GEO: how LLMs read your content and what to change so they cite you

A generative engine doesn't search for words. It joins the dots.

In February 2024 Gartner published a forecast that has since been quoted to death: by 2026, traditional search volume will fall by 25% as people turn to conversational AI agents. The figure still stands. Semrush's August 2024 report found that traffic from ChatGPT to external websites grew more than 30-fold in a year. BrightEdge estimated in January 2025 that Google's generative answers (AI Overviews) cut clicks on classic results by 30% to 60%, depending on the vertical.

The picture is simple: fewer clicks, more citations.

In a classic search engine, if you're not on the first page, you don't exist. In a generative engine, if you're not cited inside the answer, you don't exist either, even if you're indexed. The game is no longer about position. It's about being mentioned.

This has an operational name: GEO, Generative Engine Optimization. And it isn't SEO with a new logo. What changes is what the machine reads.

It isn't optional.

What GEO is and why it isn't the same old SEO with a new name

Classic SEO was built on three measurable pillars: keyword matching, domain authority measured by inbound links, and on-page user experience (Core Web Vitals). Google returned ten blue links, you fought to get into the top three, and the click was the conversion.

GEO works with a different machine. A generative model (GPT-4o, Claude 3.5 Sonnet, Gemini 2.0, Llama 3.3) doesn't retrieve ten links. It retrieves fragments of text, recombines them and synthesises a single answer. According to the Princeton and Georgia Tech study on generative search published in 2024, that answer cites between 3 and 8 sources on average. Being one of those 3 to 8 is the goal.

A quick comparison:

Translation: classic SEO rewards whoever best matches a query. GEO rewards whoever makes it easiest for the model to do its job, which is joining up entities and saying where each piece of data comes from.

An example. A firm called Consultoría Administrativo del Norte publishes a guide to electronic invoicing. For classic SEO it optimises for "facturación electrónica pymes 2025" (e-invoicing for SMEs 2025), crams the keyword in 14 times, picks up three links from sector blogs and aims for position 4. For GEO, the guide is built around three well-defined entities: the Crea y Crece law (Spain's business growth act, article 12), the Facturae format (the Spanish public sector's e-invoice standard), and the AEAT (the Spanish tax agency) as issuing authority. Each entity links to its official source, each claim carries a date and a number, and each block answers a literal question. The result: Perplexity cites it when someone asks "when does mandatory e-invoicing come into force in Spain?".

The two can coexist. The machinery behind them is different.

How an LLM reads your page: entities and relations, not keywords

This is the hardest part to absorb if you come from classic SEO. Current models don't count how many times "GEO" appears in your text. Simplified, this is what they do:

1. They tokenise the text into chunks of 200 to 800 tokens.
2. They turn each chunk into a high-dimensional vector (an embedding).
3. They store that vector with metadata: URL, title, date, author, domain, and structured schema if there is any.
4. When a query comes in, they turn it into a vector too and use cosine similarity to find the closest chunks.
5. They pass those chunks as context to a generative model, which writes the answer and decides whom to cite.

Step 3 is where you win or lose. If your chunk talks about "the regulation" without naming it, the resulting vector is semantically weak. If it talks about "the European AI Regulation (EU 2024/1689), Article 4, in force since 2 February 2025", the vector is dense and specific, and when someone asks "what does the AI Act require on training?", your chunk comes out on top.

Rule: every block needs enough named entities (by their official name, not a paraphrase) for the resulting vector to point at a specific area of the semantic space.

To an LLM, an entity is anything with a proper name and stable relations:

• People. An author identified by full name and credentials.
• Organisations. Companies, agencies, ministries, by their legal name.
• Products and services. Trade name, version, category.
• Legislation. Title, directive or regulation number, article, year.
• Dates. Absolute, never "recently".
• Metrics. Figure + unit + source.
• Places. City, region (in Spain, the autonomous community), country.

A paragraph built well for GEO names at least three distinct entities and sets out a clear relation between them. A badly built one says "many companies are adopting this technology to improve their processes", and its vector disappears into a limbo of a billion equivalent sentences.

An LLM doesn't reward what sounds good. It rewards what can be checked.

Schema.org, JSON-LD and sameAs: the signals the model reads without ambiguity

Plain text works. Plain text with structured schema works much better. The reason is simple: when a crawler finds JSON-LD on your page, it doesn't have to guess. You've handed it the schema in machine-readable form.

