AI Says Your Business Is Closed: A Source-by-Source Status Check and Correction Playbook

If an assistant says your business has closed and it has not, treat it as a sourcing fault rather than an opinion, and work it in one order: confirm you are open in your own records, find the exact answer and the exact source that says otherwise, check the public registers a careful reader would check, correct the record at its source, then rerun the same question on a fixed schedule and log every result. The number that settles whether wrong closure claims are common or rare is the count of tested answers that called a confirmed open business closed, out of all answers tested, on a stated date with the assistants named. We have not run that test. As of 29 September 2026 we hold no status accuracy sample, for any sector, so no such count appears anywhere on this page. What we do hold, in legal technology in India, is the measurement that explains the mechanism: on 27 July 2026 ChatGPT answered 78 blind buyer questions and afterwards confirmed it had run no live web search for any of the 78, which means every fact in those answers came from what the model remembered rather than from a current record.

That single fact is why a closure claim is usually not a lie and usually not current either. If your firm moved, renamed, restructured or lapsed a registration for a quarter, the answer can be confidently wrong and internally consistent at the same time. The work below does not argue with the answer. It fixes the records the answer was built from, and then measures whether the answer changed.

First verify whether the business is actually open

Start here even though it feels unnecessary, because in a surprising number of cases part of the business genuinely is closed, and arguing that the assistant is wrong will waste a week. Check the following six things and write the answer and the date next to each one in a sheet.

• Your company status on the Ministry of Corporate Affairs master data. Open the company master data page on the MCA portal for your own CIN and read the status field as it stands today. Do not quote what you remember it saying. A company can show as struck off or under liquidation while the business is trading, if a filing was missed.
• Your GST registration status. Search your own GSTIN on the GST portal's search taxpayer page and read the status. Cancelled and suspended are different states and both get read by an outsider as closed.
• Your Google Business Profile. Sign in and read whether the profile is marked permanently closed, temporarily closed, or suspended. A profile can be marked closed by a suggested edit you never saw.
• Your own website. Is there a page that says you have stopped a service, merged, or paused a product? Old announcement posts are the commonest self inflicted cause of this problem.
• Your entity names. If the trading name and the registered name differ, or if a group has more than one entity, list every name and every registration number in one place before you go further.
• Any entity in your group that did close. A dissolved subsidiary with a similar name is enough to produce this answer, and no amount of correcting your own page will fix it.

If any of those six says closed, your first job is a filing or a profile correction, not a content project.

Identify which AI answer and source says closed

You cannot correct a claim you have not captured. Reproduce it and record it properly, because a screenshot with no prompt text next to it is not evidence of anything.

One. Write the prompt down exactly. The wording that produced the closure claim, character for character, including any earlier turns in the conversation. Status answers are unusually sensitive to phrasing. Is X still operating, has X shut down, and is X open both give different answers on the same day.

Two. Record the conditions. Which assistant, which mode, signed in or out, location setting, language, device, and the timestamp. An answer from a logged in account with memory turned on is not the same observation as one from a fresh session, and mixing them makes your own log useless later.

Three. Save the full answer as text, not as a description of it. Copy the words. You will need the exact sentence when you raise a correction.

Four. Ask the assistant for its sources in the same conversation, and record the reply as a claim. This is the step people get wrong. What an assistant tells you about its own retrieval is not a fact about its retrieval. In the aiknowsus.com audit of September 2026, Perplexity was asked afterwards how many of the questions it had actually searched for, and it withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question. In another batch it said its claim to have searched all five was not adequately supported. Three separate batches produced that retraction. So log the answer in a column headed claimed sources, and log any URL you can open yourself in a second column headed verified sources.

Five. Open every verified source and read what it says about your status, noting the page, the sentence and the date the page shows.

Six. Repeat the same prompt three times in three fresh sessions. One wrong answer may not repeat. Three out of three tells you the claim is stable and worth the effort of a correction.

