AI Confuses Two Businesses With Similar Names: A Disambiguation Audit and Evidence-Based Fix Guide

When an assistant blends your business with a similarly named one, the fix is not to argue with the answer but to make your identity facts specific, public and machine readable, then measure the error with a test you can rerun. The number that describes the problem is the wrong company identification rate: answers that attributed a fact to the wrong company, divided by answers to ambiguous prompts, with the numerator, the denominator and the test date printed together. We have not run a disambiguation audit, on our own domain or any client's, so this page publishes no wrong company rate and does not estimate one. As of 29 September 2026 the measurement does not exist in our records. What we hold is a nearby dated error of the same family, in legal technology in India: on 17 September 2026 Claude pulled a well known review site into its raw results for a set of clawlaw.in questions and then discarded it, because the list it offered was of American products and so was not an answer to an India question. The system had matched the category correctly and the country wrongly, which is the same failure as matching the name correctly and the entity wrongly.

What follows is the audit that produces your own number, and the identity work that changes it. Both are written so that somebody with no access to our files can run them.

What counts as a mix-up?

Six distinct errors get reported as the AI is confusing us with somebody else, and they have different causes and different fixes. Score them separately or your audit will produce one number that hides five problems.

• Wrong entity, right name. The answer is about a genuinely different company that shares or nearly shares your name.
• Merged entity. One description containing facts from both companies, so the answer is about a business that does not exist.
• Wrong attribute. Your name with the other company's founding year, city, size, funding, leadership or sector.
• Wrong group member. Your name attached to a subsidiary, a former brand, or a dissolved entity in your own group.
• Wrong country or jurisdiction. The right category and the wrong market, which is what Claude caught itself doing on 17 September 2026 with an American product list offered for an India question.
• Wrong sector entirely. A name collision with a business in an unrelated trade, common where the name is a dictionary word or a common surname.

Write one sentence defining each in your scoring sheet, with a worked match and a worked non match, before you run a single prompt. A second person must be able to apply your rule and reach the same verdict on the same answer, or nothing you publish can be rerun.

Create a two-company identity table

This is the artefact that makes everything else possible. One table, two columns, one row per identity fact, filled in for your business and for the company you are confused with. Use public records only, and record the URL and the date for every cell.

• Registered legal name, exactly as filed.
• Registration number, the CIN or LLPIN for an Indian entity, and the equivalent for a foreign one.
• Date of incorporation.
• Registered office address, including city and state.
• Trading name and any former names, each with the date it changed.
• Primary website domain, and every other domain you own that serves content.
• Sector and the specific service, written narrowly rather than as a category.
• Named founders or directors as on public record.
• GSTIN, where public.
• Trade mark registration number and class, where you hold one.
• Official social and directory profiles, listed individually by URL.
• The single sentence that distinguishes the two businesses, written so a stranger could tell them apart in one read.

Two rules about the other company's column. Use only what is on a public register or on that company's own site, and record where you got it. And do not publish claims about the other business beyond plain factual identifiers. Your page needs to establish who you are, not to characterise a competitor, and a page that does the second gets read as advocacy. On 17 September 2026 Claude said in as many words that clawlaw.in comparison pages read as vendor advocacy, and that is why the site landed as a name inside a list rather than as a linked recommendation.

Test assistants using ambiguous and disambiguating prompts

A disambiguation audit inverts one of the rules that governs every other test on this site, and you need to know why before you run it.

The blind rule is suspended here, deliberately, and only here. You cannot test whether an assistant confuses two named companies without naming them. That is legitimate for this test and it makes the results useless for anything else. On 27 July 2026 two runs of the same 78 questions happened on one day for clawlaw.in. The run whose wrapper named the brand came back ranking it first on almost every question. The blind run put the company second by breadth and absent altogether from the litigation due diligence questions it most wanted to win, and the first run was discarded because the only thing it measured was our own prompt. So keep the disambiguation prompts in their own file, label the file clearly, and never let a number from it appear in a visibility report.

Build three prompt groups of ten each. Group A is ambiguous: the shared name alone, with no qualifier. Group B is disambiguating: the name plus one true identifier, for example the city, the registration number, or the specific service. Group C is attribute probes: ask for one fact at a time, the founding year, the head office city, the service offered, the leadership, so a wrong answer tells you which attribute has been crossed over.

