Reputation Management in the AI Era: How LLMs Judge Your Brand

LaCleo published a playbook for online reputation management in the AI era in March 2026, covering review velocity, multi-platform monitoring and response times. This article is narrower and goes deeper on one part of it: the assistants themselves, how they decide what to say, and what the evidence published since March says about getting it checked and corrected.

What is AI reputation management?

AI reputation management is the practice of monitoring how AI assistants describe, rank and recommend a brand, tracing negative or incorrect statements to the sources the assistants retrieved them from, correcting those sources, and re-checking on a schedule. It inherits the goals of online reputation management and changes the unit of work. ORM manages a list of places a buyer might look. AI reputation management manages a single synthesized answer that draws on all of those places, is written differently for each buyer, each country and each run, and is read by someone who will usually not open the pages behind it.

Three things follow. The monitoring unit is the prompt, not the platform: "best accounting software for a UK charity" is the thing being tracked, and the review sites, forums and comparison pages are inputs. The metric is position, sentiment and accuracy inside the answer rather than a star rating. And the fix lands at the cited source, which for most comparison questions is a page the brand does not own. The article on how AI engines choose what to cite covers the retrieval mechanics; this one is about what those mechanics do to a brand's standing.

How an LLM forms its view of a brand

An assistant's view of a brand comes from two layers, and a reputation program has to work on both.

Layer one: what the model learned

Every assistant carries a picture of well-known brands from its training data, and that picture shapes answers even when the assistant could look things up. A Cornell Tech study published on 20 October 2026 queried 102 models across 42 product categories and found that the brand a model recommends is often not the brand it names as the market leader: the two differed in 26 of 42 categories, and a model's own named leader was its recommendation only 31% of the time. The preference appears in base models before any fine-tuning and is strengthened by preference training. In a spot check, Claude Opus 5.5 answered brand questions from memory and never searched, with web search available. Ahrefs' October 2026 fan-out study puts a number on how often ChatGPT retrieves at all: about 14.4% of free, logged-out responses contained search queries.

For a brand, this layer is slow to change and not directly editable. What can be done is to make sure the next training snapshot contains accurate, consistent, widely replicated facts, which is the same work as layer two done early.

Layer two: what the model reads when asked

When an assistant does retrieve, it reads whatever is currently published and ranks it by its own rules. Ahrefs' test of 20,000 US prompts found the mix depends on the question. For factual questions ("does X integrate with Y"), 63.4% of citations went to official websites. For comparison questions ("X vs Y", "which is cheaper"), 46.4% went to official sites and the rest to third parties: review platforms, forums, hands-on testers such as RTINGS and RepairPal, and news. Those third parties are where a reputation forms and where it has to be managed.

The mix also moves. In the same Ahrefs data, Reddit went from being cited in about one in five US ChatGPT answers in March 2026 to about one in 43 by late September, a fall of 88%; Wikipedia fell 80%; government, research and manufacturer support pages rose. A brand that fixed its Reddit presence in spring and stopped checking would have been managing the wrong source by autumn. The Reddit citation article on this blog describes the spring picture; this is the correction.

You are always compared with someone

A buyer rarely asks an assistant about one brand in isolation, and when they do the assistant usually volunteers alternatives. The practical consequence is that reputation in AI is relative: the question is not whether the description of you is positive but whether it is more useful than the description of the competitor named next to you.

Mike Sonders, writing for Search Engine Land, ran twelve B2B software prompts 100 times each through logged-out ChatGPT from different IP addresses. A single answer named about ten brands. Across 100 answers to the same prompt, about 44 different brands appeared. Only about five, roughly 11% of the pool, appeared in 80% or more of runs; in competitive categories 72% of mentioned brands appeared in fewer than 20% of answers. Two things follow for reputation work. First, a brand can be "in" ChatGPT and still be absent from nine answers in ten, which is where the buyer went to a competitor without anyone noticing. Second, a single check tells you almost nothing; the author's recommendation is at least five runs per prompt per period, and the AI visibility tools that check once a week are measuring noise.

The comparison is also local. Two countries share fewer than one in ten cited domains for the same prompt on most engines, so the competitor named beside you in Germany is not the one named in the US. The multi-country tracking article has the figures; the reputation implication is that a brand with a clean picture in its home market can carry a stale or negative one abroad and not know.

How wrong the answers are

Enough to lose deals, and more often for smaller companies. Searchable, an AI visibility vendor, published a review in July 2026 of 32,556 facts that AI assistants had stated about 1,704 brands, each graded correct or incorrect by employees of the brand concerned. 9.2% were wrong. 75% of brands that reviewed 20 or more facts found at least one error. The worst categories were the ones buyers act on: pricing 13.5% (16.1% for US pricing), funding and financials 10.9%, integrations 10.6%, partnerships and clients 9.9%, reviews and ratings 8.8%. Companies with 50 or fewer staff saw a 9.4% error rate against 6.3% for large companies. By engine, Perplexity (11.5%) and Google AI Overviews (11.4%) were worse than ChatGPT (8.1%).

