Researchers Reverse-Engineered AI Search Rankings. Here's What They Found.

For the past year, the advice on getting your content surfaced by AI has been part intuition, part extrapolation from traditional SEO, and part educated guessing. Nobody had actually run a controlled experiment to test which specific content changes improve your chances of ranking inside an AI response.

Until now.

Key Takeaway: Query-based content — writing that directly answers a specific question with clear reasoning and a position — outperformed keyword-optimized content 77-82% of the time across four major AI models. For SMEs, the implication is straightforward: stop writing for keywords and start writing for questions.

Written by Derek Chua, digital marketing consultant and founder of Magnified Technologies. Derek runs AI content systems in production for Singapore SMEs — this is what he's seeing in practice.

Researchers published new findings under the label CORE (Content Optimization for Retrieval and Evaluation), testing which content optimizations reliably move the needle when AI language models rank sources in response to queries. The models tested include Claude 4, Gemini 2.5, GPT-4o, and Grok-3. The results are specific enough to be actionable, and surprising enough to be worth paying attention to.

The caveat up front: this research was conducted via API testing, not live consumer AI interfaces. How AI models respond in a research environment may differ from how they behave when real users are asking questions through ChatGPT, Perplexity, or Claude.ai. That gap matters, and we will come back to it. But the directional findings are still the clearest signal we have on what actually moves the needle.

What the Research Tested

The CORE team reverse-engineered the ranking logic of four leading AI models by presenting them with queries and competing documents, then systematically varying the content characteristics of those documents.

The question they were answering: if you give an AI model a pool of candidate sources to draw from when answering a question, which content attributes predict whether a given source gets cited?

The finding that matters most for anyone producing content: query-based optimization produced top-ranked results 77 to 82 percent of the time across all four models tested.

That number deserves unpacking.

What "Query-Based Optimization" Actually Means

Traditional SEO optimization is about keyword placement: get the keyword in the title, the first paragraph, the subheadings, the meta description. The logic is that search algorithms pattern-match on those signals to determine relevance.

Query-based optimization is different. Instead of placing keywords, you directly address the question the user is asking, including the sub-questions behind it, the context that informed the question, and the reasoning the user would use to evaluate an answer.

If someone asks an AI assistant "what CRM should a 10-person Singapore marketing agency use," a keyword-optimized article might hit "CRM Singapore" and "marketing agency CRM" throughout its text. A query-optimized article actually answers that specific question: here are the options that work at that team size, here is what matters at that stage, here is what to expect in terms of cost and setup time for each.

The distinction sounds subtle. In practice, it requires a different approach to structuring content. You are not writing to rank for a keyword. You are writing to be the definitive answer to a specific question a real person has.

The Finding That Should Change How You Write

The CORE research also found that reasoning-based content outperformed content that relied on keyword repetition, and that review-based content performed particularly well across AI retrieval tasks.

The reasoning finding makes sense when you think about how AI language models work. These systems were trained on human-generated content. They have developed a sense of what good reasoning looks like: claims supported by evidence, conclusions that follow from premises, acknowledgment of counterarguments, specific examples that ground abstract claims. Content that thinks clearly reads as credible to these models, in the same way it reads as credible to an intelligent human reader.

The review-based finding is more interesting. Content that includes genuine evaluation, comparison, and judgment based on experience surfaced more reliably than content that described features or explained concepts without taking a position. This is not the same as inserting an opinion for its own sake. It means that content demonstrating judgment, formed by someone who clearly understands the subject, carries more retrieval weight than neutral description.

For SME marketing content, this has a direct implication: articles that tell you what to do and why, based on an opinionated view of what works, perform better in AI retrieval than articles that present all options without committing to a recommendation. The "here are six things to consider" structure is less useful than "here is what we would do in your situation, and here is why."

At Magnified, we have been applying this framing with clients for months before this research landed. The pattern holds: the articles that get cited by AI tools are the ones where Derek has taken a position and explained the reasoning, not the ones that enumerate options and say "it depends."

What Underperformed

The research also identified what did not help, and some of these findings cut against common content marketing assumptions.

Keyword density had low predictive value. Stuffing an article with primary and secondary keyword variations did not reliably improve AI retrieval rankings. This is worth noting because a significant chunk of content created for SEO is still optimised for keyword frequency first. That approach is becoming a liability, not an asset.

Generic comprehensiveness underperformed specific depth. Articles that covered a topic broadly, touching on everything without going deep on anything, ranked worse than articles that narrowed their scope and went genuinely deep on a specific question. Being the complete guide to everything about your category is less valuable than being the definitive answer to one specific question your customer asks repeatedly.

Authoritative-sounding language without substance didn't help. Content that used confident, professional language without backing it up with specific evidence, examples, or reasoning ranked similarly to lower-quality content. The signal AI models are detecting is real substance, not the appearance of it.

