Being mentioned by an AI assistant is only half the win. The harder question is what happens when a prospect asks the model to recommend a vendor. According to a new GEO study from AnswerShare, published on Search Engine Journal, frontier models often respond by surfacing every complaint they can find and steering the user elsewhere.
Why AI models read reviews without context
AnswerShare points to policies from OpenAI, Anthropic and Google that push models to be cautious in high-stakes categories. In practice, that caution can become an overcorrection: the model has the numerator—a handful of complaints—but not the denominator of total customers, years in business and how the company responded.
In one client test, the business had 75 positive reviews, five negative reviews and two Better Business Bureau complaints. Against 20 customers that would look risky. Against 35,000 customers over 13 years, those seven complaints imply a satisfaction rate of about 99.98%.
From warning to recommendation
The team asked the same neutral reputation prompts before and after adding context. Before, the AI framed the company as risky and named competitors. After publishing an llms-full.txt file and serving it through an edge worker, answers began to include operating history and scale.
After 14 days, the result was 40 out of 40 AI recommendations. Some clients saw sentiment shift in as little as three days and stabilize in about two weeks. AnswerShare reports a regional ad agency flipped in 15 days; a Phoenix/Scottsdale luxury resort in 10.
What to feed the machine layer
The method is less about hiding negatives and more about giving AI the same full record a careful reviewer would want. AnswerShare’s “billboard” record covers five areas:
- The business: legal identity, ownership, history, customers served
- The offering: services, locations, buying information, differentiators
- The evidence: credentials, case studies, reviews, third-party sources
- The reputation record: concerns, scope, company response, context
- The boundaries: dates, attribution, unresolved questions
Each concern should be described with its source, date, the customer relationship involved and the company response. Scale and context belong next to the claim they qualify, not in a separate section.
Is this cloaking?
AnswerShare argues no: the llms-full.txt file is public and byte-for-byte identical whether served to humans, Googlebot or GPTBot. The edge worker only routes it cleanly to crawlers, similar to responsive design or markdown conversion.
Still, the tactic raises a fair debate for marketers: when does audience-specific delivery become cloaking? The line AnswerShare draws is transparency—nothing is hidden, and all sources are checkable.
What marketers should do next
Start by asking the questions a buyer would ask an AI: Is this company trustworthy? What concerns exist? Should I consider it for my needs? Save the AI’s answers and sources, then build your own llms-full.txt from supported information. Put the denominator next to every complaint. Do not try to game the model; AnswerShare says AI will read manipulation fast and penalize “marketing slop.”
If the brand story cannot be supported, no amount of optimization will fix a bad reputation. But if the record is genuinely strong, the full context can move the model from critic to advocate.
Source: Search Engine Journal
