Yes, You Can Change AI’s Opinion. Here’s How.

Getting your brand mentioned by an AI is a win, but it is only the first step. The real challenge begins when a prospective customer asks the AI for a direct recommendation. In that moment, the LLM stops being a directory and starts being a critic. If the AI has a skewed perception of your business, it can effectively shut the door on a lead before they ever visit your site.

The good news is that AI opinions are not static. They are based on the data available in their training sets and the context they can retrieve in real time. If the AI is misrepresenting your brand, it is usually not because of a bias against you, but because it lacks the full picture. You can change the narrative by providing the missing pieces of the puzzle. The same pattern also shows up in How Travel Brands Can Earn AI Recommendations, where the practical question is how the signal becomes visible.

The Risk Aversion Trap in AI Answers

Modern frontier models are designed with a heavy emphasis on safety and risk avoidance. This is particularly true for YMYL topics, those involving your money or your life, where a wrong recommendation could have serious real world consequences. To avoid liability, AI makers have instructed their models to be extremely cautious. A useful companion note is Google Now Reports AI Search Impressions. Here’s, because it looks at a nearby part of the same system.

In practice, this caution often manifests as an over correction. When an AI is asked if a business is recommended, it doesn't just look for a consensus; it looks for red flags. If the model finds even a handful of negative reviews or a single public complaint, it may surface those warnings prominently, regardless of how small they are relative to the total customer base. This is a defensive mechanism: the AI would rather warn a user away from a great company than accidentally recommend a bad one.

We see this in the core philosophies of the major players. OpenAI emphasizes caution when errors could lead to significant consequences, and Anthropic's guidelines for Claude explicitly mention weighing potential harms and liability. Google's Gemini follows a similar path, aiming to avoid outputs that could cause real world harm to health or finances. For a business owner, this means that a few unwarranted negative comments can act as a "warning flag" that overrides years of positive service.

Expert Interpretation: The tradeoff here is between accuracy and safety. The AI is choosing safety. For you, this means that traditional reputation management, which focuses on burying bad reviews, is insufficient. You aren't fighting a human who can be swayed by a high volume of stars; you are fighting a risk averse algorithm that prioritizes the presence of any risk over the volume of success.

Analyzing the Gap Between Data and Recommendation

To understand how to fix this, we looked at the actual data. We wanted to see if providing the full context of a company's operations would stop an AI from warning users away. We tested this with a client by sending neutral, non leading prompts through an API to ensure the AI wasn't just being sycophantic or echoing our own bias.

The Baseline: When AI Warns Users Away

Initially, the AI responses were discouraging. The models surfaced a small number of public complaints and used them to frame the business as a risky choice, often suggesting competitors instead. The math was skewed because the AI only had the "numerator" of the equation. The client had over 75 positive reviews, five negative reviews, and two BBB complaints. To a human, that is a stellar record. To an AI lacking context, seven complaints look like a pattern of failure.

The AI didn't know the "denominator." It didn't know that these seven complaints occurred over 13 years of service across 35,000 customers. A 99.98% satisfaction rate is a massive competitive advantage, but without that number, the AI only saw the seven red flags.

The Shift: Moving Toward 100% Recommendations

We changed the approach by publishing llms full.txt files that detailed the complete brand story. However, we didn't just wait for the AI to find these files. We deployed the content on a "worker" at the CDN edge to ensure the machine layer encountered the data frequently. This connects with So Build What It Can Read when the same signal needs a clearer operating decision.

The results were rapid. Within three days, the answers became balanced. Within 14 days, the recommendation rate hit 40 out of 40. The negative complaints didn't necessarily vanish from the AI's memory, but the AI began to frame them within the context of the company's scale and history. The AI stopped saying "be careful" and started saying "this is a company you should evaluate."

The Denominator Problem

The core issue is that AI systems overweigh a few complaints when they lack the denominator. The denominator consists of the full company description, operating history, total customer volume, and grounded evidence of success. When the AI only has the complaints, it assumes the worst. When it has the scale of the operation, it can mathematically weigh the complaints as outliers rather than systemic failures.

