Ranking in LLMs: How AI Decides What Information Matters

What Is LLM SEO?

Like mainstream SEO, LLM SEO focuses on creating content that has a better chance of being found and used by AI systems when they generate an answer. That could mean a direct citation with a link or simply being the source an AI pulls a fact from without naming it.

Rand Fishkin, founder of Moz and cofounder of SparkToro, has published research that complicates the whole idea of chasing a fixed set of ranking rules. His data found that AI systems are highly inconsistent when recommending brands or products, sometimes producing a different answer to the exact same question depending on how it is phrased. 

That does not make LLM SEO pointless. It means the goal is building real authority around a topic. 

How LLMs Decide What Matters?

Traditional search engines rank pages using signals such as backlinks, keywords, page speed and hundreds of other factors before returning a list of results. AI search works differently. While each AI system uses its own retrieval, ranking, grounding and generation processes, a simplified way to understand how AI selects information is to think of it in three stages: retrieval, source authority, and structure and extraction.

Stage 1: Retrieval 

When a visitor searches a question on ChatGPT or Perplexity, the model does not scan the web in real time and rank ten results. It pulls a set of candidate sources first. This is the retrieval step. It then writes an answer using those sources. Some models lean heavily on training data. Others, such as Perplexity, search the live web for nearly every query. Get excluded at the retrieval stage and nothing you write later will matter because the model never sees your page in the first place.

Stage 2: Source Authority

After the sources are retrieved, the model favours ones that agree with each other. A claim repeated across several independent, credible sources reads as more trustworthy than an unsupported assertion, even if it is a well-written one. This is part of why brand mentions on third-party sites carry real weight. A model forming an opinion about your brand is not merely reading your own website. It is checking what everyone else has already said.

Stage 3: Structure 

The last piece is extraction. Can the model extract a clean, self-contained answer out of a passage without needing the surrounding context to make sense? Ahrefs and Semrush have found that AI Overviews extract data from the top 20 organic results for a query. Content that answers a question directly in a tight and quotable passage gets lifted out far more often than a long and meandering paragraph that buries the point three sentences in.

How to Optimise for LLMs?

There is no fixed algorithm you can reverse engineer with certainty.  Instead, we have to look at patterns from many different studies and tests which often point to similar results. Here is what you can do: 

 1. Add Statistics and Quotes 

This was one of the strongest findings from the original GEO research. A vague claim gives an LLM very little to work with. A specific figure, that too when it is clearly attributed to a reliable source, gives it something much easier to use in an answer. If a page says a strategy improved results with a clear margin, that line is far less useful to a model than one that names an actual figure and where it came from.

2. Structure Content for Extraction

Break your content into clear and self-contained sections so that it can be extracted properly. Also use headings that state the actual question being answered. Keep separate passages readable on their own, as an LLM often lifts one paragraph without the rest of the page around it. Bullet points, short definitions and direct answers near the top of a section all help here.

3. Create Topical Authority Across a Subject

One strong article rarely earns lasting citations. A Semrush study of 50,000 brands across ChatGPT found out that visibility belongs to whichever brand covers a topic thoroughly. It is not for the page which ranks highest. Comparison content, alternatives, common questions and use cases, covered properly, creates the kind of coverage models actually reward.