How AI Assistants Decide What to Say About Your Brand, and How to Change It

11 min read
How AI Assistants Decide What to Say About Your Brand, and How to Change It
11 min read

Ask ChatGPT, Gemini or Perplexity about your company and you will get a confident paragraph. It might describe what you do accurately. It might also repeat a complaint from years ago, confuse you with a business of a similar name or leave out the clients and awards that took years to win. For the buyer, investor or candidate who asked, that paragraph is the first impression, and it may be the only one.

None of this is beyond your influence. Most of what an assistant says about a named business comes from sources it finds on the web when the question is asked, and those sources can be found, checked, corrected and added to. The work is slower than editing a web page, but it is not guesswork. It is a core part of our online reputation management work.

This article explains where AI answers about a brand come from, how to make your facts consistent, how to correct problems at source and how to track progress. For the board-level view of search reputation, see What do Google and AI say when someone searches our name?

Where an AI answer about your brand comes from

An assistant can draw on two things: what the model learned in training, and what it retrieves from the web when the question is asked.

Training is the slow layer. OpenAI says (page updated 30 April 2025) that its foundation models learn from information publicly available on the internet, from information obtained through partnerships with third parties, and from data provided by users, human trainers and researchers. What a model absorbed about you reflects the web at the time that data was gathered. You cannot edit it, and it is updated only when the provider trains a new model.

Retrieval is the fast layer, and it is the one you can influence soonest. Each of the main assistants documents how it works:

  • ChatGPT searches the web automatically when a question might benefit from it, sends search queries to third-party providers such as Bing, and shows inline citations with a Sources list (OpenAI Help Centre: ChatGPT search, checked November 2026).
  • Claude uses a search tool to ground its answers in the live web when current information helps, and includes citations so users can check them (Claude Help Centre: using web search, checked November 2026).
  • Perplexity searches the web in real time and adds numbered footnotes that link to its sources (Perplexity Help Centre, checked November 2026).
  • Gemini shows a Sources button with links when it has them, though Google notes that not every response includes sources (Gemini Apps Help, checked November 2026).
  • Google's AI Overviews are integrated with its core web ranking systems and built to show information backed up by top web results (Google, 30 May 2024). Both AI Overviews and AI Mode may run several related searches across subtopics before building a response, which Google calls query fan-out (Google Search Central: AI features and your website, updated 10 December 2025).

The practical consequence is simple. For a question about your company, the assistant runs searches, reads what it finds and summarises it. The sources tend to be the same kinds of page that rank for your name: your own site, news coverage, review platforms, Wikipedia, directories, business profiles, forum threads and comparison articles. Which ones get cited varies by assistant and by question, so the only reliable way to know is to look.

If those pages are outdated, thin or hostile, the answer will be too. When Google explained early AI Overview errors in May 2024, it pointed to misread queries, satirical and forum content, and data voids, meaning topics with little good information available. A mid-market brand with little coverage beyond its own website has a data void of its own, and an old complaint thread can fill it.

Make your facts consistent and easy to read

An assistant has to work out which entity you are before it can describe you. A company that is called one thing on its website, something slightly different on LinkedIn and a third thing in an old directory, with two founding dates and a former chief executive still listed, makes that harder. Inconsistent facts produce inconsistent answers, and they leave room for confusion with similarly named businesses.

Start with one approved set of facts and apply it everywhere:

  • Agree the facts. Legal and trading names, what you do in one sentence, founding year, headquarters, leadership, the figures you are happy to publish and the categories you want to be known for.
  • Fix your own site first. The about page, leadership pages, footer and press page should state those facts plainly in text, not only in images or PDFs.
  • Add Organization structured data. Google says it helps it understand an organisation's administrative details and disambiguate it in search results, and its sameAs property lets you list your profiles on other sites, such as social and review platforms (Google Search Central: Organization structured data, updated 8 September 2026).
  • Update the profiles you control. Google Business Profile, LinkedIn, industry directories, partner pages and award listings. If Google shows a knowledge panel for your brand, you can get verified on Google and suggest changes through the panel's feedback option.

