CORE Research: AI Search Rankings Can Be Systematically Gamed

The Black Box Cracked: CORE Research Reveals AI Ranking Vulnerabilities

AI search rankings are not impartial. A recent study by CORE researchers has demonstrated that large language models (LLMs) used in product search can be systematically influenced with a success rate of 77-82%. This is not a theoretical exercise—it is a proven methodology that businesses can deploy today to gain market share. The research exposes the black box problem at the heart of AI-driven search: the internal ranking mechanisms are opaque, but they are not immune to reverse engineering.

The study tested two primary strategies: the Query-Based Solution and the Shadow Model Solution. The Query-Based Solution, which iteratively modifies document text to observe ranking changes, achieved the high success rate. In contrast, the Shadow Model Solution—building a proxy model to predict rankings—managed only 30-34% success. For executives, the implication is clear: direct manipulation of content is far more effective than attempting to mimic the AI's internal logic.

Query-Based Solution: The 77-82% Path to Top Rankings

The Query-Based Solution treats the AI as a black box, requiring no knowledge of its internal architecture. Researchers made iterative modifications to document text and observed ranking changes. Two content expansion types were tested: reasoning-based generation and review-based generation. The effectiveness varied by AI model. GPT-4o and Claude-4 preferred reasoning-based content, while Gemini-2.5 and Grok-3 favored review-based enhancements.

This model-specific preference is a critical strategic insight. Companies can now tailor content to the dominant AI search engine in their market. For example, if a business targets users relying on GPT-4o-powered search, they should prioritize reasoning-based content—detailed, logical explanations. If the target is Gemini-2.5, review-based content—user testimonials and ratings—will yield better rankings. This granularity forces a shift from generic SEO to AI-specific optimization.

Shadow Model Solution: Predictive Power with Limits

The Shadow Model Solution involved training a smaller model (Llama-3.1-8B) to mimic GPT-4o's ranking behavior. The shadow model achieved a similarity rating of 4.5 out of 5, indicating strong predictive power. However, the optimization success rate using this method was only 30-34%. This suggests that while shadow models can predict rankings, they are poor at generating content that actually achieves top positions.

For businesses, the shadow model approach may still be valuable for testing and forecasting. It can identify which content changes are likely to improve rankings without the risk of detection. But for direct manipulation, the Query-Based Solution is far superior. Companies should invest in building shadow models for their target AI engines to guide content strategy, but rely on iterative query-based methods for execution.

Ethical and Detection Risks: The 62.1% Detection Rate

The reasoning-based optimization, while most effective, carries a detection rate of 62.1% by human raters. This means that over half of the manipulated content can be identified as artificially optimized. For brands, this poses a reputational risk. If users or regulators detect systematic manipulation, trust can be eroded. The review-based optimization raises even deeper ethical concerns: it involved generating reviews without actual product testing, essentially fabricating social proof.

Executives must weigh the short-term ranking gains against long-term brand damage. The research suggests that detection is likely, and as AI search engines evolve, they will develop countermeasures. Companies that rely on these tactics may find themselves penalized in future updates. A more sustainable approach is to use the insights to create genuinely high-quality content that aligns with AI preferences, rather than gaming the system.

Strategic Implications: A Dual-Track SEO Future

The CORE research confirms that AI search ranking optimization is becoming a dual-track game. One track is reasoning-based, favored by GPT-4o and Claude-4. The other is review-based, favored by Gemini-2.5 and Grok-3. This forces SEO strategies to become engine-specific, increasing complexity and cost. Businesses that serve multiple AI search engines will need to maintain separate content strategies for each.

Moreover, the research highlights the growing importance of generative AI in search. As more companies adopt AI-powered product search, the ability to influence rankings will become a competitive differentiator. Early adopters of these techniques will capture market share, while laggards will see their visibility decline. The window for action is narrow: as detection methods improve, the effectiveness of these tactics will diminish.

Who Gains, Who Loses

The winners are companies that invest in query-based optimization for their target AI engines. They will achieve top rankings and drive traffic and sales. The losers are those that rely on traditional SEO or shadow model approaches, which are less effective. Also at risk are ethical content creators who produce genuine reviews; they may be crowded out by fabricated review-based content that ranks higher.

AI search engine providers themselves face a dilemma. They must balance the user experience against the integrity of their rankings. If manipulation becomes widespread, user trust will decline, potentially driving users back to traditional search engines. This could trigger a arms race between manipulators and search engines, with frequent algorithm updates to counter new tactics.

Outlook: What to Watch in the Next 30 Days

In the near term, expect a surge in content optimization campaigns targeting GPT-4o and Claude-4 with reasoning-based content. Companies that move quickly will gain an edge. However, watch for announcements from AI search providers about new detection algorithms or ranking changes. The 62.1% detection rate suggests that countermeasures are already feasible. Additionally, regulatory scrutiny may increase, particularly around fabricated reviews. Businesses should monitor FTC guidelines and consider the legal risks of review-based optimization.

For executives, the bottom line is clear: AI search rankings are now a manipulable variable. The CORE research provides a blueprint for action, but it also carries risks. The strategic choice is whether to exploit the vulnerability for short-term gain or to build sustainable content strategies that align with AI preferences without crossing ethical lines. The next 30 days will reveal which path the market takes.