The Schema.org types that pay off most for GEO in 2026:

• Organization. With name, url, logo, and sameAs pointing to Wikidata, LinkedIn, GitHub and official profiles.
• Person. For the author. With jobTitle, worksFor, alumniOf, sameAs.
• Article. For posts, with headline, datePublished, dateModified, author, publisher.
• FAQPage. Each question with Question and acceptedAnswer. This is the type that most often triggers citations in Perplexity.
• HowTo. When the content is procedural, with numbered step entries.
• Product. If you sell software or a service, with brand, category, audience.
• BreadcrumbList. So the model understands where the section sits in the hierarchy.

The key field, and the most ignored, is sameAs. It's the explicit link between your entity and its representation in public graphs: Wikidata (https://www.wikidata.org/wiki/QXXXXXXX), DBpedia, LinkedIn Company, Crunchbase, GitHub Org. When the LLM sees sameAs, the ambiguity goes away. It knows Consultoría Administrativo del Norte isn't one of the five other firms with a similar name, because you've told it so with a stable identifier.

A minimal JSON-LD example for a service page:

{
  "@context": "https://schema.org",
  "@type": "Service",
  "name": "Vertical software for consultancies",
  "provider": {
    "@type": "Organization",
    "name": "APFerrer",
    "url": "https://apferrer.com",
    "sameAs": [
      "https://www.linkedin.com/company/apferrer",
      "https://github.com/apferrer"
    ]
  },
  "areaServed": "ES",
  "serviceType": "Custom software development",
  "audience": {
    "@type": "BusinessAudience",
    "audienceType": "Consultancies and professional firms"
  }
}

Rule: every page you want cited needs at least one JSON-LD block with the most specific type available, and sameAs on every Person or Organization entity.

Critical note: pasting in the JSON-LD and forgetting about it isn't enough. Run it through Google's Rich Results Test and the Schema.org Schema Markup Validator. Silent JSON-LD errors (a trailing comma, a misspelt type) get the whole block thrown out without warning.

The explicit question-answer pattern inside the content itself

This is the most counterintuitive part if you've spent years writing long posts. An LLM doesn't need a sales pitch. It needs blocks that answer a specific question in a self-contained paragraph.

The winning structure is simple:

1. An H2 or H3 phrased as a direct question.
2. A first sentence that answers it in under 30 words.
3. A paragraph of 60-120 words with the context, the data point and the source.
4. Optional: a list of sub-cases if the answer has nuances.

Here is the same content in two formats.

Bad for GEO:

The evolution of search ranking in the age of AI

Search engines have changed a lot in recent years. It used to be enough to optimise keywords, but things are different now. People don't only search on Google any more; they ask conversational assistants questions, and this forces us to rethink our content strategy from top to bottom, adapting to the new realities of the digital market.

Good for GEO:

How does GEO differ from traditional SEO?

GEO optimises for being cited in generative answers (ChatGPT, Perplexity, Gemini, Google AI Overviews) instead of ranking on a SERP of ten blue links. Classic SEO prioritises keywords, backlinks and CTR; GEO prioritises well-named entities, structured schema and self-contained question-answer blocks. The first optimises for the click, the second for the mention. Both can coexist on the same page, but the markup and the paragraph architecture change.

The second block has a dense semantic vector, answers the H2 question literally, names all four platforms and draws an operational contrast. It's the kind of block an LLM lifts and cites untouched.

Rule: a GEO-ready post isn't an essay. It's a mosaic of atomic answers held together by a single argument.

Apply this to FAQ Schema as well. Each Question in your FAQPage must match a real H3 in the body, with the answer written in the same tone it will be cited in. If the schema says one thing and the visible HTML says another, the crawler penalises the inconsistency. Google Search Central has been doing this since the March 2024 update.

Verifiable sources: why the model prefers to cite whoever shows where the data comes from

A generative model has a quiet priority: reducing the risk of hallucination. When two chunks compete to answer a query, the verifiable one wins. For a machine, verifiable means something very specific: an external link to a primary source, a figure with context, an absolute date.

The studies on grounding that OpenAI and Anthropic published in 2024 and 2025 point the same way: internal RAG (Retrieval Augmented Generation) systems give more weight to chunks with explicit citations and numerical data than to chunks that offer only opinions or unsupported claims. Perplexity has said so openly in its public documentation since November 2024: it prioritises domains that cite primary sources.

