LLM Ranking Factors: We Tested the Beliefs. Here's What We Found.

The GEO Debate Has a Credibility Problem

Confident lists of "LLM ranking factors" circulate weekly — assembled from intuition, analogies to traditional SEO, and a handful of cherry-picked examples. The community knows it, even when it doesn't say it out loud.

We decided to do something about it.

Who We've Been Working With

Over the past several months, Contxt has been running a structured testing programme in collaboration with a broad cross-section of the industry: large digital agencies with active GEO practices, independent GEO and AI search specialists, and marketing and digital teams inside medium, large, and enterprise businesses actively managing AI visibility as a commercial priority. The group spans sectors, business sizes, and levels of LLM sophistication — which was the point.

Single-vertical findings are easy to dismiss. Patterns that hold across this range of contributors start to mean something. The methodology was shaped by the practitioners who've been running this alongside us — which means it reflects the questions people are actually trying to answer, not just the ones that are easy to test.

The Shared Question

The question was straightforward: can we take the beliefs the industry has coalesced around, design a test for each one, run those tests against real businesses at scale, and find out what the data actually says?

The answer is that some beliefs hold cleanly. Some hold only for specific engines. Some don't hold at all. And a few findings are genuinely counterintuitive in ways that change how we'd prioritise GEO work.

The dataset behind this is Contxt's platform: 130+ businesses, 4 LLMs (ChatGPT, Claude, Perplexity, Google AI Overview), and thousands of prompt-response pairs run through a consistent set of test types designed to isolate individual signals.

Why This Matters Now

The SEO industry did not arrive at its understanding of Google's ranking factors through a formal announcement. It built that knowledge through obsessive testing, empirical observation, and a lot of well-organised disagreement. In 2005, Moz published its first Ranking Factors survey. Practitioners voted on what they believed was moving the needle. It was rough, crowdsourced, and still more rigorous than anything that existed before it.

That process took the best part of two decades. The question for AI visibility is whether it can be compressed.

Unlike Google, which at least acknowledges the existence of ranking signals, the major AI providers publish almost nothing. OpenAI does not have a Search Central blog. Perplexity does not run Webmaster conferences. Google AI Overview behaviour is tangled up with conventional search documentation in ways that make the two hard to separate. So the signal-finding has to happen through direct testing, controlled experiments, and careful reading of the published research that does exist.

The Testing Framework

The framework currently covers 20 identified signals across four categories: Entity & Authority, Content & Structure, Technical & Crawlability, and Engagement & Credibility. These signals did not emerge from assumption. They came from three sources: published academic research, citation pattern analysis across thousands of prompt executions, and practitioner observation across dozens of industries and business sizes.

Each test targets a single signal, controls for as many variables as possible, and runs across multiple brands, multiple industries, and multiple query types simultaneously. Results are reported in plain language that marketing teams can act on without a technical background. When a test is inconclusive, that finding is published. When a signal behaves differently across models than the hypothesis predicted, the framework is updated accordingly.

What the Data Has Shown So Far

Some signals hold cleanly across all four engines. Entity salience — how distinctly a brand exists as a named entity in training data and web content — is consistently one of the strongest predictors of AI mention. This aligns with academic research confirming that entity salience correlates directly with LLM recall probability. It also has a concrete measurement mechanism, which makes it actionable.

Some signals are engine-specific. Structured data (schema markup) has a measurably stronger influence on Google AI Overview than on ChatGPT or Claude. Citation patterns differ between Perplexity, which heavily weights recently indexed pages, and Claude, which appears to weight longer-form authoritative content differently. The practical implication is that a single optimisation checklist applied across all engines is likely leaving performance on the table.

Some widely-held beliefs don't hold. The assumption that high Domain Authority automatically translates to AI citation frequency is not supported by the data at scale. Businesses with modest DA but strong entity disambiguation and consistent structured content are outperforming high-DA generalists in AI responses. This is one of the findings that changes how we'd prioritise GEO work in practice.

A few findings are genuinely counterintuitive. Content recency matters — but not uniformly. For some query types, recency is a strong positive signal. For others, the engines appear to weight established, frequently-cited older content more heavily. The pattern correlates with query intent: decision-stage queries tend to surface more established sources; discovery-stage queries are more receptive to recent content.

The Question We're Genuinely Asking

We've published the current state of the cheat sheet — 20 signals, the evidence behind each one, and the per-engine impact ratings based on the data we have so far. It is not a finished document. It is a working one.

The question we're asking the community is this: do you disagree with any of these signals or their rated impact? Are there signals you're seeing evidence for that aren't on the list? Are there test designs we should add to the programme?

The SEO industry got better by disagreeing in public. The GEO community needs to do the same thing — and it needs data to disagree with. This is ours. We're putting it out so you can push back on it.