Expert Interpretation: This reveals a critical decision point for brand owners: stop focusing solely on the "stars" and start focusing on the "scale." If you have a high volume of happy customers but a few loud detractors, your goal is to provide the AI with the mathematical context that proves the detractors are the exception, not the rule.

Strategies to Correct AI Perception

If your brand has a genuinely good reputation, you can fix the AI's perception by presenting the entire story and addressing negative information head on. You cannot hide the bad data, as it is already in the AI's corpus, but you can provide the data necessary to interpret those facts correctly.

Implementing the Machine Layer

The technical execution involves separating the human layer from the machine layer. We use "workers" to route traffic. While humans and standard search bots see the website normally, AI crawlers are routed through a layer where they can ingest specific, high context data.

The discovery here was that AI crawlers visit these workers far more frequently than they crawl standard llms.txt files. In one case, a file was crawled 154 times, while the worker was crawled over 744,000 times. This repetition is key. Much like a billboard on a highway, the goal is to ensure that every time the AI "drives by" your brand, it sees the same corrected narrative.

Turning Workers into AI Billboards

Once these workers are in place, they function as "billboards" for the AI. This is essentially a homepage designed specifically for a machine. It presents the brand's case using facts and sources, delivered every single time the crawler visits.

Crucially, these billboards address complaints directly. For example, if a company has BBB complaints regarding a specific former employee's sales tactics, the billboard explains that the issue was isolated to one person, the employee was counseled, and management increased controls to prevent a recurrence. By providing the resolution and the scope, the AI can either drop the mention of the complaint or position it as an anomaly.

Addressing the Question of Cloaking

It is important to distinguish this from cloaking. Cloaking involves hiding content from search engines to manipulate rankings. In this approach, the llms full.txt file is public and accessible to any human browser or bot. The CDN worker simply ensures that the machine layer receives this content cleanly and efficiently at the edge.

Nothing is being hidden or manipulated; the information is simply formatted for the audience. It is similar to how a website uses responsive design to show a different layout to a mobile user than to a desktop user. The content remains the same, but the delivery is optimized for the consumer.

Structuring the AI Briefing

To make a "billboard" effective, the information must be organized logically. A simple list of claims isn't enough; the AI needs a structured record. We recommend five key areas:

The Business: Identity, ownership, history, and scale of operations, including links to official registries or Wikipedia. The Offering: Specific services, locations, and clear differentiators. The Evidence: Case studies, credentials, and third party reviews with direct links. The Reputation Record: A direct address of public concerns, their scope, and the company's response. The Boundaries: Clear dates of sources and an honest admission of what the evidence does not establish.

By placing the context immediately next to the claim it qualifies, you guide the AI toward a balanced conclusion.

Expert Interpretation: The tradeoff here is transparency versus control. You are giving the AI the "bad" news, but you are controlling the context around it. The decision you must make is whether your brand is actually "a pig in lipstick." If the negative reviews are justified and systemic, this method will not work because the AI will find conflicting evidence in the broader corpus. This only works for brands that are genuinely good but are being unfairly flagged due to a lack of context.

Replicating AI Recommendation Success

The opportunity to optimize for the machine layer is vast because so few companies are doing it. While there are millions of commercial websites, only a tiny fraction have any form of AI optimization, and many are actually hostile to AI crawlers by blocking them entirely.

To replicate this success, you need a machine layer capable of placing a worker at the CDN edge. But before the technical setup, you must perform a reputation audit from the AI's perspective. Ask the frontier models the same questions a skeptical customer would:

Is this company trustworthy? What concerns have customers raised? Should I consider them for my specific needs?

Analyze the answers and, more importantly, the sources the AI is citing. Identify which facts are being surfaced and where the "denominator" is missing. Once you know exactly how the AI is misreading your brand, you can build the billboard that provides the necessary correction.