AnswersWrong answers

What do you do when ChatGPT or another assistant gets your company wrong?

Work out which kind of wrong it is first, because each kind comes from a different place and is corrected by different people. An out-of-date fact, a product called discontinued, a wrong price, a mix-up with another company and being left out altogether can all start as one screenshot. Check how often a buyer meets the error, correct what you publish, then ask the other sites the assistants read to correct theirs.

What kinds of wrong answer do AI assistants give about a company?

The wrong answers an AI assistant gives about a company fall into a handful of kinds: a fact that was true once and is not now, a live product called discontinued or out of stock, a wrong price or a wrong claim about what you sell, a mix-up with another company of a similar name, and being left out of an answer you belong in. Each comes from a different place, so each is corrected differently.

Out of date. An old address, a retired product line, last year's prices, a service you no longer offer. The answer describes a company that existed once.
Discontinued, out of stock or the wrong price. A buyer told a product no longer exists does not go looking for it. Where each of these comes from, and which a brand can correct, is set out in wrong price and stock answers.
A wrong claim. A feature you do not have, a specification stated wrongly, a fault that belongs to somebody else. Sometimes the answer is stale and sometimes your own page is, and how to tell which is the subject of contradicted claims.
Mistaken for another company. Two companies with similar names, and an answer that blends them: the other company's products, location or reviews attached to yours.
Left out. Nothing it says is wrong, and the answer is wrong all the same: a buyer asks for the options in your category and you are not among them. Nobody sends a screenshot of this one, because there is nothing on the screen to object to.

Being described accurately and being named at all are two separate parts of AI visibility, and the kind decides who does the work. A retired product line is a job for your own web team, a stale retailer listing for whoever manages that account, and a mix-up partly for your own pages and partly for the profiles that describe you elsewhere.

Why does ChatGPT get facts about a company wrong?

An AI assistant answers from what its model learned before a fixed cut-off date and, when it searches, from the pages it can read at that moment. A wrong answer comes from one or more of these places: the model learned an older version of the facts, your own pages do not state the fact where a crawler can read it, other sites describe you differently or confuse you with somebody else, or your site refuses a crawler the assistant's search depends on.

The model's training has a date on it. OpenAI publishes a knowledge cutoff for each of its models, and a model has learned nothing published after its own. A price you changed, a product you launched or a line you retired after that date is unknown to the model until a newer one replaces it, and until then it answers from the older version whenever it does not search.

Your own pages count only if the fact is on them. A price that appears only once a script has run, a specification inside an image, or a fact your site never states at all leaves the assistant to fill the gap from somewhere else.

Other sites describe you too. Retailer listings, directories, review sites, comparison articles and forum threads, some of them written years ago, and an answer can repeat any of them. A company that shares your name can have its coverage blended into yours when nothing machine-readable on your site says which company you are.

Your site can shut the assistant out. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Google says AI is built into its Search, so robots.txt rules for Googlebot are a site owner's control over how the site is crawled for it, and that a page must be indexed and eligible to be shown with a snippet to appear as a supporting link in AI Overviews or AI Mode. A refused crawler leaves the answer built from what everybody else says about you. Which AI crawlers to allow has its own answer, including why a rule at your network edge can refuse one your robots.txt lets in.

How can a brand find out if ChatGPT is saying something wrong about it?

Ask it the questions your buyers ask, then ask it directly about your company, and write down every sentence about you that is not true. Asking ChatGPT alone a handful of buying questions takes about ten minutes; a proper check across several assistants and more questions, written down so it can be repeated, takes an afternoon.

The buyer's question and the direct question find different errors, so ask both, then ask again and read the sources. Treat the first answer as one reading rather than the verdict.

The buyer's question. The one somebody types before they know you exist: the best option for a small team, a named use, a budget. It shows whether you are named at all, and what is said about you when you are.
The direct question. What does your company sell, what does it cost, where is it, is a named product still available. It shows what the assistant believes about you, which is where out-of-date facts and mix-ups surface.
The same question again. The same question can return a different answer each time it is asked, so a wrong claim that appears once and never again is a different problem from one that comes back every time.
The sources. Where an answer shows the pages it drew on, open them and note the domains. A wrong fact with a source beside it tells you where the correction has to land.

Checking what ChatGPT says about you sets out the free check with a buyer's questions and what to write down for each answer. That check leaves your company's name out, because it asks whether you come up on your own. For errors, add the direct question above, and for a wrong answer write down the exact sentence, the assistant, the date, and any source it showed.

How do you correct what an AI assistant says about your company?

Correct each wrong answer where the assistant reads it, starting with what you control: state the fact plainly on your own pages, in text a crawler can read, then ask the retailers, directories and publishers carrying the wrong version to update theirs. You cannot edit an assistant's answer directly, and planted reviews or undisclosed posts are not a correction.

