LLM SEO: how to optimize for ChatGPT, Claude & Perplexity

Optimizing for AI engines is not an agency hunch — it is a field with peer-reviewed research and public behavioral data. Three sources form the foundation of this work.

1. The foundational research

The term Generative Engine Optimization was formalized by a team from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi in the paper GEO: Generative Engine Optimization (Aggarwal et al.), presented at ACM KDD 2024. It remains the only widely cited peer-reviewed work measuring which tactics actually move the needle on AI citations.

The study tested nine tactics on a benchmark of 10,000 real queries across nine domains. Validated tactics can increase visibility inside generative-engine answers by up to 40%. The single tactic with the largest measured effect — emphasizing cited sources inside the content itself — raised citation probability by +115%.

2. User behavior already shifted

The Pew Research Center measured the real browsing behavior of 900 U.S. adults in March 2025. When Google shows an AI-generated summary, users click a traditional result only 8% of the time — versus 15% when no summary appears. Inside the summary itself, they click the cited sources just 1% of the time.

Google position still matters, but fewer users scroll to see it. The new fight is to be inside the answer, not beneath it.

3. The scale is no longer marginal

At his Google I/O 2026 keynote, Sundar Pichai confirmed that AI Overviews now exceed 2.5 billion monthly active users. AI Mode, Google's conversational interface, exceeds 1 billion. This is no longer an experiment: it is how hundreds of millions of people get answers today.

Academic research defines how visibility is earned in generative engines. Pew confirms why that visibility now outweighs traditional ranking. Google's own scale proves the shift has already happened — it is not coming.

Aggarwal et al. didn't just theorize: they tested nine concrete tactics on 10,000 real queries and measured how each raised or lowered visibility. Ranking by measured impact — higher means closer to being cited by AI.

1. Source Emphasis. Visually highlight citations inside the content (bold, callout boxes, prominent placement) — +115% citation probability. The largest measured effect and the most underused tactic on the web.
2. Statistics Addition. Include verifiable quantitative data (percentages, figures, dates) instead of generic claims. LLMs prefer concrete numbers because they are easier to attribute and quote.
3. Quotation Addition. Quote experts or recognized sources verbatim. Quotation marks are a structural signal the model reads as backed content.
4. Authoritative Tone. Write with professional, firm language — avoid hedging ("maybe", "it could be that"). Models penalize ambiguity because it makes building clean statements harder.
5. Fluency Optimization. Well-built sentences, clean grammar, clear transitions. What reads well for a person extracts well for a model.
6. Easy-to-Understand language. Accessible vocabulary. The clearer the language, the higher the chance the AI reproduces the idea without distortion.
7. Technical Terms. In specialized content, correct domain terminology reinforces topical relevance. Combine with the previous tactic: clear but precise.
8. Unique Words / coverage. Include information that does not appear in other sources. When the model picks between ten pages that say the same thing and one that adds something new, the new one wins.
9. Keyword stuffing — does not work. The old-school SEO tactic. The paper measured that artificially repeating keywords reduces visibility in generative engines — one of the validated negative findings.

These nine tactics operate at the content level. The structural layer — how pages link, how the entity is modeled, how authority is distributed — operates at another level, and it is where the gap lives between "sometimes shows up" and "the source the AI picks by default".

Research defines the framework. These are two public cases where the framework was applied to real clients — one in Spanish, one in English — with the evidence documented in my case studies.

Case 1 · From 0 to 987 organic users in 6 months

Peña Edelweiss (Sabiñánigo, Spain) runs a hand-built website in plain PHP: no WordPress, no CMS, no plugins. Since launch it had never registered organic traffic — zero sessions, zero users.

In June 2025 I started a technical-SEO + content project applying the GEO framework. The numbers verified in Google Analytics 4:

• Jan 1 – Jun 15, 2025 (before): 0 active users, 0 new users.
• Jun 16 – Dec 31, 2025 (after start): 987 active users, 987 new users, average engagement time of 1m 59s.

Case 2 · 6 LLMs citing a client in Chile

Asesorías Integrales JAAO (Chile) is the proof that the GEO framework also works to be cited, not only to rank. Between April 1 and May 25, 2026, six different generative engines registered verified referral traffic in GA4 to the client's landing page:

• ChatGPT (chatgpt.com) — citations from Linares, Santiago and Talca.
• Gemini (gemini.google.com) — Santiago.
• Claude (claude.ai) — Santiago, logged as user's first source.
• Perplexity (perplexity.ai) — Curicó, logged as user's first source.
• Copilot (copilot.com) — Santiago.
• NotebookLM (notebooklm.google.com) — Calama. Google's assistant indexed the landing, used it to answer a query, and the user clicked the citation.

