A recent study (shout out to @benwills) spanning 100 industries, 16 large language models, ChatGPT's web interface and Google AI Overviews found a consistent pattern across more than 20,000 observations. The brands AI systems recommend most often have backlinks and brand mentions on pages relevant to their industry, and this holds even for mentions without a link. The study's author calls this the #1 ranking factor for AI. The more useful interpretation is that AI recommendations reflect category association: how often, and how clearly, the web places a brand in the context of a specific problem or industry.
What the Study Found
The core finding is a correlation between two things:
being recommended by AI systems for queries in a given industry, and
having mentions and backlinks on pages topically relevant to that industry.
Three details make it more than a restatement of standard SEO advice:
Unlinked mentions count. A brand named on a relevant page without a hyperlink still shows the pattern. Link equity alone cannot explain the effect.
Relevance is the qualifier. The signal is tied to mentions on industry-relevant pages, not to mention volume in general.
The pattern holds across very different systems. It appears in standalone LLMs, a chat interface that retrieves live web results, and Google's AI Overviews.
Why Language Models Reward Relevant Mentions
Language models learn statistical relationships between words and entities from very large volumes of text. When a brand name repeatedly appears near the vocabulary of a category, such as its problems, use cases, competitors and buyer language, the model learns to associate that brand with the category. Later, when someone asks "what's the best tool for X?", the model draws on those learned associations.
This is why unlinked mentions matter. A model reading training text does not follow hyperlinks the way a crawler computing PageRank does. It reads sentences. "Most mid-market finance teams use Brand Y for reconciliation" teaches the association whether or not "Brand Y" is linked.
Retrieval-based systems reinforce the same pattern through a different route. When an AI interface searches the web before answering, the pages it retrieves for category queries are, by definition, category-relevant pages. Brands that appear frequently on those pages show up in the retrieved context, so the model is more likely to name them.
The model's learned knowledge and its live retrieval are both downstream of the same thing: what the web collectively says about who belongs in a category.
Why "Ranking Factor" Is the Wrong Frame
Calling relevant mentions a "ranking factor" borrows a model from traditional search, where an algorithm scores documents against weighted signals. Generative AI systems do not work like that. They do not rank brands against a published or inferred factor list. They generate answers from learned associations and retrieved context.
The distinction has practical consequences:
Ranking-factor model (traditional search)Association model (AI recommendations)Unit of value is the linkUnit of value is the descriptive sentenceAuthority flows through hyperlinksAssociation forms through repeated co-occurrence in textOptimisation targets a page's positionOptimisation targets how an entity is categorisedEffects can appear relatively quicklyEffects accumulate slowly across training cycles and the wider webOwned mainly by SEO teamsShaped by PR, product, partners, customers and community
Under the association model, a link from a high-authority but off-topic site may do little. An unlinked, specific mention in a niche industry discussion may do a lot.
Why Consistency Across 16 Models Matters
The finding that the same pattern holds across 16 LLMs and two consumer AI interfaces is arguably more important than the pattern itself. These systems differ in training data, architecture, fine-tuning and retrieval design. Their agreement suggests they converge on a shared underlying source: the web's consensus view of each category.
For businesses, this weakens the case for platform-specific "AI optimisation" tactics. If ChatGPT, Gemini-powered AI Overviews and open-source models all reflect broadly the same consensus, the higher-leverage activity is influencing that consensus, not adjusting to the quirks of each assistant.
What Most People Miss
The mechanism is the sentence, not the placement. The obvious response to this study is more link building and more digital PR, now filtered for relevance. But the evidence that unlinked mentions count suggests the content of the mention is what matters most. A useful mention clearly connects the brand to a category, a problem, or a set of peers. A placement that names a brand without that context adds little association, however authoritative the site.
The correlation runs in both directions. Category leaders get discussed on category pages because they are leaders. Part of what this study measures is the market's existing judgement appearing twice: once in the web, and again in the model trained on that web. Generating relevant mentions without the reputation to support them will likely have a weaker effect than the raw correlation implies. Relevant mentions are both a lever and a symptom.
Incumbency compounds. Brands with years of category-relevant coverage start with association density that newer entrants cannot build quickly. AI recommendations may end up more concentrated than traditional search results, where a well-optimised new page can still rank. For challengers, the realistic path is often to own a narrower category clearly rather than compete for association with a broad one.
What Businesses Should Do
Define the association you want. Decide which categories, problems and peer sets the brand should be connected with. This must be agreed across SEO, PR, content and product marketing, or each team will reinforce a different association.
Audit how the brand is described, not just where it appears. Review third-party mentions in your industry. Is the brand placed in the right category? Is it described with the language buyers use?
Prioritise industry-relevant venues over general authority. Trade publications, niche communities, comparison content, practitioner forums and partner ecosystems carry more relevant context than high-authority general sites.
Treat unlinked mentions as valuable. Stop dismissing coverage because it lacks a link. For AI visibility, the mention itself carries the signal.
Invest in the underlying reputation. Because the correlation partly reflects real leadership, the most durable strategy is to be a brand practitioners mention unprompted, and then make it easy for them to describe you accurately.
Frequently Asked Questions
Do backlinks still matter for AI search?
Yes, but more as relevant mentions that happen to carry links than as conduits of authority. The study found the pattern holds without links, which suggests the surrounding text drives most of the effect.
What counts as an "industry-relevant" page?
A page whose main subject overlaps with the category a brand wants to be recommended for. Examples include trade coverage, category comparisons, practitioner discussions and use-case content.
Does this prove mentions cause AI recommendations?
No. The study shows a strong and consistent correlation. A causal role is plausible given how language models learn, but part of the relationship likely reflects existing market leadership producing both outcomes.
Should companies optimise separately for each AI platform?
The consistency across 16 models suggests limited returns from platform-specific tactics. Shaping the web's overall description of a brand is likely to carry across systems.
Key Takeaways
AI systems tend to recommend brands that the web repeatedly mentions in an industry context, and unlinked mentions count.
The mechanism is text association, not link authority. What matters is how a brand is described, not only where it is linked.
Relevance outweighs volume. Off-topic mentions build little category association.
The correlation partly reflects existing leadership, so manufactured mentions without reputation may underperform.
Agreement across 16 models suggests AI platforms reflect a shared web consensus, which makes that consensus the real target.
The Bigger Shift
Online visibility used to be about being findable: getting a page to rank for a query. AI-assisted discovery is about being associated: being the entity a model connects to a need when someone asks for a recommendation. That association is not held on any single page. It is spread across every review, forum thread, comparison and passing mention that places a brand inside a category. As a result, visibility is no longer owned by one team. It is built by everyone whose work shapes how the market describes a company. The study's quieter conclusion, that SEO, PR, content and reputation teams must agree on who they are trying to reach and how they want to be referenced, may be its most important one. In AI search, a brand is largely what the web consistently says it is.