TLDR
AI answers are far easier to influence than most people realize. New research shows that persuasive, longer content often wins in AI generated responses even when it adds no real value. This creates short term opportunity and long term risk. We are watching the beginning of an AI optimization arms race that mirrors early SEO. When platforms step in, shortcuts will collapse. Clarity, authority, and trust will outlast tactics.

If You Care About AI Visibility, Start Here

You Can Game AI Answers. That’s the Problem.

A lot of marketers are asking the wrong question right now. They keep asking how to get AI to mention their brand. The better question is why it is so easy to influence AI answers in the first place. In a recent article published on Search Engine Journal, growth strategist Kevin Indig lays out an uncomfortable truth. AI answers are not just influenceable. They are fragile by design. And that fragility is the real risk.

This breakdown translates Kevin’s analysis into plain language, connects it to real business consequences, and explains why this moment should feel familiar to anyone who remembers the early days of SEO chaos.

Why AI Answers Feel Unstable by Design

Traditional search engines are boring in a good way. You type a query. Google runs a formula. You get roughly the same results every time.AI does not work like that.

Large language models generate answers instead of retrieving them. Every response is probabilistic. Same prompt. Different outcome. Think less calculator and more jazz improvisation. Kevin outlines several reasons this makes AI answers unstable. Ask the same question multiple times and you may see completely different brands appear. In real testing, only about 20 percent of brands show up consistently across repeated prompts. Models also carry built in bias from their training data. Some lean heavily on Wikipedia. Others pull more from Reddit. That bias exists before your content ever enters the picture. Understanding how to work with these systems — not against them — is exactly what our 2026 AEO & GEO Guide covers, from platform-by-platform citation mechanics to schema and E-E-A-T signals.

Models also evolve rapidly. What worked on older versions of ChatGPT may not work on newer ones. Personalization adds another layer of volatility. Two people can ask the same question and receive different answers based on context or data access. Longer prompts narrow the possible answers, which sounds safer, but still does not eliminate manipulation. Yes, AI answers can be influenced. They are just inconsistent, unstable, and impossible to fully control.

The Research That Changed the Conversation

If instability was not concerning enough, new research shows how easy AI systems are to manipulate when someone actually tries.

A paper from Columbia University titled “E-GEO: A Testbed for Generative Engine Optimization in E-Commerce” tested this directly. Researchers built a massive simulation using over 7,000 real shopping queries sourced from Reddit and more than 50,000 Amazon product listings. They created a controlled environment with three roles. One AI acted as the seller whose job was to rewrite product descriptions to win visibility. Another AI acted as the shopping assistant that ranked products based on the query. The rest of the products kept their original descriptions and served as the control group. The seller AI was allowed to iteratively improve its descriptions based on feedback. Think of it as a persuasive copywriter that never gets tired and never questions ethics.

The Universal Strategy That Should Make Everyone Nervous

The results were eye opening.

The biggest finding was the existence of a universal strategy for AI visibility in ecommerce. Despite common belief, AI did not reward concise, factual descriptions. It consistently favored longer descriptions with a persuasive tone and inflated language. No new facts. Just better sounding words.

These rewritten descriptions beat the original ones nearly 90 percent of the time. Even more concerning, the strategy worked across categories. A method developed using home goods performed almost as well in electronics and apparel. That means sellers do not need deep expertise. They just need the right prompt and enough iterations.

In plain terms, AI shopping results can be dominated by whoever sounds the most convincing, not whoever sells the best product. If that feels familiar, it should.

This Is Not an Isolated Finding

Kevin points out that this research is part of a growing body of evidence.

Earlier studies on Generative Engine Optimization showed that adding statistics and citations could increase AI visibility by roughly 40 percent. Other research demonstrated that inserting structured data sequences into product pages could manipulate AI recommendations. Another study showed that directly instructing AI within the content itself could influence rankings. The common thread is fragility. Small stylistic changes can move a product from invisible to number one without changing its actual usefulness. Different models have different weaknesses, making large scale protection difficult.

Why This Is a Business Problem, Not a Tech Curiosity

This might sound academic. It is not. If AI driven search and shopping become flooded with bloated, persuasive content optimized for machines instead of humans, trust erodes. When trust erodes, platforms intervene. We have seen this movie before. Content farms worked until they did not. Keyword stuffing worked until Panda. Link schemes worked until Penguin. Each phase ended with a crackdown. Kevin argues that we are watching the opening moves of that same cycle, just powered by AI instead of keywords.

The Ironic Catch

There is a quiet irony in all of this. The people documenting these vulnerabilities are also making them easier to exploit. Kevin openly acknowledges that tension. He frames his work as a warning, not a how to guide. The hope is that LLM developers take action before manipulation becomes the default behavior. Whether that happens remains to be seen.

What This Means for Real Businesses Right Now

Yes, AI answers can be influenced. No, that does not mean chasing tricks is smart. Probabilistic systems punish short term thinking. What works today can vanish tomorrow. And the more aggressively a tactic is abused, the faster platforms respond. Businesses that build real authority, clear positioning, and genuinely helpful content still have the advantage. When cleanup comes, it always favors substance over shortcuts. AI may improvise. Platforms do not forgive.

The Rakos Take

This is not a call to panic. It is a call to perspective. If your growth strategy depends on gaming unstable systems, you do not have a strategy. You have a temporary loophole.

AI is exposing something marketing has avoided for years. Influence without integrity collapses under scale. The brands that win will not be the loudest. They will be the clearest. And clarity remains stubbornly hard to fake.

FAQs

Can AI search results really be manipulated?
Yes. Multiple studies show that AI responses can be influenced through writing style, structure, and persuasive language without changing factual accuracy.

Does this mean SEO is dead?
No. It means shortcuts are fragile. Long term visibility still favors authority, clarity, and trust across multiple channels.

Should businesses optimize specifically for AI answers?
They should optimize for humans first. Clear positioning, strong content, and real expertise translate better across both search and AI systems.

Will platforms stop this manipulation?
History suggests yes. As abuse scales, platforms introduce safeguards that punish artificial inflation.

What is the safest long term strategy?
Build assets that answer real questions clearly, consistently, and credibly. That foundation survives algorithm shifts.