What counts as a verifiable primary source:

• Official gazettes. The BOE (Spain's official state gazette), the Official Journal of the EU (DOUE in Spanish), the DOGV (the Valencian regional gazette) and other regional equivalents, with a direct URL to the article.
• Public registers. INE (Spain's national statistics institute), Eurostat, AEAT, CNMV (Spain's securities regulator), ministries.
• Papers with a DOI. arXiv, IEEE, ACM, Nature, with a citable identifier.
• Official product documentation. GitHub repositories, versioned technical docs.
• Consultancy reports with a public methodology. Gartner, McKinsey or Deloitte, when they link to a whitepaper with a sample and a method rather than a blog post.
• Reproducible in-house data. Internal studies with a documented methodology and a clear time window.

What does NOT count:

• Blogs with no identified author.
• Social media posts with no screenshot or stable permalink.
• "Recent studies show" with no link.
• Wikipedia on its own. It works as a bridge, not as a final source.
• AI-generated content with no identifiable human review.

Fix: every quantitative claim in your post should follow the pattern "figure + unit + linked source + date of the data". For example: "Semrush's August 2024 report found that traffic from ChatGPT grew more than 30-fold in a year". Four elements, one dense vector.

Critical note: sources age. A GEO-ready post includes dateModified in its schema and updates its figures at least every 6 months. A 2022 data point in a post dated 2026 tells the crawler nobody is looking after the content.

GEO-ready editorial structure checklist

This is the operational part. When you convert a post or a page so an LLM can cite it, put it through these fourteen points. It's not a menu. It's a checklist.

1. Short, descriptive URL. No parameters, no dates, no session numbers. /blog/geo-como-entienden-llms-contenido-empresarial works; /post?id=8471 doesn't.
2. H1 title with the main entity and an action or question. Not just the keyword.
3. An excerpt of 100-240 characters in the frontmatter and in the meta description. Identical, or nearly so, in both places.
4. Identified author. With their own /author/name page, Person schema, a real photo and sameAs pointing to LinkedIn at the very least.
5. Visible publication and last-updated dates. In the HTML and in the Article schema.
6. Table of contents with anchor links. A <nav> linking to each H2, or a TOC component. LLMs follow internal anchors.
7. Every H2 phrased as a question or a concrete statement. "Introduction", "Background" and "Conclusion" are useless for GEO.
8. The first paragraph under each H2 answers it in under 30 words. The rest expands on it.
9. Entities named by their full official name. The first time they appear in the post. After that you can shorten them.
10. At least one sourced figure per section. If a section has no hard data at all, it's either philosophical (rare) or surplus to requirements.
11. JSON-LD with Article, a Person author, an Organization publisher, and FAQPage if there's an FAQ block. Validated.
12. An explicit FAQ block near the end, with the questions you want the LLM to cite word for word.
13. Internal links to 3-5 other pages on the same domain that belong to the same topic cluster. Internal graphs help the model understand how the site fits together.
14. External links to primary sources. At least 3 per post. Yes, even though they send traffic elsewhere. To the model, they signal honesty.

If your CMS won't let you control the JSON-LD, switch CMS or install a plugin that does. In 2026, not controlling your schema is like not controlling your <title> in 2010.

Common mistakes when implementing GEO

Here are five mistakes I see again and again when sites move from SEO to GEO, each with its symptom and fix so you don't have to guess.

Mistake 1: Stuffing the post with keywords instead of entities. Symptom: the post holds position 8 on Google but never gets cited by Perplexity or ChatGPT. Fix: rewrite each H2 as a literal question, name three entities per section by their official name, and add FAQPage JSON-LD.

Mistake 2: Copying another site's JSON-LD without adapting it. Symptom: the Rich Results Test returns an error, or the sameAs IDs point to other people's organisations. Fix: write the JSON-LD from scratch for your own entity, with your real LinkedIn, GitHub and Wikidata URLs, and validate it before publishing.

Mistake 3: Relying on a single generic FAQ at the end of the post. Symptom: the FAQ block has nothing to do with the body, and the questions are beginner-level while the post is technical. Fix: each FAQ H3 should cover a real query a user would put to an LLM and be backed up by a block in the body.

Mistake 4: Publishing without dateModified, or never updating it. Symptom: the post's data is from 2023, the publication date says 2026, and the crawler trusts it less. Fix: set a 6-month refresh policy, update the figures, and bump dateModified every time you review.

Mistake 5: Blocking AI crawlers in robots.txt by default. Symptom: generative tracking tools show zero citations even though classic SEO is going well. Fix: review robots.txt and make a deliberate choice about which user-agents you allow (GPTBot, PerplexityBot, ClaudeBot, Google-Extended, CCBot). Blocking everything for fear that "they'll copy me" has a measurable cost: not being cited.

