Brand visibility in AI answers: read and shape it
Brand visibility in AI answers explained: what assistants say about your company, which sources they read, and how to correct and strengthen them.
Brand visibility in AI answers is whether an assistant names your company when a buyer asks a question, what it says when it does, and which web pages it relies on to say it. You cannot buy a place in those answers or edit them. You can read them systematically, trace each claim to a source, and improve the sources that are accurate, thin or missing.
Most advice treats this as a contest to be mentioned more often. For a company with a reputation to protect, the harder question is quality. An assistant that names you with a two-year-old price, a closed office or a complaint thread as its main evidence has made you visible in the wrong way. The work is to decide which claims about your brand matter, find the pages behind each one, and fix what you can legitimately fix.
This guide is for the person who owns the company's name, and it is about the brand level, not any individual. It uses the documentation that Google, OpenAI, Perplexity and Anthropic publish for site owners, read for this guide and named in the text. Where a page says nothing, we say so, and our own experience is written as a tendency, never as a rule of the machine.
- A brand is mentioned in four ways: named in a list, described in a sentence, compared with others, or cited as a source. Each has a different owner and a different fix, so audit them separately.
- Providers document how their crawlers fetch pages, not how they rank them. Google says no special optimization is needed, so the lever is accurate, crawlable, consistent pages, not a trick.
- Trace every claim an assistant makes about you to a page. The page tells you whether the fix sits on your own site, on a third-party profile, in press or in reviews.
- Never plant text for machines, fake reviews or invented profiles. They break platform rules and show up in the same audits your buyers and journalists run.
- Repeat the same prompts on a schedule. One answer proves little because answers vary by wording, assistant and day.
Sources: Google Search Central page on AI features, OpenAI's crawler documentation and Anthropic's crawler help page, all read for this guide.
What brand visibility in AI answers covers and what it leaves out
Brand visibility in AI answers covers every place an assistant writes about your company without you in the room: a chat reply, an AI Overview above search results, a conversational answer in an answer engine, or a recommendation inside a research summary. It includes whether you appear, how you are described, and which links sit under the answer.
Related readBrand Ambassador Scandal: The First 24 HoursIt is not traditional search ranking, although both draw on the open web, and it is not social listening, which tracks posts rather than assistant answers. An answer can mention you without sending a click.
Overlap with neighboring topics deserves one paragraph. How an assistant builds an answer about a person, how to correct a false statement, and how Google's overview about a name behaves are each a separate subject, and this guide does not repeat them. What is different at brand level is that the questions are commercial (what does this company do, is it any good, who else does this), that the sources are broader (product pages, review platforms, directories, trade press, filings), and that you control much more of the evidence than a private person does.
Providers do not publish a formula for why one company is named before another, and anyone who claims to hold one is guessing. Answers also differ between assistants and between runs, so the unit of work is a pattern across many prompts, not a single screenshot.
Four kinds of mention, each with its own owner and fix
A mention is not one thing. Before any audit, separate the four kinds, because a company can be strong in one and invisible in another.
| Kind of mention | Question that triggers it | What you read for | Who controls the evidence |
|---|---|---|---|
| Named in a list | Which companies do this in my region or niche? | Present or absent, position, who sits beside you | Mostly third parties: directories, press, reviews |
| Described | What does this company do and who is it for? | Accuracy of offer, size, location, dates, leadership | Mostly you: your site and your profiles |
| Compared | Is this company better than that one? | Which criteria are used, which side wins, on what evidence | Shared: your pages, reviews and independent write-ups |
| Cited | Any question where your page is a source | Which of your URLs is linked, with what excerpt | You, through page content and crawler access |
The described row is the one you can fix fastest, because the evidence is your own pages. The named row is the slowest, because it depends on how many independent sources mention you in the right context. The compared row is where companies most often feel misrepresented, and the honest answer there is that only changing the evidence changes the answer.
Related readEcommerce reputation management: what to fix firstSilence usually means thin evidence or blocked access; confusion with a similar name is an identity problem, solved by consistent naming and matching details.
