How do you fix what AI says about your company? We asked the AIs themselves

A customer calls you: an AI told them you had closed down. Or you ask ChatGPT who does your job in your city, and your name is not there. The first reflex is to ask the AI itself: how do I fix this?

We asked it for you on 3 August 2026, in eighteen forms, in English and in French, three times each to six AI engines: ChatGPT, Claude, Gemini, Perplexity, Mistral and Grok. 324 answers. They agree on the mechanism and part ways on the addresses.

On the mechanism, the message is the same from one engine to the next: there is no single place to submit your information, and what an AI says about you comes from what it reads on the web. So you fix an AI by fixing what it reads. On the addresses, meaning the places to go and correct, each AI gives its own list, and the lists only look alike at the top.

No single place to submit, they answer

Asked "Is there a place to submit my company's information to AI assistants?", 16 answers out of 18 say there is no single place. What they suggest instead fits in one verb: publish. On your own website, and on the pages others keep about you.

When an AI gets you wrong: report it, then fix what it reads

Five of the eighteen questions describe an AI saying something false about a company. Of their 90 answers:

• 86 advise reporting the error to the AI's publisher: the feedback button under the answer, a form, support;
• 80 send you to fix the pages the AI reads: your website, directories, Google Business Profile, Wikipedia, LinkedIn;
• 77 do both.

Perplexity sums up the gesture in two steps: report the error through the product's feedback channel, then correct the underlying source the AI is likely reading. The first step targets the wrong answer. The second targets what the AI will read next time.

Your website first, then a list that changes from one AI to the next

These are the destinations the 324 answers name, and the share of answers that name each one.

Your own website is the first address for all six engines, from 61% of answers for Claude to 83% for Grok. Google Business Profile is the most named platform for five engines, and tied with Wikipedia for Grok. Past that shared top, the lists diverge, sometimes by more than a factor of ten.

Read engine by engine, with the average number of platforms named per answer in brackets:

• ChatGPT (3.2): Google Business Profile (57%), LinkedIn (43%), Bing Places (39%), Wikipedia (37%), Apple (33%), Yelp (31%), Wikidata (28%).
• Grok (3.2): Wikipedia and Google Business Profile tied (56%), then Reddit and LinkedIn (33% each), G2 (26%) and Crunchbase (24%). It is the only engine to name Reddit, G2 and Trustpilot regularly.
• Mistral (2.0): Google Business Profile (46%), Wikipedia and LinkedIn (37% each), and the press release more than any other engine (39%).
• Gemini (1.4): Google Business Profile (39%), Bing Places (20%), Yelp (19%), and structured data in 78% of its answers.
• Claude (1.0): Google Business Profile (39%) and LinkedIn (26%); no other platform goes beyond one answer in six.
• Perplexity (0.6): 35 answers out of 54 name no platform at all. They speak of directories (65%) and structured data (70%) in general terms, and Google Business Profile (26%) is the only platform named in more than one answer in ten.

This extends our study of the sources engines share: on the same question, two engines cite few sources in common. They also share few addresses to send you to.

Structured data, in your website's code

214 answers out of 324 recommend structured data, most often as schema.org markup, for instance the Organization or LocalBusiness types written in JSON-LD. ChatGPT and Gemini do so in 78% of their answers, Claude in 39%. And when they say where to put that markup, they almost always point to your own website, in the code of its pages.

So AIs want your facts in a form machines read without ambiguity, and the place they point to for it is your website's code. To publish those facts without touching that code, Markleex publishes pages written for AI assistants to read, at a stable address, served as plain text and dated by the platform.

llms.txt, present when the question is about format

The llms.txt file, proposed as a table of contents of a site for AIs, appears in 28 answers out of 324 (8.6%). Its presence depends on the question asked:

• 3 of the 144 answers that start from a lived problem (an AI says something false, an AI does not know you), all from Grok;
• 23 of the 108 answers about where to publish or which format to use;
• 0 of ChatGPT's 54 answers, all questions combined.