The Caveat That Matters

The research was conducted via API calls to these AI models, in a controlled environment where documents were presented directly for evaluation. This is not the same as how these models behave in production, when a user asks a question through a consumer interface and the model draws from a combination of its training data, live web retrieval (where available), and its own inference.

The gap between API-tested rankings and actual consumer-facing AI search behavior is real, and honest reporting on this research requires acknowledging it. We do not know with certainty that a document scoring well in the CORE experimental framework will surface more reliably when a real user asks a related question through ChatGPT or Claude.ai.

What we can say is that the directional findings align with what practitioners are observing in practice: content with genuine reasoning, specific examples, and clear positions on contested questions tends to be cited by AI systems more frequently than content that hedges everything and describes rather than evaluates.

The research confirms a direction. It doesn't certify an outcome.

What This Changes for Your Business

For businesses that have invested in SEO and are now trying to extend that investment to AI search visibility, the CORE findings suggest some practical adjustments.

Shift your content brief from keyword-led to question-led. Before writing, articulate the specific question this article answers. Not the keyword it targets. The question. "What accounting software should a GST-registered SME use if they want InvoiceNow compliance without switching from their current system?" is a question. "Accounting software Singapore" is a keyword. These produce very different articles.

Build in positions, not just descriptions. Every article should take a stance on something. Not a manufactured controversy for its own sake, but a genuine judgment based on what you know. If you have run Google Ads for F&B businesses for three years, you have views on what works at what budget level, which industries benefit most, and what mistakes cost clients the most money. Write those views into the article, with the reasoning behind them. That is the kind of content the CORE research suggests AI models treat as source-worthy.

Write for the sub-question, not just the headline question. When someone asks an AI about a topic, the model synthesises an answer by drawing from content that addresses the main question and the questions nested inside it. An article about "how to set up Google Analytics for your SME website" that also addresses "what events to track from day one," "how to connect Search Console," and "what reports to actually look at" covers the sub-questions that a user follows up with. Articles structured this way are more likely to be retrieved across a range of related queries, not just the primary keyword.

Review your existing high-value pages. Apply the CORE lens: does each article directly answer a specific question? Does it reason through that question or just describe it? Does it take a position? These are useful diagnostic questions for an existing content audit, not just for new content production.

How This Fits the Broader Shift

The CORE findings land in the middle of a broader transition we have been tracking for months. AI systems are absorbing search behaviour. ChatGPT is growing as a Google alternative faster than TikTok ever did. AI crawlers are indexing the web with different goals than traditional search bots.

Each of those developments raises the same underlying question: what kind of content earns a place in AI-generated answers, versus what gets left out?

The CORE research is the first attempt to answer that question empirically, across multiple models, with controlled testing. The answer is not "produce more content" or "add more keywords." It is closer to: write with the specificity, reasoning, and judgment that an expert would bring to an explanation for a smart, time-pressed business owner.

Which, to be fair, is what good content should have been all along. The difference is that AI retrieval is now applying a harder filter than keyword matching did. And that filter is sorting the good from the average more efficiently than a Google ranking would.

---

Frequently Asked Questions

What is query-based optimization and how is it different from keyword SEO? Query-based optimization means structuring content to directly answer a specific question a user might ask, including the reasoning and context behind the answer. Keyword SEO focuses on placing target phrases in strategic locations. The CORE research found query-based content ranked first in AI retrieval 77-82% of the time across four models — keyword density had low predictive value by comparison.

Does this research apply to all AI tools, or just specific ones? The CORE study tested Claude 4, Gemini 2.5, GPT-4o, and Grok-3 via API. The findings were consistent across all four, which suggests the pattern reflects something fundamental about how large language models evaluate source quality — not a quirk of one platform. That said, the research was conducted in a controlled API environment, not through live consumer interfaces, so real-world results may vary.

What type of content performs best in AI search rankings? The CORE research found three content types consistently outperformed others: query-optimized content (directly answering a specific question), reasoning-based content (claims supported by evidence and clear logic), and review-based content (genuine evaluation and judgment from experience). Generic, comprehensive overviews and keyword-dense articles underperformed.

Should we stop doing keyword research for SEO? No. Keyword research still matters for traditional search and helps you understand what questions your audience is asking. The shift is in how you use that research: rather than optimizing for keyword placement, use keywords to identify the underlying questions your content should answer directly. The two approaches can coexist — one informs your topic selection, the other shapes how you write.

How do we apply this to existing content? Run a quick audit on your top-performing pages using three questions: does this article directly answer a specific question? Does it reason through that question, or just describe it? Does it take a clear position? Pages that fail all three are candidates for a rewrite or expansion. Pages that pass at least two are likely already in reasonable shape for AI retrieval.

---