A researcher ran 403,000 prompts through 10 different large language models across roughly 100 industries, including several local search categories, then scored which business attributes most often produced mentions inside the generated answers. Ben Wills published the analysis as an attempt to map the inputs behind AI recommendations. The single category dissected in the published write-up is legal services for businesses, and the factor with the highest correlation is also the most familiar one to anyone who runs technical SEO audits: a page-one presence in Google.<\/p>\n
The prompt pool spans 100 industries, with a deliberate skew toward local search verticals. That scope matters for site owners because the same prompts an LLM might receive for a personal injury lawyer in Cleveland are also being generated for plumbers, dentists, real estate agents, and HVAC companies. The legal category serves as the worked example in the published write-up, but the methodology is built to surface signals that travel across verticals.<\/p>\n
Each prompt was paired with a structured evaluation of the responding model. The output was scored for whether the model named a specific business, and if so, which attributes that business shared with other frequently named competitors. The analysis is correlational rather than causal, a point worth keeping in mind before any audit checklist gets built around it.<\/p>\n
For the legal services for businesses segment, the attributes most tightly tied to LLM mentions ranked in this order:<\/p>\n
The page-one Google correlation is the headline number. A business that ranks anywhere in the top ten organic slots appears far more often inside LLM answers than a business that ranks on page two or beyond, regardless of how polished its knowledge panel or schema markup looks in isolation.<\/p>\n
Before touching anything else, pull a clean rank report for the queries your customers actually type. Track both head terms and long-tail variations, including local modifiers. If your domain does not appear on page one for the prompts that matter, the rest of the audit is downstream work. Log which competitors own those slots, because their pages are the references the model is most likely pulling from when it composes an answer.<\/p>\n
Open your homepage with a fresh browser, ignore the design, and read the visible text. Does the H1, the first paragraph, and the navigation all reinforce the primary service and the geographic area you serve? If a crawler or a language model had to summarize your homepage in a single sentence, would the summary match what a searcher asked for? The study ranks homepage relevance ahead of backlink metrics, which suggests the model is reading the page itself, not just weighing links.<\/p>\n
Domain authority is not a Google metric, but the referring domain count, the ratio of branded to generic anchors, and the toxicity of inbound links all feed the same underlying signal. For local sites, focus on links from local news outlets, chamber of commerce directories, professional associations, and supplier pages. These pass the kind of corroborating context a model looks for when deciding whether to name a brand.<\/p>\n
Wikidata is the structured-data backbone that a surprising number of language models consult for entity resolution. Search your exact legal business name on wikidata.org. If a record exists, check that the official website field, the industry field, and the location field all match your current reality. If no record exists, Wikidata’s notability bar is lower than Wikipedia’s, and a verified business with a public address and a real-world footprint can usually qualify. Keep in mind that edits go through a community review process, so plan for a few weeks of lead time.<\/p>\n
Search site:reddit.com for the service plus city combinations you target. Note the recurring subreddits. Look at how the top replies handle recommendations: do they name specific providers, or do they describe how to evaluate one? If real buyers post in those threads and your brand never appears, that is a measurable visibility gap. Genuine participation, not astroturfing, is what the data points toward.<\/p>\n
Because the prompt pool includes local search verticals, the findings generalize. Service area businesses, multi-location operators, and single-location shops all sit inside the same correlation surface. Homepage relevance, entity consistency on Wikidata, and Reddit participation are all within reach for a small team with no enterprise budget. Link earning and Wikidata listing take longer to move, but they reinforce the same identity signal a model is looking for when it has to choose between naming your business or naming a competitor.<\/p>\n
The study measures association, not cause. A page-one Google ranking may correlate with AI mentions because both draw on the same authority and entity signals, or because the models were trained on web snapshots that already reflected Google’s ordering. Either explanation points the same way for an audit: the levers that drive traditional rankings also drive AI recommendations. Treating the two as separate problems is a mistake the data does not support.<\/p>\n
Ben Wills published the analysis after running 403,000 prompts through 10 different large language models across 100 industries, with a focus on local search categories. The legal services for businesses segment is the worked example in the published write-up.<\/p>\n
Appearing on Google page one for the target query had the highest correlation with being named inside LLM answers in the legal services for businesses category, ahead of homepage relevance, domain authority, Wikidata presence, and Reddit mentions.<\/p>\n
Start with a clean rank report for your target queries, then audit whether your homepage copy, your Wikidata record, your backlink profile, and your presence in relevant Reddit threads each reinforce the same business identity. The page-one SERP check is the gating item because every other signal stacks on top of it.<\/p>\n