Get Expanded Industry Research

This post covers the all-industry numbers and one industry example. The live platform has this same breakdown, plus over/under-performer tables, fanout queries, and persona pages, for all 100 industries, refreshed roughly every one to three weeks as new link, crawl, and Reddit data comes in.

Coupon FOUNDERS25 at checkout.

Need this built for your industry specifically?

I’m taking on custom work directly: custom personas, competitor sets, or an industry that isn’t in the 100 covered here. I’m limiting this to 5 clients at a time. Expected turnaround is 4 to 8 weeks per project, the goal is one month, but give it room to run longer.

This Is Correlation, Not Causation

Every number in this report describes what moves together. None of them prove what causes an LLM to recommend one domain over another.

Every ρ in this report is a Spearman correlation. It measures how strongly a signal and LLM visibility move together across thousands of domains, not whether one causes the other. A domain that scores well on a signal tends to also get recommended more, for reasons this report cannot fully isolate from the outside.

That distinction has a real consequence: improving a signal that correlates strongly with visibility is not a guarantee that your own visibility improves. It means you’re moving in the direction that’s associated with getting recommended more often, not pulling a lever with a known, guaranteed effect.

The single strongest signal we measure, Search Engine Outbound Links, explains about 11% of the variance in recommendation behavior on its own (R² = 11%). Even accounting for all 18 signals together, most of what determines an LLM’s recommendation happens inside the model, in places external data can’t see: training data composition, fine-tuning, RLHF preferences, and brand familiarity built up over years. We can measure what correlates with getting recommended. We cannot see the causal machinery that actually produces the recommendation.

Treat this report as a map of what to test, not a guarantee of what to fix. If you improve a signal that correlates strongly with visibility and nothing changes, that’s not a contradiction. That’s correlation and causation working exactly as different things.

How To Use This Report

Turn the data into sales conversations, client priorities, and the next three to six months of work.

The benefit is focus. Instead of generic AI visibility tactics, you get the exact industry, the exact persona, and the signals that actually correlate with getting recommended in that market.

The Largest LLM Ranking Factors Report, To Date

I believe this is the largest LLM ranking factors analysis published anywhere. If something bigger or more comprehensive exists, let me know.

• 100 industries, each scored against its own vocabulary and its own buyer language.
• 1,100 industry/persona combinations: 10 targeted personas plus one neutral persona per industry.
• 403K+ prompts run across ten models: Claude, GPT, Gemini, DeepSeek, and GLM.
• 145,289 distinct domains recommended by an LLM at least once, scored against 18 external signals.
• 150K+ organic searches, across Google and Bing.

Compared to our own May 2026 release

The industry count went down. Everything else went up, by a lot: this edition samples 10 models instead of one, analyzes 6x the raw data, and the underlying dataset is roughly 10x larger overall.

The Scale

The research required fast code, not a giant cloud budget: 3 petabytes of data analyzed, up from 500TB in May.

Despite the size, this pipeline runs without a large cloud spend. That’s a result of years spent on fast parsing and a hand-written, high-throughput HTML/text extraction pipeline. Changing how a signal is scored, as this edition did for two of them, is a recompute measured in minutes once the raw scans exist.

Research Process

How the recommendation dataset was built.

• 100 industries, each broken into 10 targeted buyer personas plus one neutral, industry-wide persona.
• Every persona gets its own weighted vocabulary: 110,221 distinct phrases across 1,200 industry/persona segments, averaging 2.7 segments per phrase.
• 352,000 tracked hostnames, mapped to the industries they compete in.
• LLM recommendation runs across 10 models from 5 providers, plus the web searches those models issue on their own while answering (fanout).
• Every recommended domain is cross-referenced against search results, homepages, web crawl data, Reddit, Wikipedia, Wikidata, and backlink graph data.

How Signals Are Scored

Every correlation in this report is a Spearman ρ between one signal and a domain’s LLM recommendation score.

• Rank-based: only the ordering of domains matters, not the scale of the raw numbers.
• R² (ρ squared) is the share of rank variation that signal explains on its own.
• A domain missing a signal is excluded from that signal’s correlation. It is never scored as zero.
• Every signal is scored globally, per industry, and per persona, using that scope’s own vocabulary. The same domain can have a different score for the same signal in three different tables.

