AI Search Visibility and Ranking Factors, July 2026

What I measured

• I covered 100 industries and 1,100 industry and persona combinations: 10 targeted personas and one neutral persona per industry.
• I sent 403,000+ prompts to 13 Claude, GPT, Gemini, DeepSeek and GLM models, and ran 150,000+ organic searches on Google and Bing.
• The models recommended 145,289 distinct domains at least once. I scored each against 18 signals.
• To measure those signals, I used 15 billion+ crawled web pages, 282 billion+ links, 3 billion+ Reddit submissions, 26 billion+ Reddit comments, 25 million+ Wikipedia articles and 120 million+ Wikidata entities.
• The May 2026 report covered 145 industries and 105,000+ prompts to one model. This time, I used fewer industries but about ten times the data.

What I found

• I did not find one ranking factor that explains the results everywhere. The signals differ by industry, and changing the buyer persona changes which businesses get recommended.
• Across all industries, search engine outbound links have the strongest Spearman correlation at 0.331. On its own, that signal explains about 11% of the variation in recommendations.
• Homepage keyword relevance was the weakest signal in May (0.072). In July, it was the second strongest (0.204).
• The correlations grew stronger for Reddit comment mentions (0.111 to 0.148) and Wikidata entities (0.120 to 0.151). Wikipedia presence weakened (0.077 to 0.055).
• Search engine appearances fell from 0.241 to 0.165, and best search engine rank fell from 0.238 to 0.148. Outbound links now have a much stronger correlation than appearances. That reverses what I found in May.
• Two of the largest changes reflect how I scored the signals. A phrase now counts 3 times in a page’s title, 2 times in its meta description and once in its body. The dataset also grew about ten times, so keep that in mind when comparing the reports.

All 18 signals

This table shows the Spearman correlation between each signal and a domain’s AI recommendation score, with all industries combined. R² tells you how much of the variation that signal explains on its own.

One industry up close: airlines

Airlines give us a useful example. This analysis includes 507 domains and 11 personas, and its strongest signals differ from the overall results.

• Search engine outbound links (0.552) and Reddit comments (0.551) are essentially tied for the strongest signal. Both Reddit signals have stronger correlations than every backlink measure.
• Delta is the most recommended airline overall. But google.com ranks fourth, because the models often suggest searching Google Flights.
• Who is asking changes the answer. Delta leads for the Business Road Warrior and the Senior Comfort Traveler. Southwest leads for the Family Vacation Planner, United for the Miles Maximizing Loyalist, and studentuniverse.com for the Student Abroad Flyer.
• The same airline domains rank differently in model recommendations, the searches models run, Google AI Overview citations and organic results. You cannot reliably use one to stand in for the others.

What this means for your marketing

• Search is one part of AI recommendations. Authority, relevance, crawlability, citations across sources and useful content still contribute to what the models have to work with. You need to measure AI visibility separately from search rankings.
• Start by defining the buyer persona. An AI visibility score means very little if you do not know who is asking.
• Measure persona coverage: how many distinct buyer contexts lead to a recommendation for your brand at all. That tells you something an overall score cannot.

What the report does not show

These figures are correlations. They do not prove why a model recommends one domain over another. Most of that happens inside the model, through training data, fine-tuning and brand familiarity that we cannot see in outside data. Improving one of these signals does not guarantee better visibility. Use the report to decide what to test in your own industry.