How assistants reach your pages: the access layer you control
Whether an assistant can read your pages at all is the one part of this subject that providers document in detail, and it is the first thing to check. Each provider runs different crawlers for different jobs, and the settings are independent, so a robots.txt file written years ago for another purpose can silently remove you from one assistant while leaving you in another.
| Provider | Crawler role in answers | Training role | What blocking does |
|---|---|---|---|
| OpenAI | OAI-SearchBot surfaces sites in ChatGPT search; ChatGPT-User acts on user requests | GPTBot, for training generative models | Blocking OAI-SearchBot keeps you out of ChatGPT search answers; blocking GPTBot signals no training use |
| Perplexity | PerplexityBot surfaces and links sites; Perplexity-User fetches pages a user asks about | Perplexity says neither is used for foundation models | Allowing PerplexityBot is recommended for results; Perplexity-User generally ignores robots.txt |
| Anthropic | Claude-SearchBot improves search results; Claude-User fetches pages for a user query | ClaudeBot, for training content | Blocking Claude-SearchBot or Claude-User reduces what Claude can index or retrieve |
| Googlebot, with the same index feeding AI Overviews and AI Mode | Google-Extended limits training and grounding in other Google systems | Blocking Googlebot or using noindex removes pages from the index those features draw on |
Every cell above comes from the provider's own page: OpenAI's overview of its crawlers, Perplexity's bot guide, Anthropic's crawler help page and Google's Search Central page on AI features. Open them before you edit anything.
Three points follow. First, training and search crawlers are separate on purpose: a company can decline training use for GPTBot or ClaudeBot and still allow the search crawlers, and OpenAI states that each setting is independent of the others.
Second, user-triggered fetchers behave differently. OpenAI says robots.txt rules may not apply to ChatGPT-User because a person started the visit, and Perplexity says Perplexity-User generally ignores robots.txt for the same reason. Blocking the search bot therefore does not make a page unreachable when a user pastes its address into a question.
Third, changes are not instant. OpenAI says its systems can take about 24 hours to adjust after a robots.txt update, and Perplexity says up to 24 hours. Google says that recrawling after you change preview controls can take days to months. Test after a day, then after a few weeks, before concluding a change did nothing.
Related readFounder reputation affects company: separate or alignRun a brand answer audit in one afternoon
An audit turns "what do the assistants say about us" into a record you can compare next month. It needs a spreadsheet, a signed-out browser and about twelve prompts. The method is ours; providers do not prescribe one.
- Write twelve prompts in four groupsThree per kind of mention. For described: what does the company do, who is it for, where is it based. For named: which firms offer your category in your region. For compared: how it differs from two named alternatives. For cited: a question your best page answers.
- Vary wording on purposePhrase each prompt two ways, formal and casual. Assistants respond to wording, and a company that appears under one phrasing and vanishes under another has a different problem from one that never appears.
- Run each prompt on each assistant you care aboutUse the same account state every time. Include one AI Overview check in regular search and one conversational check, because Google says AI Mode may use different models and show different links.
- Capture the full answer, date and linksSave text, the assistant, the date, the exact prompt and every cited address. Without the prompt, a capture cannot be reproduced.
- Mark each sentence about youLabel it accurate, outdated, wrong, unsupported or neutral. Then write the address of the page it most likely came from. Where no cited page supports it, mark it unsourced.
- Score by claim, not by answerA twelve-prompt audit usually yields a short list of recurring claims. The recurring ones, good or bad, are the real picture of your brand in answers.
Do not tune prompts until the result flatters you. Record where you tested from, because OpenAI's ChatGPT search help page says search uses a general location taken from your IP address, so answers about local providers can differ by place.
Trace each claim to the page behind it
The audit tells you what is said. The trace tells you why, and separates a plan from a complaint. For each recurring claim, open the cited links and read them as the assistant would: as text, without your own knowledge of what is true.
Four patterns recur. A claim that matches a page you wrote is simply being repeated. A claim that matches a page about another company, or an old version of yours, is a source-quality problem. A claim with no matching page is synthesized or ungrounded, and a site edit will not fix it directly. A claim that is accurate but unflattering is not an error at all.
| What the assistant says | Typical source class | Where the fix sits | Speed |
|---|---|---|---|
| Wrong offer, size or location | Your own old pages, or third-party profiles filled in years ago | Update your page, then claim and update each profile | Weeks, after recrawl |
| Outdated leadership or ownership | Press, directories, filing summaries | Ask the publisher to update, and state current facts on one page you own | Weeks to months |
| Harsh summary of customer feedback | Review platforms and forum threads | Respond to the reviews, resolve the causes, dispute only what breaks platform rules | Months |
| Confused with a similar name | Directories and thin pages that blur the two names | Identity signals: distinct naming, consistent details, a clear homepage statement | Weeks to months |
| Unsupported claim with no source | None found | Report to the provider through its feedback route; publish the accurate fact | Uncertain |
Keep a ledger with one row per recurring claim: prompt, sentence, page, page owner and planned action. In most companies, many corrections turn out to belong to a department that has never been told its pages feed what assistants say.
