What Is LLM SEO and How Does It Work?

LLM SEO (also called Generative Engine Optimization or GEO) is the practice of optimizing your website to be cited by AI-powered search engines — ChatGPT, Perplexity AI, Google AI Overviews, and Bing Copilot. The term GEO comes from Aggarwal et al.'s Princeton study (2023).

Unlike traditional SEO, where you compete for a position in a list of ten links, LLM SEO targets a different outcome: being one of 3–5 sources cited in a synthesized AI answer. If you are not cited, you get zero traffic from that query.

The most important finding so far: an empirical study across 441 domains and 14,550 domain-query pairs found that content relevance is 62x more predictive of citations than any structural optimization (schema markup, FAQ sections, etc.). Structural readiness scores showed zero correlation with citation rates (r=0.009, p=0.849).

This article provides an evidence-based assessment of what LLM SEO actually means, what has data behind it, and what remains unproven marketing. Confidence levels are labeled throughout.

What My Data Actually Shows

I built getaisearchscore.com to score websites on 26 structural factors across four categories: machine readability, content extractability, entity trust, and offering completeness. Then I tested whether those scores predict real citations.

What did predict citations? Content relevance. Same-topic citation rate was 5.17% versus 0.08% for cross-topic — a 62x difference. Domain authority acted as an amplifier: high-DA sites with relevant content got cited more, but high DA alone without relevant content did not.

For the full methodology and data, see our null-finding study.

What the GEO Paper Actually Showed (vs. How the Industry Reads It)

The GEO paper by Aggarwal et al. is the most-cited academic source in the LLM SEO space. The industry treats it as proof that “GEO works.” But what did it actually test?

The study tested content-level interventions — adding statistics, quotations, and citations to the text itself. These are writing techniques, not structural optimizations. Adding statistics improved visibility by 15–41% depending on the domain.

Here is what the industry often misses: the paper did not test Schema.org markup, robots.txt configuration, JSON-LD structured data, or any of the technical signals that most “LLM SEO tools” (including mine) check. The interventions that worked were about making the content itself more authoritative and specific.

This is consistent with my own data. Content relevance and content quality drive citations. Structural scaffolding around that content does not appear to matter — at least not in any way my study or the GEO study could measure.

How AI Search Retrieval Works

AI search engines use retrieval-augmented generation (RAG). Understanding this pipeline helps explain why content relevance dominates and structural signals do not:

User query
  ↓
Query embedding (convert to vector)
  ↓
Vector search (retrieve top-k candidate pages)
  ↓
Reranker (score relevance of candidates)
  ↓
LLM synthesis (generate answer, select citations)

The critical step is vector search. It finds pages based on semantic similarity between the query and your content. Not keywords. Not Schema.org markup. Not robots.txt rules. Semantic meaning of the actual text.

This explains the 62x content relevance finding. If your page is semantically close to the query, it gets retrieved. If it is not, no amount of structured data helps because the retrieval step never surfaces your page in the first place.

Evidence Tiers: What We Know, What We Suspect, What Is Marketing

Tier 1: Supported by Data (High Confidence)

Tier 2: Plausible but Unproven (Medium Confidence)

Tier 3: Likely Marketing (Low Confidence)

LLM SEO vs Traditional SEO: Real Differences

The differences are real even if the optimization playbook is less clear than the industry claims:

What I Would Actually Recommend (Given the Data)

Given that I built a structural optimization tool and then found structural optimization does not predict citations, here is what I would actually tell someone to do. Ordered by evidence quality, not by how easy it is to sell as a service.

1. Write Content That Directly Answers Questions in Your Domain

This is the 62x factor. If someone asks an AI “What are the best diving shops in Lisbon?” and your page is literally about diving shops in Lisbon with specific information, you have a shot. If your page is about water sports equipment generally, you probably do not.

Confidence: high. Supported by my data, the GEO paper, and RAG architecture.

2. Include Specific Data, Statistics, and Original Claims

The GEO paper found that adding statistics improved visibility by 15–41%. This makes mechanical sense: LLMs need to attribute specific factual claims to sources. Generic advice gets paraphrased without citation. Specific numbers get cited.

Confidence: medium-high. GEO paper data plus mechanical plausibility.

