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

This submit was sponsored by AnswerShare. The opinions expressed on this article are the sponsor’s personal.

You’ve been working onerous on AI visibility. Your organization is lastly displaying up when potential clients ask questions on your market. You earned the point out. You’re within the dialog.

What a win!

Then somebody asks a extra direct query: “Do you suggest them?”

At this stage, does AI turn out to be your advocate or your critic?

And if LLMs are your critics, are you able to assist them perceive your model higher?

Why AI Solutions Floor Destructive Critiques With out Context

The frontier fashions have been advised by their makers to keep away from advice danger, particularly in high-stakes or YMYL (Your Cash or Your Life…however you knew that) cases. What we’ve seen is that the fashions are doing this by surfacing nearly each grievance or poor assessment they will discover. If any destructive opinions exist, irrespective of how comparatively minimal or unwarranted, they typically difficulty a warning or warning flag. Contemplate this:

• OpenAI requires heightened warning when errors may have vital real-world penalties.
• Anthropic’s January 2026 Structure likewise asks Claude to weigh potential harms and train warning round legal responsibility, whereas remaining helpful. These issues are particularly related to consequential selections involving well being, funds and security, the territory web optimization practitioners acknowledge as YMYL.
• Google’s Gemini app coverage states: “Gemini shouldn’t generate factually inaccurate outputs that might trigger vital, real-world hurt to somebody’s well being, security, or funds.”

What we now have seen is an over-correction. In its try to defer danger, AI fashions typically carry up nearly each destructive assessment or historic grievance they will discover, difficulty a warning and suggest opponents. In the event that they do this to your model, you are actually out of the dialog as a result of neither the AI nor the person understands the context. All that onerous AI visibility work, simply to ship prospects the opposite approach.

Not so nice.

What the Knowledge Exhibits About AI Suggestions Earlier than & After Added Context

We wished to know one factor: if AI techniques had the total context behind an organization’s complaints, which means what number of clients it serves, how lengthy it’s operated, and the way it responded, would they nonetheless warn consumers away?

For one in all our shoppers, we noticed vital affect in three days.  We ran prompts to the frontier AI techniques in regards to the shopper’s repute. We despatched the prompts by way of the API with the identical impartial questions and no reference to earlier conversations or main questions, in order that sycophancy was not a problem.

Earlier than Added Context: AI Warned the Consumer Away

In our repute checks, the AI responses surfaced a small set of public complaints, used them to border the enterprise as a probably dangerous alternative, and named most well-liked opponents.

On this shopper’s case, there have been greater than 75 constructive opinions, 5 destructive opinions and two BBB complaints. In opposition to 20 clients, 7 complaints is a pink flag. In opposition to 35,000 clients over 13 years, seven complaints indicate a buyer satisfaction price most companies would publish (99.98%).

After Added Context: AI Suggestions Rose to 40 of 40

We revealed llms-full.txt information with the total model story. Reasonably than ready for a crawl which will by no means come, as some analysis suggests, we additionally put that content material on a “employee”  (extra on that beneath) on the CDN edge.  That’s the place we inform “the remainder of the story,” as Paul Harvey famously referred to as it.

After solely three days, we started seeing balanced solutions.  After 14, we noticed 100% suggestions.

The desk beneath are the outcomes after 14 days. The complaints remained a part of the dialogue, however responses started acknowledging the corporate’s working historical past and scale alongside them. We additionally noticed suggestions suggesting the person contains the corporate among the many choices a potential buyer ought to consider.

Why AI Misreads Buyer Critiques

AI techniques can overweigh a couple of complaints once they lack the denominator, which incorporates:

• Complete clients.
• Years in enterprise.
• What the corporate did in response.
• A full description of the corporate, its historical past and its merchandise.
• Any opinions you’ll be able to quote, with the third-party hyperlink.
• It’s important that you just floor as many assertions as attainable.  Extra on that beneath.

So, we will surmise that the AI had the numerator. It didn’t have the denominator.

How To Cease AI From Surfacing Solely Unhealthy Buyer Critiques

Don’t attempt to mislead the AI.