In several cases the LLMs appear as the user's first source in GA4 — the prospect reached the site directly from the AI answer, without passing through Google or any traditional search engine first. Verified geographic distribution across 9 Chilean cities: Santiago, Curicó, Temuco, Valdivia, El Carmen, San Pedro de la Paz, Calama, Linares and Talca.

The same methodology, applied in English, supports the case of Destined Energy (Texas, solar): ChatGPT recorded as a referral source with 13 sessions, 80% engagement rate and over a minute of average time on site — users arriving from Highland Village, Dallas, Flower Mound and Port Louis to a local Dallas–Fort Worth business.

All data comes from Google Analytics 4 and Google Search Console, is auditable, and the cases are publicly linked from my portfolio. What is not public is the exact methodology — what was optimized, in what order and why — because that is the work.

The three disciplines coexist and complement each other. The common confusion is treating them as synonyms or, worse, as mutually exclusive. This table separates them by what matters: what each one optimizes, where the result is measured, and which signal weighs most.

The three work in layers: a site well-optimized for SEO tends to be a solid base for AEO, and a good AEO base accelerates GEO. But none transfers automatically to the next — each layer demands its own work. That is why so many number-one sites on Google are invisible to ChatGPT.

These are the patterns that repeat most when auditing sites that "should" be cited by AI and are not. None is exotic: all are seen daily.

1. Confusing length with depth

   A 3,000-word article full of filler is not deep content: it is long content. LLMs do not reward length; they reward the density of extractable information. A 1,200-word page with concrete data, highlighted sources and a clear structure systematically beats a 3,500-word one full of generalities.

2. Talking about yourself without verifiable data

   "We are industry leaders", "we offer the best solution", "over 20 years of experience". No number, no source, no date. LLMs discard that language because they cannot attribute it with confidence. Concrete figures, dates and proper names — those they do extract.

3. Hiding the sources

   The Princeton paper found that visually emphasizing cited sources raises citation probability by +115%. And yet, most sites bury their references in tiny footnotes or omit them entirely. Citing others — well — is what makes AI cite you.

4. No clear entity

   "About us" as three generic paragraphs, with no Person or Organization schema, no links to public profiles, no Wikidata presence. To an LLM, a company without a structured entity is a company it cannot name without risking error — so it doesn't name it.

5. Optimizing only for Google

   The most expensive mistake of 2026. Having an impeccable site for traditional SEO and never asking whether that same site is citable by ChatGPT. They are two different games with rules that overlap by about 30%. The other 70% is where the traffic of the coming years is being decided.

LLM SEO is the practice of engineering pages so that large language models — ChatGPT, Claude, Gemini, Copilot, Perplexity — extract and quote them as authoritative sources. The optimization unit is a self-contained passage with one claim, one unit of measurement, one entity and one primary-source citation. Pages built this way also win Google's AI Overviews and featured snippets, because the underlying signals overlap.

• Entity grounding — JSON-LD Person, Organization and Article with sameAs to LinkedIn, GitHub and press bylines.
• Primary sources — FTC, EIA, NREL, government datasets cited inline, not blog roundups.
• Crawler access — GPTBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot allowed in robots.txt.
• Extractable passages — 40–120 words, one intent each, semantic HTML, no carousel-trapped text.
• External trust — at least one mention on a publication the model already weighs heavily.

1. Days 1–14: Prompt audit. Query ChatGPT, Claude, Gemini, Copilot and Perplexity with 20 priority intents. Log who gets cited. Identify the 5 intents where a competitor wins and you don't.
2. Days 15–45: Rewrite the first passage of each target page in 40–60 words with one primary-source link and the exact entity name a user would query. Add Article + FAQPage JSON-LD with explicit author and dateModified.
3. Days 46–75: Open crawler access. Add GPTBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot allow-lines to robots.txt. Submit a clean sitemap.xml and an llms.txt index.
4. Days 76–90: Measure GA4 referrals from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com. Re-run the prompt audit and iterate on passages that earn impressions but not citations.