Managing your LLM reputation means monitoring how large language model (LLM) products such as ChatGPT, Claude and Perplexity describe your organization, then correcting the sources behind those answers. It is a communications discipline, and no company can dictate what a model says. In a 2025 Pew Research Center analysis of Google searches, users who saw an artificial intelligence (AI) summary clicked through to a website in 8% of visits, versus 15% without one.
If an AI answer names a former chief executive officer (CEO) or calls a settled lawsuit open, most readers never see the correction on your website.
The workflow audits those answers, traces each error to its source, and maintains a Generative Knowledge Graph: a consistent, dated set of facts about the organization, published on its own site, in structured data and through independent sources that AI search tools retrieve.
What Is LLM Reputation Management?
LLM reputation management is the practice of tracking what generative AI products say about an organization and fixing the inaccurate sources behind those answers.
In practice, a team asks AI assistants the questions its audiences ask, records the answers, checks which pages they cite and corrects false, outdated or inconsistent statements where they originate, whether a company website, a news archive or a Wikipedia entry.
AI outputs are probabilistic, so a model can skip a correct source or favor a louder, newer page.
The audience is large: a June 2025 Pew Research Center survey found 34% of U.S. adults had used ChatGPT, roughly double the 2023 share.
What Shapes How ChatGPT, Claude and Perplexity Describe a Brand?
ChatGPT, Claude and Perplexity build answers about a brand from two inputs: text the model processed during training, and web pages a search tool retrieves when a user asks.
Training data has a cutoff date, so later changes stay invisible unless retrieval fills the gap. Retrieval works like a search engine: the assistant runs a query, pulls a handful of pages and summarizes them.
Communicate Magazine's analysis of LLMs and public relations argues that standalone AI search engines lean heavily on earned coverage and cross-check claims against authoritative sources. If so, one launch-day press hit counts for less than the same proof point repeated across trade press, regulatory filings and the company's own fact page.
Consider a hypothetical software company that replaced its founder as CEO in 2025. Its newsroom announced the change, but a 2022 funding story, an old conference bio and an unedited Wikipedia infobox all name the founder. An assistant retrieving three of those four pages will likely name the wrong CEO.
AI answers now belong on the list of forms of the media a comms team audits on a schedule.
A brand audit across AI platforms is a fixed set of prompts, run on a schedule in each major assistant, with every answer scored against approved facts and traced to its cited sources.
Citations make the audit possible: ChatGPT search, launched by OpenAI in October 2024, attaches links to the sources behind answers, and Perplexity shows numbered citations beside its responses.
| Step | What to do | Output |
|---|
| Build the prompt set | Write the questions customers, investors, job candidates and reporters ask, including competitor comparisons and past controversies | Versioned prompt list by audience |
| Run across platforms | Use identical wording in ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews, with search on, and repeat each prompt | Dated answer log with model names |
| Score answers | Check accuracy against approved facts, plus sentiment, omissions, cited sources and differences between models | Baseline scorecard |
| Trace errors | Follow each citation to the originating page, or search distinctive phrasing when no citation appears | Error register with source URL and owner |
| Re-run | Repeat on a fixed schedule and after major news | Trend line showing whether fixes hold |
Repeat runs matter because answers vary between sessions: one response shows what a model can say, several show what it tends to say. Score what each priority audience receives, not raw mentions.
How Does a Generative Knowledge Graph Improve AI Answers?
A Generative Knowledge Graph improves AI answers by putting the same verified facts about an organization, in the same words, on its own site, in its structured data, on reference pages and in independent coverage, so retrieval tools find agreement instead of conflict.