Error rate of AI-stated brand facts by category, July 2026 (Searchable, vendor data, n=32,556 facts)

Self-selected sample of Searchable customers, graded by brand employees; "incorrect" includes outdated facts and mistaken identity as well as fabrication. Direction, not precision.

Two features of this data matter for how reputation work is organised. "Incorrect" was mostly not invention; it was once-true information that nobody updated, and brands confused with others. Both are fixable at the source. And the error rate is highest where the brand has the least published, consistent information, which is why smaller companies fare worse: the assistant fills the gap with whatever third party wrote about them.

Why speed matters

Three clocks run at once, and the brand controls none of them.

1. The buyer's clock. Bain's December 2024 survey found about 80% of search users rely on AI summaries in at least 40% of searches and about 60% of searches end without a click. BrightLocal's 2026 survey found 23% of consumers would decide on an AI review summary alone, 77% are less likely to use a business after negative reviews, and 19% expect a reply to a review the same day, up from 6% a year earlier. A buyer who reads an unfavourable or wrong answer acts on it that session. There is no follow-up visit in which to recover.
2. The source clock. Reddit's share of ChatGPT citations fell 88% in six months; publishing partners' share doubled between March and July then fell back. The pages that shape your description this quarter are not the ones that shaped it last quarter, so a fix made against last quarter's sources may be aimed at pages the assistant no longer reads.
3. The competitor's clock. Responsive's November 2025 survey of 350 B2B buyers found 61% start with a preferred vendor but 45% are open to switching, and 80% of technology buyers use generative AI at least as much as search to research vendors. A competitor who corrects their facts, publishes the comparison page and answers the forum thread is the one the assistant describes accurately, and Searchable's report puts the outcome plainly: the buyer goes to "whichever competitor happened to be described accurately".

The operational conclusion is a weekly cadence rather than a quarterly audit, with every engine and every market in the set, and a fix pipeline that starts the day a wrong or negative answer is found.

How to audit your AI reputation

1. Write the prompts buyers use, not the ones you would like them to use. Four kinds: the category recommendation ("best X for Y"), the direct comparison ("you vs competitor"), the fact check ("X pricing", "does X integrate with Y", "X offices") and the judgement ("is X any good", "X reviews", "X problems"). At least 100, written separately for each language and country you sell in, since the answer and the competitor set change at the border.
2. Run each prompt several times on every engine. ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity, and Claude or Copilot where your buyers use them. Five runs per prompt per period is the floor that the repeated-runs data supports.
3. Grade every answer on four things. Mentioned or not; position relative to the competitors named; sentiment of the description (recommended, neutral, caveated, warned against); and the accuracy of each stated fact about pricing, products, locations, integrations and partners.
4. Trace each negative or wrong claim to its source. Open the citations. For comparison and judgement prompts most will be third-party pages: a review profile, a forum thread, a directory listing, a competitor's comparison page, a news item. Record the URL, the claim, and whether it is stale, mistaken or a live complaint.
5. Fix at the source and re-check. Your own pages first, then the third-party pages, then publish what is missing. Re-run the affected prompts weekly until the answer moves, then keep them in the tracked set.

This can be done by hand for a handful of prompts on one engine. Across 100 or more prompts, five engines, several countries and five runs each, it is a tracking system. LaCleo GEO runs this loop as a platform: prompts tracked across the major LLMs and 37 geographies, each answer graded for mention, position against competitors and sentiment, with the cited sources attached and the specific fix recommended for each gap. It is available on a free trial, and the audit steps above are what it automates.

What to fix first

Priority order for AI reputation fixes, by where the assistant is reading

How we do it

LaCleo runs AI reputation monitoring inside its agentic reputation management and agentic SEO and GEO services, on our own platform, with a named person reviewing every deliverable.

1. Prompt set. 100 or more prompts per market across the four prompt types above, written in the buyer's language and reviewed by the client, run on ChatGPT, Google AI Overviews, Gemini and Perplexity, with Claude on request, from each of the client's countries on LaCleo GEO.
2. Grading. Every answer is scored for mention, position against the named competitors, sentiment and factual accuracy, with the cited sources captured, so a negative or wrong statement comes with the page it came from.
3. Competitor view. The same prompts show which competitors are recommended instead, how they are described, and which sources earned them the position, which is the gap list.
4. Fixes. Each gap gets a specific recommended fix: the page to correct, the listing to claim, the thread to answer, the comparison or pricing page to publish, with the schema and copy delivered as an implementation pack the client's developers paste in. Nothing to install.
5. Cadence. Weekly re-runs, with on-demand re-checks after a fix ships, and a monthly report of mention rate, position, sentiment and accuracy by engine and country, next to referral sessions from AI sources.

