LLM Ranking Factors Report Data (May 2026)

I spent three weeks running what might be the most comprehensive LLM ranking factors analysis to date. It tracked and scored 29,562 unique domains across 145 industries, 1,595 buyer personas and 105k+ ChatGPT prompts. It covers over 500TB of data and correlates 13 external signals against rank-weighted LLM recommendation scores.

This is a large-scale correlation study: what external signals actually predict whether a brand gets recommended by ChatGPT, across 145 industries and 1,595 buyer personas. The point is not to reverse engineer a neural network. The point is to measure which observable signals are associated with LLM recommendation visibility, by industry and by persona.

The full report is free: oppalerts.com/LLM-Ranking-Factors. I first shared it in this LinkedIn post on May 11, 2026.

How To Use This Report

The report is built to be used now, whether you work in-house, run an agency, sell strategy, or manage ongoing client work.

The benefit is focus. Instead of guessing at generic AI visibility tactics, you can look at the exact industry, the exact persona, and the signals that appear to matter for that market.

The Largest LLM Ranking Factors Report, to Date

I believe this is the largest LLM ranking factors analysis published so far. I don't know of anything else that analyzes this many industries, personas, prompts, web pages, search results, Reddit posts, Wikimedia records, and backlinks at the same time, trying to understand an LLM's ranking factors. If something larger or more comprehensive exists, please let me know.

Here's how the research started:

• 145 industries, narrowed from about 500 candidate industries.
• 1,595 industry and persona combinations: 10 targeted personas plus one neutral persona per industry.
• 105k+ prompts run through ChatGPT 5.4, including repeated recommendation prompts for statistical stability.
• 29,562 unique recommended domains scored against 13 external technical signals.

The Scale

The research required fast code, not a giant cloud budget. This project analyzed more than 500TB of text data. Despite the size, the analysis took only about three days. This is a result of spending significant time over the last 12+ years obsessed with fast parsing, including writing an HTML parser and article extractor that is 100x faster than anything in the public domain.

Research Process

How the recommendation dataset was built.

• I started with about 500 candidate industries and refined them to 145 with sufficient domain coverage.
• I generated 1,595 buyer personas: a neutral baseline plus 10 targeted market segments for each industry.
• For each persona, I collected top recommended domains, top search phrases, and top on-page phrases.
• I ran recommendation prompts repeatedly and scored each domain with rank weighting, so a #1 recommendation counted more than a #10.
• I downloaded and analyzed search results, recommended homepages, web crawl data, Reddit data, Wikimedia data, and backlink graph data.

Analysis Process

How the signals were compared against LLM recommendation scores.

• Each domain received a rank-weighted LLM recommendation score per persona and per industry.
• 13 signals were extracted per domain: SERP appearances, best rank, average rank, outbound links, backlink count, PageRank, backlink authority, harmonic centrality, Common Crawl, Reddit comments, Reddit posts, Wikipedia citations, Wikidata, and homepage keywords.
• Spearman rank correlation was used because recommendation scores and authority metrics are heavily skewed.
• R² estimates how much variance each signal explains. Lift shows how much more common a signal is among the top 10% most-recommended domains.
• Every signal is evaluated globally, per industry, and per persona where data coverage is sufficient.

All-Industry Signal Correlations

Spearman rho versus LLM recommendation score across all industries. Search appears at the top globally, but the effect is partial. SERP appearances explain 5.8% of variance, which is meaningful at this scale but nowhere near a complete explanation.

• Search Engine Appearances: rho = +0.241, R² = 5.8%, lift = 2.00x.
• Best Search Engine Rank: rho = +0.238, R² = 5.7%.
• SE Outbound Links: rho = +0.230, R² = 5.3%, lift = 2.08x.
• Backlink Count: rho = +0.204, R² = 4.2%, lift = 2.84x.

The full table of signals is in the report.

Example Industry: Hotels & Resorts

Hotels and resorts is a useful example because it includes major brands, OTAs, luxury chains, travel guides, and direct booking sites.

What predicts hotel recommendations?