Then check that assistants can read your pages. OpenAI and Anthropic both run the crawler that finds pages for search answers separately from the one that gathers training data, so a business can opt out of training and still be found. Blocking the wrong one removes you from answers. OpenAI's crawler documentation (checked November 2026) says sites that opt out of OAI-SearchBot will not appear in ChatGPT search answers, though they can still appear as navigational links. Perplexity uses PerplexityBot to surface and link websites in its results (Perplexity crawler documentation). Anthropic says (7 April 2026) that blocking Claude-SearchBot may reduce your visibility and accuracy in its search results. For Google's AI features, a page needs to be indexed and eligible to appear in Search with a snippet.

Correct problems at source, then add to them

You cannot edit an AI answer. You can change what it reads. When an answer is wrong or unfair, follow the citation back to the page it came from and work out which of three situations you are in.

  • The source is yours. An outdated about page, an old press release, a leadership page listing someone who left years ago. Fix it, update it or redirect it. This is the quickest win and the one most often overlooked.
  • The source is a listing you can influence. Directories, review platforms, partner pages and award pages usually offer a way to claim a listing or request a correction. Use it, and keep a record of what you changed and when, so you can match it against your tracking later.
  • The source is independent. A news article, a blog post or a forum thread. If it contains a factual error, a polite correction request to the publisher, with evidence, sometimes works. If it is accurate but negative, it will usually stay, and our guide to removing negative content from Google sets out what is realistically possible.

Wikipedia needs care. Its conflict of interest guideline (checked November 2026) strongly discourages people connected with a company from editing its article directly, asks them to propose changes on the article's talk page, and requires paid editors to disclose who is paying them. Follow it. Edits made around the rules tend to be undone and can draw attention to the very thing you wanted to correct.

Correcting errors removes bad material, but it does not change the balance of an answer on its own. The assistant also needs credible positive sources to cite. That means filling the data void: case studies with named clients, leadership interviews, expert commentary, awards and coverage on sites you do not own. Independent coverage is the hardest for a sceptical reader to dismiss, which is why Digital PR sits at the centre of this work. Genuine reviews matter too, because review platforms often rank for brand names, and our guide to online reviews and your search reputation covers how to earn more of them.

The same pages also help push negative results down in Google. Our guides to building positive content that ranks for your brand name and pushing negative search results down cover that side in detail.

Track a fixed set of prompts

AI answers vary from one run to the next. OpenAI notes that the same question may produce different answers on different occasions. One screenshot proves little. A fixed set of prompts, run the same way each month, shows whether the picture is moving.

A useful set covers:

  • The company name alone, and the name with words people add, such as reviews, complaints, pricing or jobs.
  • Direct questions: what the company does, whether it is trustworthy and who runs it.
  • Comparisons with two or three named competitors.
  • Category questions where you would expect to be recommended, such as the best providers of your service in the UK.
  • The names of the chief executive and founders, where they are in scope.

Run each prompt in ChatGPT, Gemini, Perplexity, Claude and Google's AI Overviews and AI Mode, more than once, from a browser that is not signed in. Record four things: whether the answer is accurate, its tone, whether you are mentioned at all, and which sources are cited. The sources column is the most useful, because it tells you which pages to fix, strengthen or outrank. Our article on the first ten things we check in an AI visibility audit covers the wider set-up.

Expect change to be gradual. Retrieval responds as pages are updated, recrawled and re-ranked. Google says that after you change how a page may be shown, recrawling can take anywhere from days to months. OpenAI says changes to its search crawler settings in robots.txt take about 24 hours to take effect, but that only controls access, not what the answer says.

