Original insights by Ben Wills
30-second rundown
Key learnings:
- Search visibility led the broad results: Across more than 105,000 ChatGPT prompts, appearing in Google results had the strongest overall correlation with being recommended by ChatGPT, but it explained only 5.8% of the difference between brands.
- The dominant signal changes by industry: Wikidata stood out for hotels, Wikipedia citations for CRM software, and Reddit posts for enterprise AI platforms.
- Use the study as a map, not a recipe: Strengthen the AI search ranking factors that matter in your market instead of chasing one universal ranking factor.
This week: Open the report for your industry, record its three strongest signals and their sample sizes, compare your brand with five frequently recommended domains in the signal matrix, and turn the clearest credible gap into one specific improvement brief.
Coming 3 months: Set an AI Search monitoring workflow to test the same buyer questions daily, combine it with monthly rank and external-source checks, and give one owner responsibility for investigating changes. The useful result is a record showing which visibility gains repeat and which were merely desert mirages.
Ben Wills and OppAlerts dragged more than 105,000 ChatGPT prompts into a laboratory with 1.15 billion crawled pages, five billion Reddit posts and comments, 15,000-plus Google searches, 300 million Wikimedia entities, and a four-billion-link web graph. It is the sort of apparatus that should require goggles, legal counsel, and a priest.
The report asked which visible signals tend to travel with ChatGPT recommendations across 145 industries and 1,595 buyer personas. The answer is useful precisely because it refuses to become a magic button: search visibility leads overall, old-fashioned web authority still matters, and every industry hides a different animal under the floorboards.
1. Google Wore the Crown, but It Was Made of Tin
Search visibility was the strongest broad signal. Search engine appearances had a Spearman correlation of 0.241 with ChatGPT recommendation scores and explained 5.8% of the variation. Spearman correlation measures whether two things tend to rise together, while the 5.8% figure tells us the relationship leaves most of the outcome unexplained.
A brand’s best Google rank followed closely at 0.238, while links from pages that rank in search results reached 0.230. That means traditional search visibility is a serious clue, not proof that Google rankings directly cause ChatGPT recommendations. Anyone selling a single secret lever has arrived at the casino with a painted roulette wheel.
2. The Machine Recognized the Web’s Old Scaffolding
Backlink count, meaning links from other websites, produced a 0.204 correlation and explained 4.2% of the variation. PageRank, Common Crawl coverage, Wikidata records, Reddit comments, and Wikipedia citations also showed positive relationships, while homepage keyword relevance managed only 0.072 and explained 0.5%.
Broader web evidence beat homepage repetition. The machine appears more comfortable when search results, independent sites, public reference databases, and communities all leave compatible tracks. Stuffing your homepage with the right phrase is not authority; it is a man shouting his own name into an empty fuel tank.
3. Every Industry Hired a Different Informant
The global averages conceal violent local differences. For hotels and resorts, Wikidata was the leading signal at 0.650 and explained 42.3% of the variation among the 37 domains with that data; for CRM software, Wikipedia citations led at 0.577 across 37 observed domains, followed by search appearances at 0.530.
Enterprise AI platforms told another story: Reddit posts led at 0.528 across 87 domains, followed by Wikidata at 0.517. The winning signal changes with the market. Even inside an industry, the report shows different buyer personas producing different patterns, so a universal checklist can become a very efficient machine for digging the wrong hole.
These remain correlations, not controlled experiments proving that changing one signal will force ChatGPT to recommend a brand. The report itself warns that coverage varies and small samples demand caution. This is reconnaissance from a vast battlefield, not a signed confession from the model.
The honest strategy is less glamorous than the gold rush and far more likely to survive it. Build search visibility, crawlable pages, credible links, accurate public facts, and real discussion, then use your own industry data to decide where the next hour and dollar belong.
The machine did not reveal one ranking factor. It revealed a room full of witnesses, all unreliable in different ways, and left us to discover which one knows the road out.