Improved Keyword Relevance Scoring

Search Engine Outbound Links and Homepage Keyword Relevance use an improved scoring approach this edition: instead of a flat keyword match, a phrase now counts 3x if it appears in a page’s title tag, 2x in its meta description, and 1x in the visible body. A title tag holds a handful of words, so what a page spends those words on is a stronger signal of what the page is actually about than a body-text mention is. This more accurate scoring is a large part of why both signals moved so much since May, see Why Search Signals Jumped.

How The Rankings Moved

All 14 signals we can directly compare across both editions. ▲ teal = strengthened since May. ▼ red = weakened.

Rows marked with a methodology note compare May’s single blended backlink metric against July’s domain-level equivalent; July also reports a separate host-level variant of each (see the full table below). Two of these rows also changed because of the improved scoring described above, not just dataset size.

Why The Rankings Moved

Two structural changes explain most of it, visualized below: every signal’s change in ρ since May, sorted worst to best.

The dataset grew roughly 10x since May, which tightens every correlation by reducing noise, and two signals were rescored with the title/meta/body weighting described above. Both are real, and together they explain the movement.

All 18 Signals, Ranked

All 18 signals, ranked by ρ against LLM visibility. Color marks the tier.

These are the all-industry pooled tier boundaries. Individual industry pages (see the Airlines example below) use the same signals but their own tier cutoffs relative to that industry’s own distribution.

Full July 2026 Data Table

The complete correlation table, all-industry pooled, for readers who want every number.

ρ = Spearman correlation vs. LLM visibility score. R² = share of variance explained by this signal alone. Coverage = % of sampled domains with a non-null value for this signal.

What’s Actually Driving This

The 18 signals fall into four natural families. Here’s how each one is trending.

Search Engine Signals

Search Engine Outbound Links, Search Engine Appearances, Best Search Engine Rank, Average Search Engine Rank. Being cited (outbound links) now outperforms merely appearing (appearances) by a wide margin, the reverse of May. Part of that is the title/meta/body scoring change described above, not just market movement. Rank position within results (average rank, best rank) matters less than either.

Backlink & Authority Signals

PageRank, Harmonic Centrality, and Backlink Count, each split into domain-level and host-level variants. Nine of the eighteen signals in this table are backlink-graph metrics. Domain-level cuts beat host-level cuts on every single pair, the split itself is the finding.

Content & Relevance Signals

Homepage Keyword Relevance. The one on-page signal we measure, and the biggest riser in the whole study. A domain’s homepage now needs to actually be topically relevant, not just technically indexed.

Community & Knowledge Graph Signals

Reddit Comment Mentions, Reddit Submission Mentions, Wikidata Entities, Common Crawl Presence, Wikipedia Presence. Reddit and Wikidata both strengthened. Wikipedia article presence, on its own, weakened: structured data and live discussion appear to be more useful to LLMs right now than encyclopedic prose alone.

Why 100 Industries, Not 145

This edition covers 100 industries, down from 145 in May. 45 were cut, 0 were added, the new list is a strict subset of the old one.

The 45 industries removed since May 2026

Every Industry In This Report

Click any industry to jump straight to its live report.

Airlines

A concrete walkthrough of what a single industry report looks like, end to end.

Airlines is a useful example because the brands are ones almost everyone recognizes, and because the industry shows nearly every finding in this report at once: a search-visibility mismatch, a persona-driven recommendation set, and a wide gap between LLM recommendations and Google’s AI Overviews.

The following sections use this one industry to show what the full paid report looks like: the signal correlations, the cross-channel comparison, the over/under-performer analysis, the literal search queries models run, and how much the answer changes by persona. Every number below is pulled from the live report site, the same thing a paying visitor sees.

What Predicts Airline Recommendations?

Every top-15 signal reads “Dominant” on this industry’s own scale, more than usual is explained by what we measure.

Reddit is essentially tied with search citation as the top predictor, both Reddit signals outrank every backlink metric. Source: oppalerts.com/AI-Search-Visibility/airlines/.

Four Channels, Four Different Winners

The same airline domains, ranked four different ways: what LLMs recommend directly, what LLMs search for while answering (fanout), what shows up in Google’s AI Overviews, and plain organic search rank. Each column is a 0–100 percentile within that channel, a dash means the domain never appeared there at all.

