First-party data · Published

India AI Reputation Index

What an AI assistant, and the open web behind it, say about the companies of the Nifty 100. One automated scan of each; 99 of the 100 are counted and reported in total. No company is named with a figure.

01Key findings

What does an AI assistant say about India’s largest companies?

In November 2026, an AI assistant was asked about each company in the Nifty 100: who it is, what it is known for, and whether there is anything to know before dealing with it. Answering from memory, it recognised 99 of the 99 companies counted and raised an adverse matter, such as a penalty, an investigation, a lawsuit or a controversy, about 96 of them (97%). 73 of the 99 answers from memory (74%) carried at least one point that did not match what a web search showed the same day. Searching for trouble, the scan read 1,943 web results about the companies; 1,254 of them (65%) were harmful or mixed, and 64% of those sit on pages where no statutory takedown clock applies: a publisher’s own article, a court record or an official record.

97%

of answers from memory raise an adverse matter about the company

74%

of answers from memory had a point that did not match the web that day

65%

of the web results read were harmful or mixed

64%

of those sit where no statutory takedown clock applies

02What the assistant says

An AI answer is now a first impression

The same question was put to the assistant about each company: who it is, what it is known for, and whether there is anything to know before dealing with it. Of 99 companies:

  • Recognised from memory99 · 100%
  • Answer from memory raises an adverse matter96 · 97%
  • Answer with web search raises an adverse matter99 · 100%

    For 3 of them the adverse matter appeared only once the assistant searched.

  • Answer from memory had a point that did not match the web that day73 · 74%

    Out of date or wrong, as marked by the same model against what the search showed. A flag, not a ruling, and not checked by a person.

03Where the harmful results sit

Where the harmful results sit, and which have a legal clock

The scan read 1943 results about the companies, a median of 20 each. 1254 were harmful or mixed: they make or repeat an allegation, a complaint or an adverse report. That is a description of the page, not a finding that it is true or false.

All results, by how they reflect on the company

  • Harmful841 · 43%
  • Mixed413 · 21%
  • Neutral496 · 26%
  • Favourable193 · 10%

Harmful and mixed results, by kind of page

  • News article717 · 57%
  • Complaint board143 · 11%
  • Blog post113 · 9%
  • Review89 · 7%
  • Court record or case-law database64 · 5%
  • Other page50 · 4%
  • Government or regulator record26 · 2%
  • Forum post18 · 1%
  • Social media post12 · 1%
  • Business listing11 · 1%
  • The company’s own site or profile6 · 0%
  • Video5 · 0%

The same results, by the route that applies under Indian law

RouteResultsShareTime the rules allowProvision
Impersonation of a business10%Within 36 hours of the grievance complaintIT Rules 2021, Rule 3(2)(a)(i), proviso
Allegations on a platform44035%Acknowledged within 24 hours, decided within seven daysIT Rules 2021, Rule 3(2)(a)(i)
News coverage71757%No statutory clock applies to a publisher's own articleNone
Court records and case-law databases645%No fixed statutory clockNone
Government and regulator records262%No fixed statutory clockNone

The takedown deadlines in the IT Rules 2021 bind platforms that host other people's content. A news publisher's own article, a court record and a regulator's record are outside them, which is why 64% of the harmful and mixed results here have no statutory clock. The route table is the one behind every RepuLex Scan report.

04By sector

How the sectors compare

Every one of the 99 companies had harmful or mixed results: a median of 13 among the twenty or so read for each. The sectors differ only a little.

SectorCompaniesHarmful or mixed results, eachAdverse matter, from memoryAnswer did not match the web
Automobile and Auto Components913.4100%44%
Capital Goods1110.382%73%
Fast Moving Consumer Goods712.1100%86%
Financial Services2213.0100%73%
Healthcare814.5100%100%
Information Technology613.383%67%
Metals & Mining813.0100%88%
Oil Gas & Consumable Fuels611.0100%50%
Power613.2100%83%
Other sectors1612.8100%75%

Sector names are the stock exchange's own. Sectors with fewer than five companies are reported together. With this few companies in a sector, a gap of a few points means little.

On the exposure scale a visitor's own scan uses, 98 of the 99 companies come out High. That scale is built for a person or a smaller business. For a company this size it mostly measures how much is written about it, so no level is reported here. Results about a different company of the same name turned up for 19 of them and are left out of every figure.

05Method and limits

How this was measured

Each company was run once through RepuLex Scan, the same automated check a visitor runs on the website. For each company an AI model (claude-opus-5-5) searched the open web with up to eight searches, wrote notes on up to twenty results, and sorted each result by the kind of page it is and by whether it reflects on the company badly (harmful), partly (mixed), neutrally or well. The same model was asked "Who is this company, what is it known for, and is there anything I should know before dealing with it?" once from memory with no search, and a short answer was written from the search results. A second, smaller model read both answers and recorded whether each raises an adverse matter. The company list is the Nifty 100 as published by the National Stock Exchange of India and read on 2026-11-11. 99 of the 100 scans are counted; 1 was held back by the scan’s own rule that passes sensitive matters to a person.

What it cannot tell you

One run per company on one day, by one AI assistant: another day, another assistant or another wording can give a different answer. The question put to the assistant asks whether there is anything to know before dealing with the company, which invites a caution, so the share of answers that raise an adverse matter belongs to that question and not to every question a person might ask. A point in an answer from memory is marked as looking wrong or out of date by the same model, by comparing it with what the search showed that day; it was not checked by a person. The scan looks for trouble on purpose: it searches the company name with words such as fraud, complaint, case and controversy, so the share of harmful results shows what turns up when someone goes looking for problems, not the balance of everything written about a company. It reads at most twenty results per company, chosen by the model as the most consequential, so the counts describe what a first look surfaces and not everything published. "Harmful" means the page makes or repeats an allegation, complaint or adverse report; it is not a finding that anything on the page is true or false. Sorting is done by an AI model and was not checked result by result. No company is named with its figures, and sectors with fewer than five companies are reported together.

Companies counted
99 of 100
List
Nifty 100, read 2026-11-11
Scans run
2026-11-11
AI model
claude-opus-5-5
Web searches made
752
Scans held back, not counted
1

Citation: “India AI Reputation Index, RepuLex, November 2026, https://www.repulex.com/resources/india-ai-reputation-index”. Reuse is permitted with attribution. Machine-readable copy: JSON. Corrections: hello@repulex.com.

Run the same scan on your own name

The scan behind this index is free for any person or business. It takes about two minutes and shows what the web and an AI assistant say about you, with the route for each harmful result.