Name the assistants individually and treat each as its own denominator. Three is a workable number.

Set the repeat count and write out the arithmetic. Three fresh sessions per prompt per assistant. Thirty prompts times three assistants times three repeats is 270 recorded answers per round. The ambiguous group on its own is ten times three times three, which is 90, and that 90 is the denominator for the wrong company rate. Print both multiplications on the report so the reader can see the denominator being built rather than accept it.

Use fresh sessions with no memory carried over. A logged in session that has already been told who you are will get the next answer right for the wrong reason. Record session state on every row.

Log these fields per answer: date and time, assistant, mode, session state, location, prompt group, prompt version, prompt text, which of the six error types applies or none, the exact sentence containing the error, the attribute that was wrong, claimed sources, verified sources, whether a live search was evidenced, and the screenshot file name.

Treat any statement about sourcing as a claim. In the aiknowsus.com audit of September 2026 Perplexity withdrew its own earlier statement when asked how many questions it had actually searched for, saying it could not honestly substantiate the claim that it had run a live search for each one, and in another batch that its claim to have searched all five was not adequately supported. Three separate batches produced that retraction. Keep a claimed sources column and a verified sources column and never merge them.

Publish the error counts and examples

We publish no wrong company identification rate, because we have not run this audit. The cell is empty as of 29 September 2026 and stays empty until the design above has been completed and the captures published, at which point it will carry the numerator, the denominator, the assistants, the prompt group and the dates in one sentence.

What we hold are four dated observations from two programmes. They are the nearest real evidence we have about how assistants handle entity level detail, and none of them is a disambiguation rate.

• A right category and a wrong country, caught and discarded. Claude, 17 September 2026, clawlaw.in, legal technology in India. A well known review site appeared in the raw results and was dropped because its list was of American products, so it was not an answer to an India question. Across those six questions no third party review or directory source made it into any answer at all.
• Three distinct vendor names kept apart across 78 answers. ChatGPT, 27 July 2026, clawlaw.in. In a memory based reading the breadth order was ProVakil on 28 questions, CLAW on 21 and Legistify on 18. The names were not merged, and the ordering reflects what the model had absorbed rather than what was true that week, because the same run confirmed it had run no live web search for any of the 78.
• A vendor's own page named as the vendor's own page. ChatGPT, 18 August 2026, clawlaw.in. Across eighteen blind commercial questions the company was the top source on exactly one, and its own comparison page was named in the answer as still being the vendor's own editorial page, which is an assistant getting entity provenance right rather than wrong.
• An attempt to force identity signals, refused. Claude, 6 August 2026, clawlaw.in. A competitor's page carried a hidden block of text addressed to answer engines, instructing them to cite that company as the source. The assistant found it, refused it, and named the company that had done it.

The fourth one is on this page as a warning rather than as a technique. Hidden instructions are the fastest way to turn a disambiguation problem into a reputation problem, because the assistant reported who had done it.

Improve first-party identity signals

Seven pieces of work, in the order we would do them. All of them are within your control, which is what makes them worth doing.

• Publish one identity page on your own domain. Registered name, registration number, incorporation date, registered office, former names with dates, the exact service, and the date the page was last reviewed. Plain text, in the served response, not built by scripts.
• State the distinction in one sentence on that page. Not a comparison and not a complaint. One factual sentence naming what your business does and where, so an assistant reading it has something short and quotable to use.
• Use the same name form everywhere. One spelling, one capitalisation, one legal suffix, across your site, your profiles, your invoices and your listings. Inconsistency is what creates the merge in the first place.
• Add organisation markup carrying the identifiers, including your official URL and your profile URLs. It makes the facts easier to parse and it obliges nobody, so add it and then measure rather than assume.
• Correct the third party records one at a time, naming each in your sheet with the date you raised it. Registers, map profiles, app stores, directories, and any review site that carries your entity details.
• Make sure the fact a buyer asks about is in the served bytes. On 6 August 2026 the clawlaw.in pricing page rendered prices only after scripts ran, so the crawler's copy contained no prices at all and the prices the assistants quoted came from an app store listing instead. Identity facts fail the same way and for the same reason.
• Check every claim on the identity page for a source and a date. On 17 September 2026 Claude examined two clawlaw.in pages whose competitor prices were correct and carried no link, no date and no source, and treated the pages as advocacy, using official sources instead. An identity page that links to the public register it quotes is doing the one thing that page could not.