Fix what you own first, in this order
Start with pages you control, because they need no one's permission and they feed every assistant at once. The order matters: access, then facts, then consistency, then structure.
Related readLayoffs and employer brand reputation: what stays publicAccess. Open robots.txt and confirm that you are not blocking search crawlers by accident. A blanket block added to stop scraping can also stop OAI-SearchBot, PerplexityBot or Claude-SearchBot. Google asks for crawling to be allowed and for important text to be available in the page itself, so check that your key facts are not only inside images, scripts or a login wall.
Facts. Publish one plain page stating what the company is, what it sells, where it operates, when it was founded, who leads it and how to contact it. Write it for a stranger, in short specific sentences, with a visible date, and make sure other pages do not contradict it. Link it from the footer.
Consistency. Use the same company name, legal name, address and description across your site, social profiles, directories and registries. A company that describes itself three ways gives an assistant three versions to choose from.
Structure. Google's page on Organization structured data says the markup helps it understand administrative details of your organization and tell it apart from others, with name, url, logo, sameAs and legalName among the properties. It also says Google does not assure that any feature built on structured data will appear, and that guideline violations can lead to a manual action. Use it for true facts that match the visible text, and expect help with identification, not a ranking effect.
Shape what you do not own, without crossing lines
The second half of the picture sits on pages that belong to others: review platforms, directories, trade press, partner sites, comparison articles and forums. You cannot edit them. You can make accurate material easier to find and ask for corrections where they are due.
Related readLegal Dispute and Brand Reputation: A Company PlaybookDirectories, profiles and partner pages
Claim every profile a customer or journalist would plausibly open, fill it with the same facts as your facts page, and ask partners, resellers and investors to use your current description. These low-glamour tasks usually move a described answer more than any new content.
Press and independent coverage
Earned coverage, such as a trade article about a real launch or a customer case written with permission, gives an answer something outside your site to quote. Pitch only what is true and newsworthy. Paid placement presented as editorial risks platform penalties and a worse second story if exposed.
Reviews and customer feedback
If assistants summarize your reviews, the summary reflects what customers wrote. Answer calmly, fix recurring causes, and use each platform's dispute route for reviews that break its rules. Never write reviews yourself or recruit a crowd of accounts: it is against platform policy, often against consumer law, and the first thing a journalist checks.
Comparison and best-of prompts
When a buyer asks which company is best, the assistant draws on lists, reviews and comparison pages written by others. You cannot dictate the verdict. You can publish an honest comparison with stated criteria and fair descriptions, and make sure independent sources in your category hold your current facts. A page that says you are the best, with no criteria, is easy for any reader to discount.
A worked scenario: a regional software company, eight weeks
This is an invented scenario with generic parties. A company that sells scheduling software to clinics in one region finds, through a twelve-prompt audit on three assistants, that most answers describe it as serving "small businesses in general", name a headquarters city it left two years ago, and omit it entirely from "which vendors serve clinics in this region".
Related readNew company online reputation with no reviewsAssumptions: the website has product pages but no facts page; two directory profiles and an older trade article give the old city; and robots.txt, inherited from a former agency, disallows several crawlers.
The trace shows three separate causes. The "small businesses" description comes from the homepage, which never says clinics. The old city comes from the two profiles and the article. The omission from the vendor list comes partly from the thin evidence that anyone independent describes the company as a clinic vendor, and partly from the robots.txt rules, which turn out to block a search crawler. Each cause has a different owner and a different timeline.
| When | Action | What usually happens |
|---|---|---|
| Day 1 to 2 | Remove the accidental crawler block, publish a facts page stating the clinic focus and current city, link it from the footer | Access restored within about a day on providers that document a 24 hour adjustment; no change in answers yet |
| Week 1 to 2 | Update both directory profiles, request the publication to correct the address, rewrite the homepage opening | Profile edits appear within days; the publisher may take longer or decline |
| Week 3 to 5 | Publish two customer stories with permission, give a trade body the current description, re-run the twelve prompts | Descriptions often begin to match the facts page; the vendor-list question may still omit the company |
| Week 6 to 8 | Re-run the audit on all assistants, update the ledger, decide what to chase and what to leave | Usually a clear gain on described answers, partial gain on cited, little proof on named |
The company gains most reliably where the evidence was its own, description and location, and least where the evidence belongs to others, as with appearing on lists. A plan that promises the second outcome in eight weeks is selling something it cannot control.