What these answers change for you

The AIs agree on the mechanism and part ways on the addresses. Two consequences follow.

The first: the reporting that almost all of them suggest comes, in most of their answers, with work on the pages they read. That work is what can be measured, answer after answer.

The second: fixing "the AIs" as a block means following the list of just one. Depending on which AI your customer asks, the advice will send you to Bing Places, to Reddit or to Wikipedia. A correction is therefore run engine by engine: record the wrong answer, trace it back to the pages that feed it, and re-read each engine until the error stops coming back. For the pages others keep about you, our guide to third-party mentions explains how to get them.

Method

Protocol. On 3 August 2026, eighteen questions were put to AI engines through their official APIs, three times each on every engine: ChatGPT, Claude, Gemini, Perplexity, Mistral and Grok. 324 answers in all, 54 per engine. Thirteen questions are in English, five in French. Word for word, by family:

• An AI says something false: "ChatGPT is saying something wrong about my company, how do I fix it?"; "How do I correct false information an AI gives about my business?"; "An AI assistant gave a customer the wrong information about my services, what can I do?"; "Une IA raconte des choses fausses sur mon entreprise, comment corriger ?"; "Comment corriger ce que ChatGPT dit de mon entreprise ?"
• An AI does not know you: "Why doesn't ChatGPT know my business exists?"; "My company doesn't appear when I ask AI about my industry, what should I do?"; "Mon entreprise n'apparaît pas dans les réponses des IA, que faire ?"
• Where to publish: "Where can I publish official information about my company for AI assistants to read?"; "Is there a place to submit my company's information to AI assistants?"; "How do I create a page about my business that AI assistants can read?"; "Où publier les informations officielles de mon entreprise pour que les IA les lisent ?"
• Which format: "What format should my company information be in so AI reads it accurately?"; "How do I make my company facts machine readable for AI?"
• Getting recommended, and who holds the record: "How do I get my business recommended by AI?"; "What is an AI readable company registry?"; "Comment se faire connaître des IA ?"; "Who keeps the official record of a company that AI assistants trust?"

Unit of count. The answer. An answer counts for a destination as soon as it names it, once or more. The citation links some engines insert into their text are removed before counting: we count what an answer advises, not what it cites.

What counts for each destination. The thirteen platforms count by their name: Google Business Profile or Google My Business (GBP, fiche Google), Wikipedia, Wikidata, LinkedIn, Crunchbase, Yelp, Bing Places, Apple Business Connect or Apple Maps, Facebook, Reddit, G2, Trustpilot and YouTube. The llms.txt file counts by its name too. Five destinations count by a word, in both languages of the questions:

• your own website: your website, your site, official website, votre site, site officiel;
• directories: directory, directories, annuaire;
• structured data: structured data, schema.org, schema markup, JSON-LD, données structurées;
• press release: press release, communiqué de presse;
• official company registry: business registry, Companies House, LEI, D-U-N-S, Dun & Bradstreet, OpenCorporates, GLEIF, SIRENE, INSEE, Infogreffe, registre du commerce, RCS, Secretary of State.

The number of platforms named per answer counts, for each answer, how many of the thirteen platforms it names.

When an AI gets you wrong. On the 90 answers of the first family, "report" counts an answer that advises reporting the error to the AI's publisher (feedback button, report form, support); "fix what it reads" counts an answer that names your website, a directory, Google Business Profile, Wikipedia, LinkedIn or a listing.

The single place. The 18 answers to "Is there a place to submit my company's information to AI assistants?" were read one by one: 16 say there is no single place.

Scope. Three bounds define what these figures say:

• the questions are asked in English and in French;
• the counts say what an answer names, and the order in which it names it does not enter the count;
• they describe one day of measurement, 3 August 2026.

The protocol replays identically, same questions, same repetitions, same engines: this wave is its reference.