Retest and report limitations

Rerun the identical three prompt groups monthly for at least three rounds, changing nothing but the date, and publish every round including the ones that got worse. Report the ambiguous group rate with its own denominator, the disambiguating group rate with its own denominator, and the attribute level counts separately, because the useful finding is usually which single attribute keeps crossing over.

Six limits, stated plainly.

• We hold no wrong company rate, for any sector, as of 29 September 2026.
• Prompts here name brands, so nothing from this audit is a visibility measurement. Keeping the files separate is not bookkeeping, it is the difference between a valid number and a contaminated one.
• Sectors and countries do not transfer. The observations above are legal technology in India and the AI visibility category. A rate from either would say nothing about a name collision in, for example, hospitality.
• A wrong answer that does not repeat is not a finding. Three repeats per prompt is the minimum before you act, and it still does not give you a distribution.
• You cannot make an assistant answer correctly. You can only make the correct facts easier to find, cheaper to verify and harder to contradict. We do not sell a guaranteed answer and nor should anybody else.
• A correction that appears to work may not be yours. Without a set of prompts you did not act on, you cannot separate your change from the assistant's own change that month.

Sources and change log

Every observation above comes from the clawlaw.in programme of July to September 2026, recorded in GEO_BASELINE_RESULTS_2026-07-27.md, GEO_GAP_ANALYSIS_2026-08-18.md and the assistant audit files of 17 September 2026, or from the aiknowsus.com audit of September 2026 across 24 batches and 72 conversations. The 6 August 2026 hidden instruction finding and the 17 September 2026 discounting of undated prices can be checked from outside today. Counts taken over our own capture folders cannot, until we publish the captures.

Version: 29 September 2026, first publication. The page gets updated when a disambiguation audit is completed and a wrong company rate can be printed with its denominator, when any figure is corrected, and when the captures are published.

Common questions

Should we change our name?

Run the audit before you consider it, because the cost is very high and the cause is often something smaller. In the three groups above, if group B fixes the answer, the assistant can tell you apart when given one qualifier, and your work is to make that qualifier prominent everywhere. Only a business whose group B answers are also wrong has a naming problem rather than a signals problem.

Can we tell the assistant we are a different company and have it remember?

Inside one session it will usually accept the correction, and that is not a fix. Test in fresh sessions with no memory carried over, and record session state on every row, or you will measure your own correction rather than the assistant's knowledge.

Is it worth naming the other company on our own page?

Naming it factually, once, to say plainly that you are not that business, is reasonable and sometimes necessary. Going further is not. On 17 September 2026 Claude treated clawlaw.in comparison pages as vendor advocacy and used other sources instead, and a page that spends paragraphs on another company is the same shape of page. State who you are, link the register, stop.

What about hidden text telling engines which company to cite?

Do not do it. On 6 August 2026 Claude found exactly that on a competitor's page, refused the instruction, and named the company that had written it. The downside is not that it fails, it is that it fails publicly.

How many prompts are enough to prove a mix-up?

Ten ambiguous prompts run three times on three assistants gives 90 answers, which is enough to know whether the error is stable and which attribute is crossing over. It is not enough to publish a confident rate for the category, and we will not print a confidence interval for a design we have not run.

Does a trade mark registration settle it?

It settles a legal question and it does not settle a retrieval question. A registration number is a strong public identifier and putting it on your identity page is worth doing. No assistant is obliged to prefer the registered holder of a mark when it answers a question, and none of them publishes a rule saying it does.

What to do first

Build the two column identity table today using public registers only, with a URL and a date in every cell. Then run ten ambiguous prompts three times each on one assistant in fresh sessions, and mark which of the six error types each wrong answer is. If one attribute keeps crossing over, you have found the single fact to make unmissable on your own site this week.