AI brand perception and reputation: how ChatGPT, Claude, Gemini, Google and Meta AI describe your brand, and what to do when they get it wrong

Visibility measures whether AI names you. Perception measures what it says when it does. They are measured separately because they fail separately, and together they produce four very different situations:

• Visible and well described: AI names you and describes you the way you want. The job is to keep it that way and check that it doesn’t drift.
• Visible and poorly described: AI names you, but with an outdated fact, an exaggerated flaw, or without what sets you apart. This is the worst of the four, because AI amplifies a story that works against you.
• Invisible, but well described when asked by name: AI knows you but doesn’t bring you up when the question is about the category. That is a visibility problem, worked on through share of voice and content.
• Invisible and undescribed: AI has no material about you. First you have to exist.

For visibility, read the guide to AI share of voice. This guide covers perception: what gets measured, how to read it, what to do about it, and what to do when AI gets a fact wrong.

Sentiment analysis scores the tone in which each answer talks about your brand on a 0–100 scale. In Refractia the bands are: 70 or higher, positive; 40 to 69, mixed; below 40, negative. A high average is good news, but it doesn’t tell you why, and two brands with the same score can need opposite actions:

Seguros Lince has to address one specific pain point, with facts AI can repeat. Aseguradora Brisa needs a story of its own. The score is identical; the work isn’t. That is why sentiment is always read next to the descriptors, and per engine: a chat engine can carry a criticism from years ago that AI search no longer shows.

Descriptors are the adjectives and short phrases engines use to characterize you: “reliable,” “expensive,” “pioneering,” “hard to reach.” Aggregated across hundreds of answers, they show the brand image AI has put together, in more detail than any score.

• Frequency: which words repeat most. The ones that appear in most answers are the core image; the ones that appear once are noise.
• Tone: which descriptors are positive, negative or neutral. One heavily repeated negative weighs more than ten scattered positives.
• Vocabulary diversity: how many different terms AI uses to describe you. A few repeated terms point to a consistent image; many scattered terms point to a blurry one, where every answer says something different.
• Differences by engine: if ChatGPT calls you “traditional” and Gemini calls you “innovative,” there isn’t one AI perception but several, and you want to know which one your customer sees.

Every brand has traits it wants to be known for: fast, approachable, safe, premium. Trait alignment compares those desired traits with what AI actually says. The trap is counting only whether the trait appears: if “innovative” shows up twice among three hundred mentions, it is technically there, but nobody perceives it.

That is why alignment is weighted by prominence: a trait counts in proportion to how often it repeats compared with the rest of the vocabulary. A brand that wants to be “fast, approachable and innovative” and gets “fast” all the time, “approachable” now and then and “innovative” never is not aligned on three out of three, even though two of the three traits “appear.”

The comparison with competitors completes the picture. A perception matrix shows, trait by trait, whether AI assigns it to you more, equally or less than to each rival: wins, ties and losses. Sometimes the trait you want is exactly the one AI gives to someone else.

Beyond individual words, AI tells a story about your brand: where it comes from, who it is for, what it stands out for, and what it gets criticized for. That narrative sorts into three parts that help you decide:

• Strengths: what AI consistently says in your favor. This is what you need to keep backing with fresh evidence.
• Vulnerabilities: the negatives that repeat, whether a real problem, an outdated fact or an unfavorable comparison. This is the first thing to investigate.
• Opportunities: what AI doesn’t say and could. Desired traits that are missing, topics where your competitors show up and you don’t, uses nobody attributes to you.

A fictional example: a gym chain in Colombia whose AI narrative is “cheap and with lots of locations, but with old equipment and crowded at peak hours.” The strength is price and coverage; the vulnerability is the in-gym experience; the opportunity is that no answer mentions its group classes, which are its big bet for the year.

AI doesn’t hold a single perception of your brand: it answers differently depending on who is asking and what they ask. A buyer, an investor and a job candidate ask different questions, and engines look in different sources to answer them.

To measure it, build a set of questions for each audience and compare them on the same engines. Add “who is … for?” questions too: they show which segment AI associates you with. If you want to sell to small businesses and AI only recommends you for large companies, you have a positioning problem that no visibility metric will show.