The mechanism is statistical. A model summarizing six retrieved pages is more likely to repeat the version most share, so consistency raises the odds of an accurate answer without guaranteeing one.
| Layer | What it contains | Why it affects AI answers |
|---|
| Source-of-truth hub | A fact page with legal name, leadership, founding date, products, policies and dated corrections | Gives retrieval tools one current page to cite and gives the team a record to score answers against |
| Structured data | Schema.org Organization, Person and Product markup with sameAs links to official profiles | Google's structured data documentation says markup gives explicit clues about a page's meaning |
| Reference entries | Wikipedia, Wikidata, regulatory filings and business directories | Assistants retrieve these often, so one stale entry spreads |
| Spokesperson profiles | Executive bios, speaker pages and social profiles with matching titles and dates | Stops old bios from contradicting the leadership page |
| Earned coverage | Independent reporting from a sustained earned media strategy that repeats the proof points | Supplies third-party corroboration that owned pages cannot |
Wikipedia needs care. Its conflict-of-interest guideline discourages paid editors from changing a company's article directly and directs them to propose edits on the talk page with independent citations. In the software example, the team would update the leadership page, schema and old bios, then file a talk-page request citing the press release and trade coverage.
Seeding content to fake corroboration is astroturfing; planted reviews or undisclosed paid articles create a worse story than the original error.
How Should Comms Teams Respond When an AI Answer Is Wrong?
When an AI answer about a brand is wrong, the comms team first decides whether the fault sits with the model, with a source the model cited, or with a real problem the answer reports accurately.
What the Air Canada Chatbot Ruling Shows
In Moffatt v. Air Canada, decided in February 2024, the Civil Resolution Tribunal of British Columbia found the airline liable after its website chatbot told passenger Jake Moffatt he could claim a bereavement fare retroactively, though Air Canada's own policy page said he could not.
Air Canada argued the chatbot was a separate legal entity responsible for its own actions. Tribunal member Christopher Rivers rejected that, writing that the airline is "responsible for all the information on its website," and ordered it to pay Moffatt damages plus interest and fees.
Two lessons carry over to third-party models: conflicting versions of a fact on one website create legal and reputational exposure, and fixing the source is a far stronger public position than distancing the brand from what an AI tool said.
Match the Response to the Cause
| Cause | What to correct | Response |
|---|
| Model error: the answer contradicts accurate, available sources | The output and its corroboration | Report it through the platform's feedback tools, strengthen corroborating pages and log it; send defamatory claims to legal |
| Source error: the answer repeats an outdated or false page | The originating source | Update owned pages, request corrections from publishers and directories, and use Wikipedia's talk page |
| Underlying issue: the answer reports a real lawsuit, layoff or recall | The business problem and the public record of the response | Run standard crisis communications and publish dated facts on what changed; do not try to bury accurate reporting |
Set escalation thresholds before a crisis. Claims about safety, legal liability, financial results or named executives go to legal review and an executive owner before any public response; a wrong founding year goes into the error register. Keep an approved-facts sheet and source documents in your crisis communications plan so a correction request can go out the same day.
What Does Success Look Like for Comms Leaders?
Success in LLM reputation management means priority audiences receive accurate, current and fairly sourced answers about the organization across several assistants, and the error rate falls over time.
Measure Narrative Quality Over Mention Volume
Track these measures each audit cycle:
- Share of audited answers that state priority facts correctly
- Open high-severity errors and the days each has stayed open
- Proportion of citations from owned pages or authoritative independent sources
- Recurring misconceptions, broken out by platform
- Time from error discovery to source correction
Keep Human Judgment in the Loop
Automation can run prompts at scale, but an Altavoz analysis of reputation management in the AI era warns that handing corporate communications fully to AI produces impersonal messaging cut off from strategy. Deciding whether an answer is a model error or a fair summary of a bad quarter takes someone who knows the business.
LLM reputation work cannot rescue a weak product or an unresolved controversy; it makes the existing record easier for machines to read, for better and worse. Start with the audit, fix the sources you own this quarter, and request corrections on the rest. Teams weighing outside help, including the best PR firms in the US, should ask to see the prompt set, error register and correction log before any mention count.