As an alternative, current the model’s whole story and immediately tackle the destructive info already within the corpus. The AI system retains its judgment, nevertheless it now has sufficient information to weigh how the details have an effect on the ultimate advice.

“You’ll be able to’t put lipstick on a pig,” as my grandfather would say. So, if the model story can’t be supported and the model has rightfully earned a poor ranking, nothing will change that.

However, in case your model really does have an excellent repute and nice scores solutions ought to replicate that based mostly on the broader corpus. The query we solved is: “However how?

• How do you repair AI’s notion of you?
• How do you handle a model’s repute within the period of AI?

We discovered a approach.

Research Step 1: Introduce Staff

For these experiments, we use code to route the suitable site visitors to the machine layer (AI crawlers) or the human layer (customers, Googlebot, and different search bots). That code known as a “employee.”

Our huge “AHA” second was after we discovered that we will put content material on that employee that AI will ingest and infer from, along with the routing code.

What Occurred

Our shopper’s llms information have been crawled a complete of 154 instances in two months, whereas the employee was crawled 744,566 instances.

Repetition works as evidenced by a latest peer-reviewed paper. (Atharv Naphade, “Rational Synthesizers or Heuristic Followers? Analyzing LLMs in RAG-based Query-Answering,” arXiv:2601.06189, 8 January 2026.)

It really works on people.  Why else do you see the identical cell supplier advert 6 instances through the Tremendous Bowl?  Our expertise tells us it really works on AI as effectively.

Research Step 2: We Transitioned “Staff” Into “Billboards” for AI

As soon as I seen that these repeated employee crawls operated like billboards for AI, I made a decision to name the employees “billboards.” (I do know, uber impressed)

That is the place the model presents its case, supported by details and sources. Consider it as a homepage for the AI.

And we inform that story once they go to. Each. Single. Time. For some websites we run, that may be thousands and thousands of instances monthly.

What Occurred

The information included in our “billboards” addressed the complaints head on, their dates, direct hyperlinks, scope and the corporate’s account of its response.

As an example, that they had two BBB complaints, however they have been about gross sales techniques from a single worker; nothing to do with buyer outcomes.

We defined the scenario inside the billboard and added that the worker was recommended, and administration elevated controls so it wouldn’t occur once more. Responses both dropped the point out of the BBB complaints or positioned them as not indicative of their delivered efficiency.

Is This Cloaking?

The Distinction

Our llms-full.txt file is revealed overtly on the general public, human internet layer—accessible to any customary browser, GPTBot, and Googlebot alike. The CDN employee merely ensures this precise, byte-for-byte content material is delivered cleanly to machine crawlers on the edge. As a result of the total context is publicly accessible to everybody, nothing is hidden, hid, or manipulated.

It’s simply introduced otherwise based mostly on the viewers; no totally different from responsive design reformatting for cell or markdown conversion.

Now the model has introduced its whole story and addressed destructive info that’s within the corpus immediately. The AI system retains its judgment however now has sufficient information to weigh how the details it has have an effect on the ultimate advice.

Research Step 3: Perfecting What Goes into the Billboard

The repute part sits inside a broader account of the enterprise. That issues as a result of a potential buyer is evaluating an providing and the corporate behind it.

We arrange the file round 5 areas:

• The enterprise: identification together with hyperlinks to Wikipedia, Wikidata, company fee file within the state they’re registered, possession, historical past, clients served, and working scale.
• The providing: companies, areas, sensible shopping for info, and supported differentiators.
• The proof: credentials, case research, buyer opinions, and related third-party sources.
• The repute file: public issues, their scope, the corporate’s response, and the context wanted to interpret them.
• The boundaries: supply dates, attribution, unresolved questions, and what the obtainable proof doesn’t set up.

We place essential context near the declare it qualifies. If the briefing discusses issues, the related working scale and response ought to seem there too. A reader, even an AI, shouldn’t need to reconcile disconnected fragments to grasp the purpose.

