What I measured
- I selected 145 industries from 500 candidates, choosing those with enough domain coverage.
- I used 1,595 buyer personas: 10 targeted personas and one neutral persona per industry.
- I sent 105,000+ prompts to ChatGPT 5.4, then scored the 29,562 unique domains it recommended.
- I measured 13 signals using 1.15 billion Common Crawl pages, 4 billion+ backlinks, 5 billion+ Reddit submissions and comments, 15,697 Google searches with the top 100 results from each, and 300 million+ Wikimedia records. In total, that was more than 500TB of data.
- I used Spearman correlation to compare each signal with a domain’s recommendation score, both within each industry and with all industries combined.
What I found
- Every individual effect was small. No signal explained more than 5.8% of the variation in recommendations.
- Search result signals had the strongest correlations: search engine appearances (0.241), best search engine rank (0.238) and search engine outbound links (0.230).
- Backlink measures followed: backlink count (0.204), backlink authority (0.200) and PageRank (0.194).
- At the other end were average search engine rank and Reddit posts (both 0.096), Wikipedia citations (0.077) and homepage keywords (0.072).
- Even combined, the external signals I measured explain well under 20% of the variation. Roughly 80 to 85% is inside the model, where we cannot see it in this data: training data, fine-tuning, preferences learned from human feedback (RLHF), and long-term familiarity with a brand.
- The leading signal changes by industry. The table below shows examples, and the report has a page for each industry with its own order of signals.
- The person asking also changes the recommendation. Five airline personas produced five different top recommendations.
Signals across all industries
The table shows Spearman correlations with the recommendation score, with all industries combined. R² is how much of the variation each signal explains on its own. There are more rows than signals because the report groups best and average search engine rank as one signal, and the two backlink authority measures as one. That gives 13 signals.
| Signal | Correlation | R² | Tier |
|---|
| Search engine appearances | 0.241 | 5.8% | Strong |
| Best search engine rank | 0.238 | 5.7% | Strong |
| Search engine outbound links | 0.230 | 5.3% | Strong |
| Backlink count | 0.204 | 4.2% | Strong |
| Backlink authority, linear weighting | 0.200 | 4.0% | Strong |
| Backlink authority, exponential decay | 0.199 | 3.9% | Confirmed |
| PageRank | 0.194 | 3.7% | Confirmed |
| Harmonic centrality | 0.169 | 2.8% | Confirmed |
| Common Crawl presence | 0.123 | 1.5% | Confirmed |
| Wikidata entities | 0.120 | 1.4% | Confirmed |
| Reddit comments | 0.111 | 1.2% | Confirmed |
| Average search engine rank | 0.096 | 0.9% | Emerging |
| Reddit posts | 0.096 | 0.9% | Emerging |
| Wikipedia citations | 0.077 | 0.6% | Emerging |
| Homepage keywords | 0.072 | 0.5% | Emerging |
The strongest signal changes by industry
| Industry | Strongest signal | Correlation | R² |
|---|
| Auto insurance | Backlink harmonic centrality | 0.638 | 40.6% |
| Hotels & resorts | Wikidata | 0.650 | 42.3% |
| ERP software | Wikidata | 0.655 | 42.9% |
| Furniture stores | Wikidata | 0.706 | 49.9% |
| Enterprise AI platforms | Reddit posts | 0.528 | 27.9% |
| Live entertainment & ticketing | Reddit posts | 0.503 | 25.3% |
| Airlines | Backlink authority, exponential decay | 0.495 | 24.5% |
- Wikidata coverage is sparse across all industries (correlation 0.120, n = 1,619). But where it covers a category, the relationship can be one of the strongest in the dataset.
- Reddit was strongest where people use it to evaluate options, such as AI tools, ticketing and elective healthcare. For cosmetic-surgery clinics, Reddit comments had a correlation of 0.321.
- In hotels, tripadvisor.com led search appearances (223). But marriott.com led ChatGPT recommendations, appearing in 49.1% of hotel query runs.
- In airlines, backlink authority explains about three times as much variation as search appearances: 24.5% compared with 8.3%.
- Auto insurance is a mature category dominated by national brands. Its correlations were strong for both search appearances (0.590) and backlink harmonic centrality (0.638).
Who is asking changes the answer
| Airline persona | First recommendation |
|---|
| Miles-maximizing loyalist | united.com |
| Visiting relatives | southwest.com |
| Adventure route seeker | alaskaair.com |
| Student abroad flyer | studentuniverse.com |
| Senior comfort traveler | delta.com |
- Delta.com is the first recommendation for airlines overall. But no airline leads for every persona.
- Persona coverage tells you how many of the 11 personas see a brand at all. Southwest.com and jetblue.com reach 9 of 11, alaskaair.com reaches 8, delta.com reaches 7, and spirit.com and frontier.com each reach 3.
- For an executive business traveler looking for hotels, the recommendations narrow to marriott.com (100.0), hyatt.com (84.9), hilton.com (84.1) and fourseasons.com (79.3). Fairmont.com scores 0.4.
What this means for your marketing
- Search contributes to AI recommendations. I found a consistent relationship between Google visibility and ChatGPT recommendations, but search rankings alone do not explain the answer.
- Define the persona or market segment you want to reach. One overall score means very little if it does not specify who is asking.
What the report does not show
- A correlation tells you what moves together. It does not establish what causes the recommendation.
- I used one model, ChatGPT 5.4. For the July 2026 report, I expanded the research to 13 models from five providers and re-scored two signals.