Check the official records

These are the records an assistant is most likely to reach, and the ones a human buyer will check to settle the question. Open each one for your own business today and write the status and the date in your sheet. We deliberately do not restate what any of these registers currently says about any company, including ours, because a status copied onto a web page goes stale and then becomes a false claim.

• MCA company master data, for company status and the date of the last filing on record.
• The GST portal search taxpayer page, for registration status against your GSTIN.
• Your Udyam or MSME registration, if you hold one, because a lapsed registration reads as inactivity.
• Sector registrations. A bar council enrolment, a council registration, a licence number, whatever applies to your trade.
• Google Business Profile and Apple Business Connect, which are the two map records most often mirrored elsewhere.
• Your listings on the two or three largest directories in your sector, named individually in your sheet rather than counted.

Record each one as a row with the register name, the field you read, the value, the date you read it, and the URL. That sheet is the evidence pack for every correction request below.

Follow the platform's correction or appeal path

There is no single place to correct an AI answer, and anybody who tells you otherwise is selling something. There are four separate paths and they work at different speeds.

Correct the underlying record. This is the only one that reliably works, because it changes what every downstream system reads. A filing with the registrar, a GST status update, a Business Profile edit, a directory correction. Slow, and permanent.

Use the assistant's own feedback mechanism. Each major assistant provides a way to report an answer from inside the interface. Open the current help page for the assistant you are dealing with on the day you read this and follow what it says. We are not restating those instructions here with a date on them, because they are edited and a stale paraphrase of a correction path is worse than no paraphrase.

Raise it where the assistant sources from a specific platform. If your closure claim traces to a map profile, a directory, or an app store listing, the correction belongs to that platform's own process, and your evidence pack is what makes it quick.

Publish a dated status page of your own. One page on your own domain, stating plainly that the business is operating, with the entity name, the registration number, the date of the statement, and links to the public registers a reader can check for themselves. Update the date when you review it. This does not force any assistant to do anything. It gives one that does read your site something specific and current to read, which is more than most businesses give it.

Check inconsistent or stale public pages

A closure claim often has a source you own. Go looking for it in the following five places before you conclude the assistant invented it.

• Old announcement posts saying a service is being discontinued, a branch is closing, or a product is being sunset.
• Two different addresses or phone numbers across your site, your map profile and your invoices. Contradiction reads as abandonment.
• An app store listing that has not been updated, or that has been unpublished while the web product runs.
• A subsidiary or old brand name that did close, with content still live under it.
• Pages served differently to a crawler than to a browser, which you can only find by fetching the page as a crawler and reading what comes back.

Contradiction is not a neutral state. On 6 August 2026 the clawlaw.in website and its app store listing carried different plan names and different prices for the same product, and ChatGPT noticed the contradiction and said so in its answer. An assistant that finds two versions of a fact about you will often report the disagreement rather than pick a side, and on a status question the cautious side of that disagreement is the one that hurts you.

Re-test status answers

Correction without measurement is faith. Set up a small repeat test and run it on a schedule you can actually keep.

The prompt set. Six prompts, written down and frozen. Three that name your business directly, because a status question is one of the few legitimate reasons to name yourself in a prompt. Three that do not name you, asking about your category, to see whether you are described as operating when somebody is choosing a supplier. Save the file with a version number and the date.

The engines. Name them individually in your sheet, because each one is a separate denominator. Never write all major assistants.

The repeat count. Three fresh sessions per prompt per assistant, on the same day, because a single answer can be a fluke. With six prompts, three assistants and three repeats the arithmetic is six times three times three, which is 54 recorded answers per round. Print that arithmetic next to any rate you report, so a reader can see the denominator being built.

The schedule. Once a fortnight for two months after a correction, then monthly. Choose a schedule you will complete, not the one that sounds thorough. An incomplete series is worse than a shorter complete one, because the gaps get read as stability.

The fields to log, one row per answer: date and time, assistant, mode, session state, location, prompt version, prompt text, verdict of open, closed or unclear, the exact sentence about status, claimed sources, verified sources, whether a live search was evidenced, and the screenshot file name.