By average organic rank, YouTube and Travelocity lead the airlines category, not an airline. By LLM recommendation, Delta leads outright. Google’s AI Overviews cite a mix of airline sites, aggregators, and ranking authorities that barely overlaps with either list. Each channel measures something different, none of them is a reliable proxy for the other three.

Who AI Actually Recommends

Recommendation score is a weighted reciprocal-rank sum across every model and run, scaled by each model’s influence weight.

Delta, Southwest, United, and Alaska lead, in line with conventional US market share. The outlier: google.com ranks #4, ahead of American Airlines, models frequently recommend “search Google Flights” as the answer itself, not just an airline brand. Full ranked list (25 domains) available in the paid report.

Where LLM Recommendations Diverge

Domains ranked by percentile within each channel. Over-performers rank far better in pure LLM recommendations than across fanout, AI Overviews, and organic combined. Under-performers are the mirror case.

Generic platforms and aggregator/content sites dominate organic results and AI Overview citations but barely register as direct LLM recommendations. Niche, persona-specific travel brands overperform in pure LLM recommendations relative to their weaker footprint everywhere else.

What The Models Search For

When a model needs the web to answer, it writes and runs its own search queries. These are captured verbatim.

Notice how many of these queries target a specific third-party authority by name, Skytrax, OAG, The Points Guy, NerdWallet, J.D. Power, AARP, not the airline’s own site. Ranking well with the review sites and industry-authority lists your buyers’ models actually search for matters as much as your own homepage.

Same Industry. Different Buyer, Different Answer.

The industry-wide #1 recommended airline is Delta. No single airline is #1 for every persona.

The Student Abroad Flyer result is worth calling out: it’s a student-fare booking site, not an airline, and it was the #1 recommended domain for this persona in our May 2026 report too. That’s a real, repeated finding, not noise: a persona built around a specific need pulls in a specialist brand that never shows up as a top recommendation for anyone else asking about airlines.

This is the central measurement problem this whole report keeps coming back to: a single aggregate “who’s #1 in airlines” number hides five different correct answers, depending on who’s asking.

Key Findings From The Report

The data doesn’t point to one universal ranking factor. Industries lean on different evidence layers, and persona context changes the recommendation set entirely.

1. Search engine outbound links are now the single strongest predictor across all industries, ρ = 0.331, up from rank #3 in May, partly a real shift and partly a scoring change.
2. Homepage relevance went from the weakest signal we measured to a top-2 signal, a 3× increase, for the same reason.
3. Backlink authority got more granular this edition (domain-level vs. host-level) and mostly held its ground, the extra detail is additive, not a correction.
4. In Airlines specifically, Reddit discussion is essentially tied with search citation as the top predictor, ahead of every backlink metric.
5. A brand’s LLM recommendation score can diverge sharply from its organic rank and its AI Overview citations, in either direction.
6. The literal searches models run while answering name specific third-party authorities far more often than they name the brand’s own site.
7. No single brand is the top recommendation across every buyer persona in an industry, even a category as consolidated as US airlines splits five different ways by persona.
8. Persona coverage, how many distinct buyer segments recommend you at all, is a different and arguably more useful number than one aggregate visibility score.

Why Search Signals Jumped

Two of the biggest movers since May jumped because we improved how we score them.

In May, a keyword match anywhere on a page counted the same whether it was in the title tag or buried in the footer. In July, a phrase match counts 3x if it’s in the page’s title, 2x if it’s in the meta description, and 1x if it’s in the visible body. The reasoning: a title tag holds a handful of words, so what a page spends those words on is a much stronger signal of what the page is actually about than a body-text mention is.

That improved scoring explains a large part of why SE Outbound Links jumped from rank #3 (ρ = 0.230) to rank #1 (ρ = 0.331), and why Homepage Relevance jumped from the weakest signal in the whole May table (ρ = 0.072) to a top-2 signal (ρ = 0.204), a 3× increase.

The dataset also grew roughly 10x since May, which independently tightens every correlation by reducing noise. Both the better scoring and the bigger dataset are real, measurable improvements, and together they explain the jump.