Your brand’s perception in AI moves even if you do nothing: models get updated, the web publishes new articles, and your competitors change their content. You detect that drift by comparing scans, on one condition: the same questions, on the same engines, in the same country. If you change the questions, you can’t tell whether the perception changed or the way of asking did.

When you compare, look at the sample. A jump in sentiment across ten answers can be noise; the same jump across three hundred is a signal. And look at which descriptors came in and which dropped out: it is usually the first clue to what new source AI is reading.

Each perception finding calls for a different kind of action. Before you produce any content, identify which one you are dealing with:

• A false or outdated fact: find the source and correct it. The next section explains how, step by step.
• A fair criticism: if AI says your service is slow and it is, the work is operational, not communications. Content doesn’t cover up a real problem.
• A desired trait that’s missing: publish evidence that proves it, with data, cases and third-party coverage, not adjectives about yourself. AI repeats facts more than promises.
• An unfavorable comparison: check which comparison pages exist in your category and your country, and whether your brand appears in them with current data.
• Differences between chat and AI search: if AI search already describes you well and chat doesn’t, the problem is the model’s memory. The web is already fixed; it takes time and sustained presence.

This is where perception becomes reputation. An AI assistant doesn’t make things up for the fun of it. It builds its answer from what it has at hand, and that comes from two places.

• The model’s memory. ChatGPT, Claude, Gemini and Meta AI answer largely from what they learned during training, which has a cutoff date. If for years the web repeated a price, a CEO or a product that no longer exists, the model learned it that way. What was repeated most weighs most, even if it is old.
• The pages it finds today. Google’s AI Overviews, and chat engines when they search the web, build the answer from pages that rank right now for that question: a five-year-old press article, a forum thread, a directory listing with your old address, a competitor’s comparison page.

When information is missing, the model also fills the gaps with whatever seems likely, and that is where the strangest errors come from: locations that don’t exist, services you never offered, or facts about another company with a similar name.

Typical errors

• Prices, plans or terms from years ago.
• Discontinued products presented as current.
• Executives, owners or an address that have since changed.
• An old controversy presented as current.
• Confusion with a similarly named company, sometimes from another country.
• Wrong coverage: “doesn’t operate in Mexico” when it does, or the other way around.

Why online reputation now includes AI

Online reputation used to mean what showed up on the first page of Google and in reviews. Now it also includes what an assistant answers when someone asks “is this company trustworthy?” ChatGPT alone has 900 million weekly users, according to OpenAI. The difference from a search result is that the answer arrives already written, in a voice of authority and often without the link that would let someone check it. The person asking doesn’t see the 2019 article the fact came from; they see a statement. And unless you ask, you never find out.

Before you correct anything, you need to know what to correct. These steps take you from the answer to the source:

1. Reproduce the error with several phrasings. “What do people think of X?”, “Is X trustworthy?”, “problems with X.” If it shows up once, it may be chance; if it keeps showing up, there is a source behind it.
2. Test it on several engines and in your country. Note which ones show it. If the error appears in AI search and not in chat, the source is a page that ranks well today; if it appears only in chat, it most likely comes from the model’s memory.
3. Follow the cited links. Google’s AI Overviews and chat engines that search show their sources. Open them: the false fact is often right there, word for word.
4. If there is no citation, search for the phrase. Search the web for the exact claim. The fact the model repeats is often on several pages that copied one another.
5. Look at which domains AI cites for your brand in general. Not just for this answer: the sites engines rely on when they talk about you are the ones worth keeping up to date.

• Fix your own pages first. About us, FAQ, pricing, locations, leadership team. Each fact should appear once, with a date, and no old pages should contradict the new ones: redirect them or update them.
• Update third-party profiles. Google Business Profile, industry directories, marketplaces, professional networks, app stores and Wikidata. On Wikipedia, if you have a conflict of interest, the community’s rules ask you to propose changes on the article’s talk page instead of editing the article yourself.
• Publish a clear facts page. A page with the company’s core facts, in short verifiable sentences, with an update date and the answer first. If you use organization structured data, it should say the same thing as the visible page.
• Ask whoever published the error to correct it. The outlet, the directory or the comparison site. An article corrected at the original source is worth more than ten clarifications of your own.
• Earn new, verifiable coverage. Interviews, reviews, a place in the comparisons for your category. AI cites third parties: if the only recent source about you is the old article, that is the one it will use.
• Use the engines’ feedback channels. ChatGPT, Gemini and Google’s AI Overviews let you rate or report an answer. It puts the problem on record, but there is no promise that the answer will change, or when.
• If there is defamation or exposed personal data, talk to a lawyer. That is a legal matter, not a content one.