In hotels and resorts, the strongest measured relationship is Wikidata. In plain English: hotel brands with clearer structured reference data are much more likely to be recommended. In this industry, Wikidata explains 42.3% of the measurable difference in LLM recommendation scores.

Search still matters, but not as a simple ranking copy. Search engine outbound links explain 21.3% of the measurable difference, which suggests the pages and sources surfaced around hotel queries are part of the evidence layer the model is drawing from.

Wikipedia citations explain 19.3%, and Reddit posts explain 15.2%. That combination is important: the model is not only responding to search visibility. It is also responding to structured references, encyclopedia-level citations, and community discussion.

Top LLM domains

Marriott, Hilton, Hyatt, Booking.com, and Expedia lead the LLM recommendation list, but the search list is not identical. That gap is where the analysis gets useful.

• Top LLM domain: marriott.com, 100.0% normalized LLM score, recommended in 49.1% of hotel query runs.
• Next strongest hotel domains: hilton.com, hyatt.com, booking.com, expedia.com.
• The leading domains are not just search winners. They are brands with durable visibility across multiple evidence layers.

Domain signals matrix

A domain can be strong in search, weak in Wikidata, strong in backlinks, or visible on Reddit. LLM visibility usually comes from a mix, not one universal factor. High LLM visibility comes from different combinations of search, backlinks, Wikidata, Wikipedia, Reddit, Common Crawl, and homepage signals. The matrix shows which evidence layers are strong or weak for each domain.

Top search engine domains

Tripadvisor is #1 in search appearances for this industry, while Marriott is #1 in LLM recommendations. This is the core reason LLM visibility cannot be inferred from search rank alone.

• Top search domain: tripadvisor.com, 223 hotel SERP appearances.
• Top LLM domain: marriott.com, 100.0% normalized LLM score.
• The mismatch is the point: search visibility is important, but it is not the same thing as LLM recommendation visibility.

Backlink authority

Authority still matters, but authority alone is not the ranking. Google.com has the largest backlink authority in the hotel industry data but only low hotel LLM visibility. Authority is a signal. Relevance and category fit still matter.

• Top authority domain in the hotel dataset: google.com, PR = 100, 16.9M backlinks.
• Top hotel-relevant authority domains include tripadvisor.com, booking.com, marriott.com, airbnb.com, hilton.com, and hyatt.com.
• Authority is one part of the signal stack, not the whole explanation.

Key Findings From The Report

The report does not point to one universal LLM ranking factor. It shows that different industries lean on different evidence layers, and that persona context changes the recommendation set.

1. The search engine inside the AI

Search visibility is the most consistent global external predictor, but it is not the whole system.

LLMs are not just guessing. Search results are part of the recommendation environment, whether through training data or live retrieval. The correlation is statistically significant across the full dataset. But 5.8% R² means search explains less than 6% of the variance. Search visibility matters, but it is one input into a much larger system.

2. Backlinks still matter, but the effect is different

The relationship is real, but it behaves differently than traditional SEO reporting.

Backlink distributions are extremely skewed. The most-recommended brands are not just slightly more cited. They are often structurally more present across the web. This does not reduce LLM optimization to link building. It means broad citation and authority signals have already been absorbed into the model's brand priors.

3. The 80 to 85% black box

External signals are measurable, but most recommendation behavior is inside the model.

Training data distribution, fine-tuning, RLHF preferences, and long-term brand familiarity are not directly observable from the outside. That is why LLM visibility cannot be inferred from search rankings or backlink scores. It has to be tested directly, repeatedly, by persona.

4. Different industries, different signal hierarchies

There is no universal ranking factor list that works across every category.

This is why broad meta-analyses miss the practical point. The signal that matters in one industry can be secondary or irrelevant in another.

5. Wikipedia and Wikidata are underrated LLM signals

For the industries where structured-reference coverage exists, the effects are large.

Wikipedia and Wikidata are part of the foundational reference layer for LLMs. When a brand has a clear structured entity, the model has a stronger durable representation of that brand. The caveat is coverage. Wikidata is sparse globally, but where it applies, the relationship can be one of the strongest in the dataset.