Out of date. Change it on your own pages first, in plain text on the page that covers the subject, and say what changed. Then find the other pages still carrying the old version, starting with any the answers cite.
Discontinued, stock or price. Publish the price and availability on each product page, in the text and in structured data, and keep both current. A retailer listing is corrected through whoever manages that account. When a lower quoted price is not an error at all is covered in why an assistant says your product is discontinued.
A wrong claim. Read your own page before you blame the assistant. An answer that disagrees with your site can be repeating an older version of it, or your page can be the one that is out of date. Who decides which side is right covers both cases, and how the evidence decides which end to correct.
A mix-up. Say who you are in a form a machine can read: your name, what you do and where, the same on every page and every profile, with Organization markup on your home page linking to the profiles that describe you elsewhere, such as a Wikidata entry. Coverage of the other company is theirs, not yours to change.
Left out. Check the crawler first, then whether your pages answer the situations buyers describe, then the third-party pages cited in answers that name your rivals. Why a competitor shows up first takes that case apart.

Google's own guidance for its AI features, last updated on 10 December 2025, lists the same basics: keep Merchant Center and Business Profile information up to date, make sure structured data matches the visible text on the page, and make important content available as text.

Outside your own site the work is asking rather than editing. Supply the correction and the evidence for it, and what a publisher, a retailer or a forum then decides is theirs. Delphi never recommends fake reviews, undisclosed staff posts or seeded threads.

How do you tell whether the correction reached the answers?

Ask the same questions again, the same way and more than once, and compare how often the wrong claim appears rather than whether one answer still carries it. Allow for the clocks: search picks up a corrected page only once it reads the page again, and a fact the model learned in training changes only when a newer model replaces it.

Compare rates, not single answers. A claim most of the assistants made that only one still makes is a result. A claim that is missing from one set of answers can come back in the next, because the answers vary from one asking to the next, so one clean check is not a correction. Whether a change is real and how much evidence it needs each have their own answer.

Each source runs on its own clock. OpenAI says its search takes about a day to adjust after a robots.txt change. Google says crawling can take anywhere from several days to several months, depending on how often its systems decide a page needs refreshing, and that a site owner can request a recrawl. A third-party page changes when its publisher changes it, and what a model learned in training changes only when a model with a later knowledge cutoff replaces it.

Ask the before and the after the same way: the same questions, in the same words, to the same assistants, in a fresh conversation each time.

What does Delphi do about wrong answers?

Delphi puts your category's buying questions to each assistant it measures, more than once, keeps every answer word for word, pulls out the factual claims the answers make about you and checks the most repeated of them against your home page and the product names your product pages declare. Where an answer and your site disagree, it quotes both sides and names the assistants carrying the claim, without ruling on which side is right.

The questions go to ChatGPT, Gemini, Claude, Perplexity and Grok, each asked more than once in a measurement. The claims the answers make about you are set against the visible text of your home page, read the way a crawler that runs no JavaScript reads it, and the product names your product pages declare in their structured data. Each claim checked gets a verdict: the two disagree, the two agree, or your site says nothing on the subject. Silence is never counted as agreement or as an error, and where nothing could be read from your site at all, nothing is compared. Which claims can be compared is set out under what makes a claim comparable.

In the product, the Claims view lists every claim your site contradicts, sorted into the kind of wrong it is (made up entirely, true once but not now, a fault you do not have, a detail stated wrongly) and by what it costs you, from costing you the sale to worth correcting. Claims answered by the same wording on your site are grouped as one problem, each group names the assistants carrying it, and a claim lists the sites the answers that made it cited, where they cited any. An identical claim from several assistants and several askings is checked once, the most repeated first, and a claim the check did not reach never counts for you or against you.

Across measurements it shows which contradicted claims survived from one run to the next, and which assistants carried each one most recently. A claim missing from one run is not marked corrected, because the answers vary between runs.

A contradicted claim can put the correction on the plan: your own source of truth first, then the third-party page repeating the claim, with the correction pursued on evidence where the page is not yours and checked again on the next measurement. A claim that describes something your site does not can also bring on the work of making the company's identity clear, and the site audit reads whether your home page declares the company in Organization markup with sameAs links to its other profiles. If your robots.txt refuses OAI-SearchBot, Claude-SearchBot, PerplexityBot or Google-Extended, opening it comes first in the plan's order, ahead of all of it.

The limits, stated plainly. Delphi asks each assistant through its API rather than the consumer app, and most of its questions reach ChatGPT without a live web search, so a correction on your pages shows in a Delphi reading first where an assistant searches as it answers, and elsewhere only once a newer model has learned it. And there is no separate check for being mistaken for another company: what Delphi catches is what the mix-up put in an answer, where the comparison with your own site catches it.