Wrong Facts About an Indian Company in AI Answers: The Public Register Is Usually the Origin, and the Order to Correct It In
For an Indian business the sentence an assistant repeats often started on the registration record or on a site that copies it, so correcting your own website leaves the original standing.
Published by AI Knows Us · Updated 16 October 2026
When an assistant states something wrong about an Indian company, the first instinct is to edit the website. That is usually the last step rather than the first. The facts an answer reaches for about a registered Indian company are mostly filed facts: the corporate identity number, the incorporation date, the registered office, the registrar, the status on the register, the last annual general meeting and the date the last balance sheet was filed. Those are published on the Ministry of Corporate Affairs portal, and a layer of company data websites copies them into a page per company that ranks for your own name. Change your home page and every copy is still there, still carrying the old value, still dated.
So the order matters more than the effort. Work from the filed record outwards: the register, the sites that mirror it, the listings that carry your trading details, then your own pages, then measure again. The capture and classification work that comes first is already written up in the five step correction, and the scoring sheet for running it as an audit is in the 30 query fact accuracy audit. This page is about the Indian chain those two pages tell you to trace.
The Indian source chain, in the order an answer gets built from it
| Layer | What it carries about you | What to check |
|---|---|---|
| The MCA portal, through its company and LLP master data lookup | CIN or LLPIN, incorporation date, registrar, registered office, status, last AGM date, last filed balance sheet date | Open your own entry and write today's date beside every field you read |
| Company data websites that republish the filed record, such as Zauba Corp and Tofler | The same fields, plus directors, charges and, on some, a previous CIN field | Search your company name in quotation marks and open everything on the first three pages that is not yours |
| The GST portal entry for your GSTIN | Legal name, trade name, registration status, principal place of business | Whether the trade name there matches the name you actually trade under |
| Directories and sector aggregators such as JustDial, IndiaMART and Sulekha | Address, phone, categories, services, hours, photographs | Categories and service lists, the fields that go stale first |
| Your Google Business Profile and map data | Status, hours, category, address, phone | The status field, because a closed marker costs enquiries daily |
| Your own site, feeds, old posts and public PDFs | Everything you have ever published about yourself | Whether the correct version is in the served HTML rather than loaded by a script |
We opened one of those company data pages on 16 October 2026 to see what it publishes. It states that its figures come from records of the Ministry of Corporate Affairs, then prints the CIN, the registrar, the incorporation date, the registered address, the current status, the last AGM date, the last filed balance sheet date and a previous CIN section. Every one is a short, declarative, quotable fact.
Why the copies matter more than they look
Three reasons, each of which changes what you do next.
- They read as authoritative because they are copies of an authoritative record, and they say so.
- They are indexed and linked better than most company About pages, so they are often the first thing returned for your registered name.
- They carry dates. A date beside a fact is what makes it quotable, and that holds for an old fact too.
The copies follow the register rather than leading it, so the register is the thing to put right first. That filing work belongs with your company secretary or chartered accountant, and this page does not describe it. What a founder or marketing team can do is everything after it: list the copies, raise the correction with each one citing your corrected record, and publish your own dated version.
The order to work in
- Open your own master data entry and read every field. Write the value and today's date in a sheet, one row per field.
- Have the filed record corrected where a field is wrong, through whoever handles your filings.
- List every copy and listing carrying the old value, one row each with its URL, the value shown and the date you read it. That contradiction map is the whole job.
- Correct the listings you control yourself, starting with the ones that publish a status or a category.
- Publish one identity page on your own domain, in plain text in the served response, carrying the current CIN, the registered office, any former names with their change dates, and the date you last checked the lot.
- Take the reading again a month later with the prompt text unchanged, on both engines, with your brand named nowhere in the question.
Where a different company is being described under your name, that is a disambiguation job with its own method, set out in the disambiguation audit. The Google side profiles are worth opening in one sitting, and the accounts and the order to open them in are in the five accounts to register your company on Google.
What the Indian readings showed
The Indian figures here come from a run on 16 October 2026: 210 questions per market, 630 in all, put to both engines, giving 1,260 readings with none discarded. The Indian half came back 2 of 210 on ChatGPT naming our own domain and 0 of 210 on Gemini. Those readings threw up 20,143 citations over 4,947 domains. Somebody's own commercial website is where 82.0% of them resolved. Review directories took 0.7%, and official platform documentation 12.3%. Run twice, eleven days apart on two question designs, the result barely moved: 3 of 597 then, 6 of 1,260 now, with the vendor own site share at 85.6% and 82.0%. On Gemini the count was 0 in 630 readings here and 0 in 291 there, 0 in 921 in total. Both runs are set out in what AI assistants actually cite across 597 readings, and the Indian half is broken out in the Indian baseline your before and after has to beat.
The city layer was measured separately on 15 and 16 October 2026: ten named cities, six buyer questions apiece, both engines, and our domain absent from every one of the 120 readings. Mumbai, Delhi and Bangalore were three of the ten at 12 readings each, and the domains that won there were small and local: searchmetric.in, nicodigital.com, onecity.co.in and aiseo.solutions.
Three limits belong beside those numbers. The questions went to ChatGPT and Gemini alone: Perplexity, Claude and AI Overviews are part of what we sell and no part of what these figures measure. Neither run counted how often an assistant states a wrong fact about a company, so we hold no accuracy rate. And there is no number for a city we did not measure: Pune, Hyderabad and Ahmedabad have no figure here, so carry the market figure instead.
The register chain questions Indian founders actually ask
An assistant says our company is struck off or not active. Where does that come from?
Almost always from the status field on the filed record or from a site copying it. Open your own master data entry, read the status and the dates beside it, then search your registered name in quotation marks to find every copy. Fix the register first, because the copies follow it.
Can I get a company data website to take my page down?
Ask through the contact route on that site, with the corrected record and the date, and log what you sent and when. The durable fix is the entry those pages copy, which is why the register comes first.
Our old CIN still appears in answers. What fixes that?
Some of those company pages carry a previous CIN field, so an old identity stays visible long after it changed. Publish one dated identity page naming the current CIN, the former one and the change date, and use the same name form on every listing.
Should we spend on JustDial and review sites to fix this?
Keep them accurate because they carry your address, categories and phone, and put the quarter elsewhere. In the 16 October 2026 count a review directory appeared in 0.7% of the 20,143 citations, against 82.0% on somebody's own website, so the listing is housekeeping and your dated pages are the work.
How long before an answer changes?
Re-measure monthly on a frozen set of blind questions across both engines and read the series rather than one reading. The public record is the part you control, and the series is what tells you a correction reached an answer.
Does this cover a wrong product price or MRP?
That is a different chain, running through your listings, feeds and marketplace pages, and the Indian product data version is in the declarations you already print.
The filed fact is the one that travels
An Indian company is unusually well documented by the time anybody asks an assistant about it. The identity, the address, the status and the filing dates are already public, already dated, and already copied onto pages that outrank the company's own site for its own name. That is why editing the About page feels like work and changes nothing: the sentence in the answer was never taken from there. Treat your registered record as the first page you publish and your identity page as the second. That measurement, and the pages that follow it, is what AI Knows Us charges for, so open your own register entry before taking our word for any of this.