6. Reddit drives AI recommendations in surprise categories

Community discussion is a strong signal in categories where people ask for real opinions.

These are categories where people use Reddit to vet options: AI tools, ticketing platforms, elective healthcare, and other high-trust or high-friction decisions. If a brand dominates real community discussion, that conversation can become part of the model's recommendation behavior.

7. Personas produce completely different recommendation sets

Same industry. Same model. Different person asking. Different result.

The industry-wide #1 for airlines is delta.com, but no single airline is #1 across all personas. A student abroad flyer gets StudentUniverse, not a major carrier. This is the central measurement problem: LLM visibility is persona-specific by default.

8. Executive persona recommendations are radically compressed

High-value personas can have very narrow recommendation sets.

For a high-end executive traveler, the recommendation set collapses toward known luxury and business chains. Anything outside that set is effectively invisible for the persona.

9. Auto insurance is remarkably predictable

A mature category with dominant national brands produces strong external correlations.

In mature markets, LLM recommendations often reflect existing market structure. The model has absorbed years of brand authority, advertising, links, reviews, and search presence.

10. Airlines: backlinks over rankings

For airlines, backlink authority explains about 3x more variance than SERP appearances.

Major airlines accumulate links from news, travel guides, booking flows, airports, credit card pages, and comparison sites. That broad web citation graph is part of what the model has learned.

11. Live entertainment: Reddit is the ticket

Ticketing is a category where community reputation is unusually visible.

People debate fees, trust, scams, venues, resale prices, and app experiences on Reddit. The model appears to have absorbed that category-specific conversation.

12. Persona coverage is a new kind of market share

A brand's LLM visibility is a distribution across personas, not one score.

Persona coverage measures how many distinct buyer contexts can see you at all. For brands that compete across multiple segments, broad persona coverage may matter more than a single aggregate LLM score.

What Conclusions Can We Draw?

The research does not produce one magic ranking factor. It produces a measurement framework.

• Search visibility matters, but it explains only part of LLM recommendation behavior.
• Backlink authority, Wikipedia and Wikidata presence, Reddit discussion, Common Crawl coverage, and homepage relevance vary dramatically by industry.
• The best signal for one industry can be irrelevant in another.
• The most important measurement unit is not the keyword. It is the persona or market segment.

That means the right question is not "What is the LLM ranking factor?" The right question is "For this industry, for this persona, what signals distinguish the domains that get recommended?"

Conclusion 1: SEO principles still apply

Search is not dead. It is becoming one layer of AI recommendations.

The data shows a consistent relationship between Google search visibility and ChatGPT recommendations. ChatGPT is using search results today, and OpenAI is building its own search infrastructure. That makes SEO fundamentals more important, not less important. Authority, relevance, crawlability, citation breadth, structured references, and useful content still contribute to the web evidence LLMs consume.

The mistake is treating SEO rank as a proxy for LLM visibility. It is a signal, not the answer.

Conclusion 2: Personas change the measurement

You cannot monitor LLM visibility like a search ranking report.

In SEO, personalization was usually a modifier. You could track a keyword and get a useful approximation of rank. With LLMs, personalization and user context are the center of the recommendation. A frequent flyer, a student traveling abroad, a family budget planner, and a senior comfort traveler can all ask for airline recommendations and receive materially different answers.

A single aggregate LLM rank is not a meaningful monitoring unit. Brands need to define the personas or market segments they care about, test each one, and track movement over time. This is probably the most important finding in the report: if you are measuring LLM visibility without specifying who is asking, you are measuring noise.

Free Full Report

The public report includes all 145 industries, per-persona breakdowns, signal tables, and industry-specific analysis: oppalerts.com/LLM-Ranking-Factors. For sales teams and agencies, it can help show prospects where they are missing from LLM recommendations and which competitors are being surfaced instead. For service teams, it can help define the next three to six months of work with a client: which personas to monitor, which signals to improve, and which gaps to prioritize.