For a longer walkthrough of specific error types, from wrong pricing to confusion with a competitor, read how to fix AI misinformation about your brand.

• Hide text for AI or bury instructions in your pages. Search engines treat it as spam, and it can cost you your indexing.
• Publish dozens of near-identical pages repeating the clarification. Google warns that generating pages at scale for every variation of a question can violate its policies.
• Buy reviews or mentions. If it comes out, the reputational damage is bigger than the original error.
• Argue with the chatbot. Convincing ChatGPT in your own conversation doesn’t change what it tells other users.
• Try to erase legitimate criticism. Pushing to get a fair review taken down usually generates more coverage of the problem.

When the error is serious (AI claims your company went bankrupt, that a product is dangerous, or that you are involved in a scam), handle it as a crisis, in this order:

1. 01

   Record the evidence

   The exact question, the engine, the date, the country and the full answer, with a screenshot. Without this you can’t measure whether it improves, or show it to anyone.

2. 02

   Measure the scope

   How many engines show it, and in how many of the relevant questions: “3 of 5 engines, 12 of 40 questions.” An error in one answer is not the same as an error in half of them.

3. 03

   Find the source

   Follow the steps in how to find where the error comes from. Separate what comes from current pages (it gets corrected faster) from what comes from the model’s memory (it takes longer).

4. 04

   Fix your source of truth

   Your site has to state the correct fact, on a page that is easy to find, with a date. It is the only thing you fully control.

5. 05

   Bring in third parties

   Ask outlets and directories for corrections, update profiles and, if it makes sense, publish a statement the press can quote.

6. 06

   Report it to the engines

   Use each engine’s feedback channels with the answer you recorded. Don’t expect a timeline.

7. 07

   Measure again with the same questions

   After the changes, repeat exactly the same questions on the same engines and in the same country. If you change the questions, you can’t tell whether the answer improved or the query changed.

Most errors are caught late because nobody is looking for them. Three habits reduce the risk:

• Include risk questions in your measurements: “Is X trustworthy?”, “Is X a scam?”, “X complaints,” “Does X still operate in Chile?” These are the questions people ask before they buy or sign.
• Strengthen your entity. A consistent name across every source, a Wikidata item where it applies, and a correct Google knowledge panel reduce confusion with similarly named companies.
• Look at the vulnerabilities, not just the average. Good overall sentiment can hide a criticism that repeats heavily on one specific topic. That is why the narrative and the descriptors matter more than the score.

The AI Perception module shows, for each engine, sentiment from 0 to 100, the descriptors it uses, trait alignment weighted by prominence, vocabulary diversity, a positioning radar, and a perception matrix against your competitors with wins, ties and losses per trait. The narrative is laid out as Strengths, Vulnerabilities and Opportunities. The desired traits come from Identity & DNA, the brand profile Refractia builds by reading your website.

Evolution shows how sentiment changes for your brand and each competitor between scans, and the Prompts Library stores every question with each engine’s answer, so you can read the source behind any number, or an error word for word and on which engines it shows up. Influence Sources shows which domains engines cite when they talk about your brand, separating your own site from media and other third parties, and whether those sources mention you or link to you. Identity & DNA shows entity strength signals, such as whether a Wikidata item exists and whether Google’s knowledge panel matches your brand.

For people (leaders, candidates, executives) the product switches to a public-figures mode with a reputation framing: AI reputation, performance by issue, presence against opponents, and a War Room. We measure ChatGPT, Claude, Gemini, Google’s AI Overviews and Meta AI. One important limit: every scan is started by a person. Refractia is a deep diagnostic, not an alarm that watches the answers all day.