What to measure and what not to read into it
Measurement here is soft. Providers publish no feed of how often you were mentioned, so any "share of voice in AI answers" figure comes from a sampling method that decides the number. Ask a vendor for the prompts, assistants, dates and number of runs.
Use three measures you can defend. First, your own ledger: the share of audit sentences that are accurate, outdated or wrong, repeated on the same prompts each month. Second, the list of cited addresses: which of your pages are linked, which third-party pages are linked beside you. Third, traffic that Google reports. Google's AI features page says traffic from AI Overviews and AI Mode is included in the Web search type of the Performance report in Search Console, though it is not separated from ordinary search traffic there. Your own referral logs may also show visits from assistants that send links.
Related readPartner and supplier due diligence on your companyDo not read one missing mention as failure or one new mention as success. A monthly pattern across many prompts is the smallest unit worth acting on. Monthly monitoring of search and AI answers, which some companies prefer to run through a specialist, is described on our reputation monitoring page.
Where this work stops, and what to do about accurate criticism
Some of what an assistant says will be accurate and unwelcome: a regulator's finding, a product failure covered by the press, a pattern in complaints. This work does not remove it and should not try. Accurate, lawful, public-interest information about wrongdoing is not ours to hide, and the attempt tends to produce a second story.
What you can do is add true context: publish what you changed, with dates and specifics, where an assistant looking for the current position can find it. Keep it factual and expect the older material to remain.
For claims that are false and damaging, the route is different: the provider's feedback channels, correction requests to publishers, and in serious cases legal advice from counsel in your country. That work is a distinct subject, and nothing in this guide is legal advice.
Edge cases: a very common name makes separation hard, a new company has little evidence to read, a regulated company should have counsel review its facts page, and a company that blocks all AI crawlers should accept being described by what others write.
Your next step: one prompt set and one robots.txt read
Do two things this week. Open your robots.txt and list what each rule blocks against the crawler names above. Then write the twelve prompts, run them on the assistants your customers use, and save everything. Those two hours give you a baseline.
Related readRebrand reputation search results: what carries overIf an assistant is describing your company wrongly to buyers or investors and you need a structured read of the answers and the sources behind them, a private audit sets out what is said, why, and which fixes are yours to make. Our page on AI search reputation describes how we approach it.
At Reputation Partners we begin the same way: read the answers, trace the pages, rank the fixes by who owns them, and say plainly which ones cannot be promised.
Questions people ask
Can a company pay to be mentioned in AI answers?
The providers we read describe crawlers and search access, not a paid placement for organic answers. Where products show labeled ads, their own terms apply. A payment meant to alter an organic answer without disclosure is manipulation; ask any vendor offering it for the provider policy that permits it.
Should I block AI crawlers?
It is a business choice with a trade-off. The providers separate training crawlers from search crawlers, so you can decline training use and still allow the search ones. Blocking a search crawler, such as OAI-SearchBot, keeps your site out of that product's answers, so decide with the cost in view and write each rule on purpose.
Why does the assistant mention a competitor and not us?
Usually because more independent sources describe the competitor in that context, or because your pages were hard to read or blocked. No provider we read publishes a ranking formula, so the reason is not knowable from outside. The pages cited for the competitor are the most useful clue.
How long before a change on my site shows up in answers?
It depends on the provider and the page. OpenAI and Perplexity say about 24 hours for robots.txt changes to take effect, and Google says a recrawl after preview-control changes can take days to months. Content changes need a recrawl first, so check again after a few weeks and compare against the same saved prompts.
Is a wrong answer about my company defamation?
Not necessarily. An answer can be outdated, merged with another company or a loose summary without being unlawful, and the legal tests differ by country. Save the evidence, correct the sources you can, report the specific response to the provider, and ask a lawyer in the relevant country before any legal step.