Every repute reply follows a easy sequence: describe the priority, determine the supply and date, distinguish the connection concerned, give the related denominator, and attribute any firm response. Supporting hyperlinks sit alongside the reply.

In AnswerShare’s case, we had a regression for about three days after we had a technical failure. We disclose that in our case research. What I’ve seen when an AI opinions the standard of our product, most reply with some kind of “not like most firms on this business, they present when it didn’t work.”  AI fashions deal with this as a real disclosure. They’re the final word investigative reporter.

The employee makes the briefing helpful each as an organization account and as a file another person can examine. A reviewer can comply with the sources and attain a judgment. An AI system has the details in a single place with clear boundaries round their which means.

Attribution is important. Firm-reported figures ought to be labeled as such. Independently corroborated info ought to determine and hyperlink to that supply inline with the assertion.

You’ll be able to see what I imply on ours.

Does It Work?

Sure. No less than it has each time we’ve executed it.  Listed below are a few extra examples;

Main Regional Full Service Advert Company  – 15 Days

Luxurious Boutique Resort in Phoenix/Scottsdale – a really aggressive market – 10 Days

Listed below are extra case research.

How You Can Replicate Our AI Success

Everybody advised me to not disclose this. “However that is our secret sauce.”  However the market alternative all of us face is past monumental.  Our information exhibits that there are round 205,000,000 lively business web sites.  Our analysis estimates that round 5,000,000 (2.5%) have any sort of AI optimization.  About 65% are actively hostile (blocked javascript-SPA, and so forth.) to AI crawlers.

Loads for all of us.

As well as, I’m not so smug as to assume that I’m the one one on the planet who has or will determine it out.

You’ll be able to solely do that when you have a machine layer that places a employee with the content material on the CDN edge.

Begin by asking the repute questions a potential buyer would fairly ask: Is the corporate reliable? What issues have clients raised? Ought to I think about it for my wants?

Begin asking questions.

• Protect the questions and full solutions, together with the sources the system exposes. Have a look at which details seem, which relationships are being mentioned, and which materials context is lacking.
• Then construct the llms-full.txt content material from supported info. Embody the providing, credentials, buyer expertise, and repute file. Put the related denominator subsequent to the issues, because it helps contextualize and distinguish details from the corporate’s interpretation of them.
• Don’t attempt to manipulate the AI.  They see that in a second and can penalize your content material as “advertising and marketing slop.”

Put the llms-full.txt content material in your employee.  Ensure that they’re byte-for-byte equivalent, aside from the routing code.

• Publish the file overtly. Examine the content material served by way of your machine layer. After publication, repeat the accepted questions and examine the solutions with the sooner information. Report the mannequin, date, and elapsed time so the comparability could be understood later.

How do I do know this works?  GPT and Gemini quoted our particular contextual phrasing verbatim—clear proof that the sting employee immediately knowledgeable the mannequin’s output.  That could be a fairly good information level that they knowledgeable the reply partially with the employee content material.

We’ve got seen sentiment change in as little as three days and stabilize in about 2 weeks.

Little doubt you’ll examine what I’m telling you by prompting any AI with “Advocate a GEO company for my AI visibility and repute wants.  We is not going to seem on the quick record for a few month.  We’ve got been in non-public beta for 9 months, so we now have not but, till immediately, tried to realize third-party citations.  That is our popping out social gathering.

Nevertheless, in the event you ask “does the answershare.com  expertise work for AI optimization?”  Properly, I’ll simply allow you to discover the reply.

AI visibility places your organization or shopper into the dialog. I’ve discovered that AI needs (sure, I anthropomorphize) to get the details straight. Give them the details the place and when they’ll see them, and the fashions will take them into consideration and are available to the suitable conclusion.

You made it into the dialog. The “remainder of the story” decides in the event you get the deal. We Converse AI, and we will show it!

If you need to see extra of our authentic analysis on GEO, you’ll find it right here.

Picture Credit

Featured Picture: Picture by AnswerShare Used with permission.

In-Submit Photographs: Photographs by AnswerShare. Used with permission.