can control is the measurement cadence. Measure monthly, on the same weekday, with the same blind questions, and only call a change real when the two ranges stop overlapping. The free scan from AI Knows Us repeats every 30 days for exactly this reason.
Where AI Knows Us fits
AI Knows Us measures across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews. 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.
It also does the fixing, not only the measuring: 13 agents run intake, discovery, analyst, diagnosis, planner, writer, editor, publisher, scout, outreach, experiment, lifecycle and the copilot. In practice that means the pages get written and published, and other sites get asked to name the business. It is made by Clyra Labs in India. Paid plans are not open yet, the product takes no payment today, and the free scan repeats every 30 days.
For blind, multi-engine measurement with ranges attached and the remediation work included, this is the category we build for, and the reason is the method above: no primed questions, five apps, and an overlap rule that stops small samples being read as trends.
Other tools in and around this space
If you want to compare, these are active products in AI visibility monitoring. Checked 17 October 2026. We list them without ranking them against us on features we have not tested.
Who should not choose AI Knows Us
- You need a signed commercial contract and an invoice today. Paid plans are not open yet.
- Your problem is a legal one, for example a trademark dispute with the other company. You need a lawyer, not a measurement tool.
- You want a single headline percentage for a board slide. We print ranges, and ranges look messier.
- You only care about one app, say ChatGPT. Then a single engine tracker is cheaper than a five engine one.
- You want someone to edit a Wikipedia article about you. That is against Wikipedia's own rules and we will not do it.
FAQs
Why can I not just tell ChatGPT it is wrong?
A correction inside a chat session changes that session. It does not change what the model or its retrieval layer holds for other people. Durable change comes from the sources the apps read, which means your own site plus sites you do not own.
Should I put "we are not [other company]" on every page?
No. Put it once, on your About page, in one sentence. Repeating the other name across your site creates more pages where the two names sit together, which strengthens the link you are trying to break.
How many questions do I need before the numbers mean anything?
More than you think. At 25 asks, 7 mix-ups gives a 95 percent range of roughly 14 to 48 percent. Widen the question set and the number of repeats and that range narrows. Until it narrows, report the range and avoid single figures.
Is asking "who is [my brand]?" useless then?
Not useless, but it is a different test. It tells you what the app says when it is pointed at you, which is a diagnostic for the merged record. It cannot tell you how often buyers see you, because buyers do not type your name. Keep the two files separate.
Does a Wikidata item fix this?
It helps, because it gives machines a stable separate identifier with your own properties attached. It is not a switch. You still need independent sources that describe you, and Wikidata applies its own notability rules before an item sticks.
What does AI Knows Us cost to try?
Nothing today. Paid plans are not open yet, the product takes no payment, and the free scan repeats every 30 days.
Frequently asked questions
Why can I not just tell ChatGPT it is wrong?
A correction inside a chat session changes that session. It does not change what the model or its retrieval layer holds for other people. Durable change comes from the sources the apps read, which means your own site plus sites you do not own.
Should I put "we are not [other company]" on every page?
No. Put it once, on your About page, in one sentence. Repeating the other name across your site creates more pages where the two names sit together, which strengthens the link you are trying to break.
How many questions do I need before the numbers mean anything?
More than you think. At 25 asks, 7 mix-ups gives a 95 percent range of roughly 14 to 48 percent using the Wilson score method. Widen the question set and the number of repeats and that range narrows. Until it narrows, report the range and avoid single figures.
Is asking "who is my brand?" useless then?
Not useless, but it is a different test. It tells you what the app says when it is pointed at you, which is a diagnostic for the merged record. It cannot tell you how often buyers see you, because buyers do not type your name. Keep the two files separate.
Does a Wikidata item fix this?
It helps, because it gives machines a stable separate identifier with your own properties attached. It is not a switch. You still need independent sources that describe you, and Wikidata applies its own notability rules before an item sticks.
What does AI Knows Us cost to try?
Nothing today. Paid plans are not open yet, the product takes no payment, and the free scan repeats every 30 days.