The rate to publish. Answers calling you closed, over answers recorded, for the round, with the date range and the assistants named. Never a single percentage with nothing beside it.

What AI Knows Us can measure, and what it cannot change

We are the vendor of AI Knows Us, so read this section as a disclosure rather than a recommendation. What a monitoring tool of any kind, ours included, can do here is narrow. It can run a frozen prompt set across named assistants on a schedule, record every answer with the fields above, and report a dated count of answers calling you closed over answers tested. That turns an alarming screenshot somebody forwarded you into a series.

What no tool can do, ours included, is the following four things. It cannot change an assistant's answer. It cannot correct a public register on your behalf. It cannot tell you why an answer said what it said, because the selection logic is not published. And it cannot promise you a position, a phrasing or a timeline in any AI answer, and we do not sell that.

Limits of everything above

Five limits, each of which has cost somebody a wrong conclusion.

• We publish no status accuracy rate. Not an estimate, not a range, not a figure borrowed from another study. We have not tested it, in any sector, as of 29 September 2026.
• The figures on this page are legal technology in India and the AI visibility category. 78 blind questions with 0 live searches on 27 July 2026 describes one legal technology programme on one assistant. It is not a status finding and it is not about your sector.
• Three repeats is not a distribution. It tells you whether a claim is stable in one sitting. It does not give you a confidence interval, and we will not print one.
• A change after a correction is not proof the correction caused it. The assistant may have changed its own behaviour that month, and without a comparison set you cannot separate the two.
• Several counts here cannot yet be checked by you. The clawlaw.in citation of 6 August 2026 can be verified from outside. Counts taken over our own capture folders cannot, until we publish the captures.

Sources and change log

Every figure above comes from one of two recorded bodies of work. The clawlaw.in programme from July 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. And the aiknowsus.com audit of September 2026, 24 batches and 72 conversations, captured outside this repository. Counts were produced by a script over those files rather than from memory.

Version: 29 September 2026, first publication. This page gets updated when we run a status accuracy test and can print a numerator and a denominator, when any figure here is corrected, and when the capture files are published so the unverifiable rows become checkable.

Common questions

How long until the corrected status shows up in answers?

Nobody publishes a refresh interval and we have not measured one, so any specific number you are given for this is invented. What we hold is one publication timing from a different test: fourteen days from publishing a page to ChatGPT citing it, on 6 August 2026, on clawlaw.in. Treat that as one observation on one assistant and not as a schedule.

Our registered name and our trading name are different. Does that cause this?

It can, and it is worth testing directly. Run three prompts using the registered name and three using the trading name, in fresh sessions, and compare the verdicts. If only one name produces a closure claim, you have found where the record is wrong. Then make sure both names appear together, with the registration number, on one page on your own site.

Should we name ourselves in these test prompts?

For a status test, yes, because you are asking about a specific entity and there is no way to ask without naming it. That is an exception, and it matters that you know why. On 27 July 2026 two runs of the same 78 questions happened on the same 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 had measured was our own prompt. So a named prompt is fine for a status check and is useless for a visibility rate. Keep the two sets in separate files and never mix their results.

Is a closure claim actually damaging if buyers rarely see it?

Assume it is, for one measured reason. On 6 August 2026, on clawlaw.in, a headline figure on the main site could not be true, and when Claude checked it against public numbers it advised a buyer against the product, after which the company's accurate claims stopped counting for that answer. One unreliable fact about you can move an assistant from neutral to negative for a whole answer. A wrong closure claim is the same shape of problem.

What to do first

Today, open your MCA master data page and your GST taxpayer status, read them, and write the values and the date in a sheet. Then reproduce the closure claim three times in three fresh sessions, save the exact prompt and the exact sentence, and ask for sources while recording the reply as a claim rather than a fact. If the registers say open and the answer says closed, you have a sourcing fault with an evidence pack behind it, which is the only position from which a correction request gets acted on.