3. Do Not Block AI Crawlers

Check your robots.txt. If GPTBot, PerplexityBot, or ClaudeBot are blocked, you are invisible to those platforms. This is the one structural factor with a clear binary effect.

User-agent: GPTBot
Allow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: ClaudeBot
Allow: /

Confidence: high. Binary access control. No access = no citations.

4. Build Domain Authority (the Hard Way)

Domain authority amplifies content relevance. This is not news — it is the same lever that traditional SEO has used for decades. Backlinks, brand mentions, consistent publication. There is no shortcut here, and anyone selling one is probably selling you something else.

Confidence: medium. My data shows DA as an amplifier, not a direct driver.

5. Structural Optimization (Low Priority, Low Cost)

Schema.org markup, FAQ sections, BLUF answer blocks, clean heading hierarchy. I still think these are good practice — they make your content clearer for humans too. But I cannot honestly claim they will get you cited based on my data.

If you want to do them anyway (and they are cheap to implement), the highest-value actions take under an hour:

1. Add Organization JSON-LD with sameAs links (15 min)
2. Add FAQPage schema to one key page (20 min)
3. Write a 40–60 word BLUF summary under your main H2 (10 min)
4. Ensure Product schema has price, currency, and availability (15 min)

Confidence: low. Mechanistically plausible, not empirically validated for citations. Good for general content quality regardless.

What the LLM SEO Industry Gets Wrong

Most LLM SEO advice (including what I published before running the study) assumes that structural readiness causes citations. The reasoning goes: AI engines need structured data to parse your site → better structure = more citations.

This misunderstands the RAG pipeline. The retrieval step uses vector similarity on content embeddings. Structured data is not part of that embedding. By the time the LLM is deciding what to cite, the candidate pages have already been selected — and they were selected based on content meaning, not structural signals.

The exception might be Google AI Overviews, which has a tighter integration with Google's existing search infrastructure (including Knowledge Graph and structured data parsing). But even there, I lack controlled data.

The industry also confuses “correlates with good websites” with “causes citations.” Sites with clean Schema.org markup tend to also have relevant, well-written content. The markup did not cause the citation — the content did. The markup was just along for the ride.

An Honest Assessment of My Own Tool

I built getaisearchscore.com to score websites on 26 AI readiness factors. The tool works as designed — it accurately measures structural readiness. The problem is that structural readiness does not predict the outcome people care about: getting cited.

The tool is still useful as a diagnostic. It catches broken robots.txt rules, missing Schema.org markup, JS-rendering issues that prevent crawling. These are real technical problems worth fixing. But I would no longer claim that improving your score will improve your citation rate.

You can run a free audit to check for technical issues. Just understand that the score measures readiness, not citation probability. Those turned out to be different things.

What Remains Unknown

My study has limitations. 441 domains is meaningful but not enormous. I tested Perplexity citations specifically — Google AI Overviews and ChatGPT may weight structural signals differently. The study was cross-sectional, not longitudinal.

Open questions I cannot answer yet:

• Does Schema.org markup specifically help Google AI Overviews citations? (Plausible, untested in my data.)
• Does structural optimization have a delayed effect that a cross-sectional study would miss?
• Are there interaction effects — does structure matter only when combined with high relevance and high DA?
• How much does the citation landscape change as AI search engines update their retrieval pipelines?

Anyone claiming certainty about LLM SEO in 2026 is selling something. The field is two years old, the platforms change their pipelines constantly, and the published research is thin. The honest answer to “What is LLM SEO?” is: a set of practices with limited empirical validation, built on reasonable but untested assumptions about how AI search works.

The Bottom Line

If I had to compress everything I learned into three sentences: Write content that directly and specifically answers the questions your customers ask. Include concrete data and original claims that LLMs need to attribute. Everything else — Schema.org, BLUF blocks, FAQ sections — is good practice but has no proven causal link to citation outcomes.

The LLM SEO industry will not love this assessment. But I would rather be honest about what my data shows than sell advice I cannot back up.

For the full data behind these conclusions, read the null-finding study. For practical content advice grounded in the GEO research, see how to improve your citation rate. To check your site for technical issues regardless, try the free AI readiness audit.