Persona Coverage Is A New Kind Of Market Share

A brand’s AI visibility is a distribution across personas, not one score.

Persona coverage measures how many distinct buyer contexts can see you at all, not just your single best-case score. For a brand that competes across multiple segments, broad persona coverage may matter more than one high aggregate number that’s actually being carried by a single strong segment.

What Conclusions Can We Draw?

The research doesn’t produce one magic ranking factor. It produces a measurement framework.

• Search visibility matters, but it explains part of the picture, not the whole thing.
• Backlink authority, Reddit discussion, Wikipedia and Wikidata presence, and homepage relevance vary dramatically by industry, Airlines leans on Reddit and search citation, other industries lean elsewhere.
• The best signal for one industry can be minor in another.
• The most useful unit of measurement isn’t the keyword. It’s the persona.

That means the right question isn’t “what’s the LLM ranking factor?” It’s “for this industry, for this persona, what actually distinguishes the domains that get recommended?”

SEO Principles Still Apply

Search isn’t dead. It’s one layer of AI recommendations.

The data shows a consistent relationship between search visibility and LLM recommendations, and models are already searching the web live while answering. That makes SEO fundamentals more important, not less: authority, relevance, crawlability, citation breadth, and useful content still feed the evidence layer LLMs draw from.

The mistake is treating search rank as a stand-in for AI visibility. It’s one signal, not the answer.

Personas Change The Measurement

You can’t monitor AI visibility like a single search-ranking report.

In SEO, personalization was a modifier on top of a keyword rank you could still track. With LLMs, the persona and the context of who’s asking are the center of the recommendation, not a modifier on it. Airlines alone splits five different ways by persona in this report.

A single aggregate score isn’t a meaningful thing to monitor. If you’re tracking AI visibility without specifying who’s asking, you’re measuring noise.

Methodology & Dataset Scale

Every correlation in this report is drawn from a single pooled, all-industry dataset rebuilt from scratch for July 2026.

Models Sampled

LLM visibility scores are pooled across every model listed above, sampled against 403K+ prompts spanning 1,100 buyer personas across 100 industries. Backlink, search, and content signals are cross-referenced against nearly 15 billion crawled web pages, the full history of Reddit, all of Wikipedia, and 282B+ links.

How To Work With Me

Custom analysis for industries, portfolios, agency clients, link prospecting, and reputation monitoring. I’m taking on 5 clients at a time, expected turnaround is 4 to 8 weeks per project.

• Industry-specific AI visibility analysis: define industries and personas, get your sign-off, run the prompts, collect the data, deliver the reports.
• Multi-client agency research: repeat the process across several large clients or related industries.
• Large-scale link prospecting: Common Crawl, backlink graph data, and industry relevance scoring to prioritize outreach targets.
• Large-scale reputation monitoring: Reddit, search results, Wikipedia/Wikidata, web crawl data, and backlink signals combined to monitor category reputation at scale.
• Custom high-throughput data pipelines: process large datasets quickly without turning the project into a large cloud spend.

Fixed costs depend on scope and data sources, but the engineering approach is built to keep variable processing cost low, this project analyzed 3PB of data without a large server bill.

Get The Full Picture For Your Industry

This post is the free, all-industry summary plus one industry example. The live platform has this same depth of report, per industry, for all 100 industries, refreshed roughly every one to three weeks.

Want this built custom for an industry that isn’t in the 100, or want it done as a consulting engagement? I’m taking 3 to 5 clients at a time, 4 to 8 weeks per project. OppAlerts.com/contact/

SEO & Engineering Background

Building the data pipeline, running the analysis, and turning it into strategy.

This project sits exactly where my background overlaps: SEO, search systems, high-throughput data engineering, and practical business reporting.

From about 2001 through 2013, I worked in the SEO industry: directed over 1,400 clients, a team of more than 70, and spoke at a number of the major conferences. I designed and executed SEO projects for Lowe’s Home Improvement, Motorola, Marriott, Salesforce, and others.

Since about 2012, I’ve been on the engineering side: everything from search engines indexing over 10M documents, to full-stack development, to a year writing embedded software for a high-end audio company.

If you’re looking for large-scale data collection and analysis like this, large-scale link prospecting, or large-scale reputation monitoring, let’s talk about your project.