Mistakes, and when this will not work

Most failures in this area come from a handful of avoidable mistakes:

  • Blocking the search crawlers. A blanket block on AI bots, added to stop training, can also keep your pages out of search answers. Check robots.txt and any bot rules at your content delivery network against each provider's documentation.
  • Looking for a hidden switch. Google says there are no additional requirements or special optimisations needed to appear in AI Overviews or AI Mode, and no special files or markup. Accurate, consistent, well-linked pages do the work.
  • Manufacturing sources. Fake reviews, sock puppet forum posts and thin sites written to praise the brand are dishonest, and fake reviews have been banned in the UK since 6 April 2025. When they are found, they become the negative story.
  • Judging on one answer. Answers vary between runs, products and locations. Make decisions on the tracked set, not on a single screenshot.
  • Treating AI as a separate problem. If an article ranks on page one for your name, assistants can find it too. Search reputation and AI reputation are mostly the same work, checked in two places.

This approach also has limits. If the negative material is recent, accurate and widely reported, assistants will reflect it, and they should. The realistic aim is an answer that is balanced, current and accurate, and that includes your side, not an answer that leaves the story out.

Questions to ask your team or agency

  • Which sources do ChatGPT, Gemini, Perplexity and Google's AI features cite when asked about us, and which of them do we control?
  • Do our website, profiles and listings state the same facts about who we are, what we do and who runs the company?
  • Does our robots.txt or content delivery network block any of the crawlers that AI search features use to find pages?
  • What fixed set of prompts do we track, how often do we run it and who sees the results?
  • Where an answer is wrong, have we traced it back to its source and asked for a correction?

How we help

Our online reputation management service applies the same principle to AI answers as to Google: highlight the positive to push down the negative. We trace what assistants say about you back to the sources they cite, correct what is wrong where there is a legitimate route to do so, make your facts consistent across your site and profiles, and promote and create positive content on your own site and on third-party sites, so assistants have better material to draw on. It runs alongside our AI search optimisation work for commercial searches. If you want to know what AI assistants say about your brand today, talk to us.

Frequently asked questions

Where does ChatGPT get its information about a company?

From two places. Its underlying model learned from publicly available web content and licensed data during training, which changes only when a new model is released. For many questions, ChatGPT also searches the web at the moment of asking, sending queries to third-party search providers such as Bing, and cites the pages it uses. The pages that rank for a company's name therefore shape much of what ChatGPT says about it.

Why does an AI assistant say something wrong about my business?

Usually because of the sources it found. An assistant may misread the question, confuse you with a similarly named company, or rely on outdated pages, old forum threads or a single negative article when little else exists. Google has described these gaps as data voids. Checking which sources the answer cites, then correcting or outweighing them with accurate, credible pages, is the most reliable way to change it.

Does blocking AI crawlers stop assistants talking about my company?

No. It mainly stops them reading your own pages. OpenAI says sites that opt out of its OAI-SearchBot crawler will not appear in ChatGPT search answers, and Anthropic says blocking Claude-SearchBot may reduce visibility in Claude's search results. The assistants can still describe you using other people's pages. A business that wants to stay out of model training can usually block the training crawler and still allow the search crawler.

Do I need special markup or files to appear in Google's AI Overviews?

Google says no. A page needs to be indexed and eligible to appear in Google Search with a snippet, and there are no additional technical requirements or special optimisations for AI Overviews or AI Mode. Standard good practice still applies: let Google crawl the page, link to it internally, keep structured data consistent with visible text and keep your Business Profile up to date.

How do you track what AI assistants say about a brand?

Use a fixed set of prompts covering the brand name, direct questions about the company, competitor comparisons, category recommendations and leaders' names. Run each one in ChatGPT, Gemini, Perplexity, Claude and Google's AI features, several times, from a browser that is not signed in. Record accuracy, tone, whether the brand is mentioned and which sources are cited, and repeat monthly to see the trend.

Sources

Let’s create your future growth plan

Get My Growth Plan
Visibility Playbooks

Related Playbooks

All Playbooks
More Insights

More articles we
think you’ll like

Explore More Insights