Statistics with sources

Figures used in this article, with the source and the caveat that applies

Frequently asked questions

What is AI reputation management?

Monitoring how AI assistants such as ChatGPT, Gemini, Perplexity and Google AI Overviews describe, rank and recommend a brand, tracing negative or incorrect statements to the sources the assistants retrieved them from, correcting those sources, and re-checking on a schedule. It extends review and search reputation management to a channel where the assistant writes one answer for the buyer and the brand does not get to reply.

How do LLMs decide what to say about a brand?

From two places. Training data sets a baseline view and, for well-known brands, a preference that persists: across 102 models, the recommended brand differed from the named category leader in 26 of 42 categories. When the assistant retrieves live, it reads what is currently published; for comparison questions, Ahrefs found 54% of ChatGPT's citations went to sites other than the brands' own.

Why does speed matter in AI reputation management?

The buyer gets one answer, acts on it, and the brand never sees the lost enquiry. 45% of US consumers used AI for business recommendations in 2026, 19% expect a review reply the same day, and a typical recommendation answer names about ten brands. A wrong price or stale negative thread is read by every buyer who asks until the source is corrected, and the sources themselves change inside months: Reddit's share of US ChatGPT citations fell 88% between March and September 2026.

How accurate are AI assistants about companies?

Imperfect. In a July 2026 review of 32,556 brand facts graded by the brands themselves, 9.2% were wrong, with pricing worst at 13.5%. 75% of brands that reviewed 20 or more facts found at least one error, small companies fared worse than large (9.4% against 6.3%), and Perplexity and Google AI Overviews were worse than ChatGPT. The data is vendor-collected and self-selected, so treat it as indicative.

Can I fix what ChatGPT says about my company?

Not directly, but you can change what it reads. Correct and expand the facts on your own site, which carry 46% to 63% of citations depending on question type; update directories, review profiles and partner pages; respond to and resolve the review and forum threads the assistant cites; publish the comparison and pricing pages buyers ask about. Then re-run the prompts to confirm the answer moved.

How is this different from traditional online reputation management?

Traditional ORM manages a list of places a buyer might look. AI reputation management manages one synthesized answer that draws on all of them, varies between runs and countries, and is read by a buyer who will not open the underlying pages. The monitoring unit is the prompt, the metric is position and sentiment in the answer, and the fix is at the source the assistant cited.

How often should AI answers about a brand be checked?

Weekly, with several runs per prompt and every engine and country in the set, plus an on-demand re-check after each fix ships. The repeated-runs data shows a single check is close to meaningless, and the six-month collapse of Reddit's citation share shows a quarterly audit can be aimed at sources the assistant has stopped reading.

Working with LaCleo

LaCleo runs agentic reputation management and agentic SEO and GEO as human-led managed services, with the monitoring on LaCleo GEO, our own platform. A program includes 100 or more tracked prompts per market across the major LLMs and 37 geographies, sentiment and position scoring against named competitors, the cited sources behind every negative or incorrect statement, Reddit and community citation tracking, a full competitor comparison and gap analysis, and a specific fix for each gap delivered as an implementation pack the client's developers paste in. Agentic SEO and GEO starts at $750 a month with tool licenses included. The platform is available on a free trial for teams that want to run the checks themselves. A named person reviews every deliverable, and there is nothing to install.

Connect to know more Book a call Get a free AI audit

Sources

BrightLocal, "Local Consumer Review Survey 2026", n=1,002 US adults. Bain & Company, "Goodbye Clicks, Hello AI: Zero-Click Search Redefines Marketing", February 19, 2025. Responsive, "Inside the Buyer's Mind: What Shapes B2B Decisions Today", November 11, 2025, as reported by Demand Gen Report. Louise Linehan, Ahrefs, "How ChatGPT searches: 66.6M responses and 15M fan-out queries", October 12, 2026, updated October 22, 2026. Mike Sonders, Search Engine Land, "What repeated ChatGPT runs reveal about brand visibility". Searchable, "The AI Brand Misinformation Report", updated July 30, 2026 (vendor research). Tapan Parikh, Cornell Tech, "How Language Models Learn Taste: Brand Preference in LLMs", arXiv preprint, October 20, 2026. All figures verified October 2026.

Two sources are published by vendors of AI visibility software (Searchable) or RFP software (Responsive) and are labelled as such. The repeated-runs and fan-out studies cover ChatGPT's free tier only. The Cornell Tech paper is a preprint and has not been peer reviewed.