None of these mistakes is exotic. They turn up in 70% of the GEO audits I review. They take an afternoon to fix. You just have to decide to do it.

How to measure whether you're being cited without installing a plugin

This is the least glamorous part and the one most often neglected. Without a metric, GEO becomes an act of faith. And faith doesn't pay the bills.

There are three layers of measurement, ordered by cost and precision.

Layer 1: manual checks with reproducible prompts

The bare minimum. You define 15 to 30 queries an ideal client would put to a generative model, run them every week on ChatGPT, Perplexity and Gemini with exactly the same wording, and note whether your domain appears among the citations.

The audit prompt looks like this:

"I'm researching [topic]. Give me the 3 best sources with links and explain what each one contributes."

Log the results in a spreadsheet: date, platform, query, domains cited, position of the citation, and whether yours is there. At 30 minutes a week, you'll have a time series within three months.

Layer 2: server log analysis

AI crawlers identify themselves by user-agent. The main ones in 2026:

• GPTBot and OAI-SearchBot (OpenAI).
• PerplexityBot and Perplexity-User (Perplexity).
• ClaudeBot and anthropic-ai (Anthropic).
• Google-Extended (Google, for training) and GoogleOther (for AI Overviews).
• CCBot (Common Crawl, used by almost every open model).
• Bytespider (ByteDance).
• Amazonbot (Amazon).

A simple script over your Nginx or Cloudflare logs tells you how many times a week each bot crawls which URLs. Steady growth in PerplexityBot visits to your new posts is an early sign that citations are on the way.

Layer 3: generative tracking tools

By 2026 there are established options: Otterly.AI, Peec.ai, Profound, AthenaHQ. All of them track mentions of your brand in generative answers across a defined set of queries and give you time series. Market pricing ranges from cheap monthly subscriptions for SMEs to annual contracts for agencies, depending on the number of queries and platforms covered. Don't compare them on headline price. Compare them on how many platforms they monitor (at least 4 in 2026), how granular the queries are and the quality of the citation data (whether they capture the exact snippet or just the domain).

Rule: even if you pay for a tool, keep layer 1 going by hand. It's your quality control. Aggregators break and tools shut down, but your spreadsheet with 30 queries checked every week won't.

Metrics to track, with indicative thresholds for a single-language B2B SME website:

If you're below these thresholds, the problem is usually in one of three places: badly implemented schema, vaguely defined entities, or missing external sources. It's rarely the topic of the content.

Frequently asked questions

How do I get my website into ChatGPT and Perplexity answers?

Publish content with well-named entities, valid JSON-LD (Article, Person, Organization, FAQPage, sameAs), H2 blocks phrased as questions with a self-contained answer in the first 30 words, and at least three external primary sources per post. Check that robots.txt doesn't block GPTBot, OAI-SearchBot, PerplexityBot or ClaudeBot. Reindex. The first citation usually arrives within 3 to 8 weeks.

What is GEO and how does it differ from traditional SEO?

GEO (Generative Engine Optimization) optimises for being cited in answers generated by LLMs (ChatGPT, Perplexity, Gemini, Google AI Overviews) instead of ranking in the classic SERP. Traditional SEO prioritises keywords, backlinks and CTR across ten blue links. GEO prioritises named entities, structured schema and question-answer blocks. The first measures clicks, the second mentions. You can combine them on the same page.

How do I optimise content for generative engines?

Follow the checklist: a descriptive URL, H2s as questions, a first sentence that answers in under 30 words, entities by their official name, a sourced figure in every section, valid JSON-LD, FAQ Schema aligned with the body, internal links to the topic cluster, external links to primary sources, an identified author with Person schema, and an up-to-date dateModified. Review everything every 6 months.

What signals does an LLM use to decide which source to cite?

Vector similarity between the chunk and the query, density of entities named by their official name, sourced numerical data, structured schema (especially FAQPage and Article), last-updated date, consistency between the visible HTML and the JSON-LD, domain authority measured by cross-mentions in primary sources, and sameAs linking the entity to public graphs such as Wikidata and LinkedIn.

Can I do GEO without touching classic SEO?

Not properly. GEO and classic SEO share the same foundations (semantic HTML, speed, an identified author, useful content) and part ways at the markup layer and in how paragraphs are structured. The efficient approach is to keep classic SEO as the base and add the GEO layer on top. For SMEs with limited resources, apply GEO first to the 10-15 pages that convert best and leave the rest on classic SEO.

How long does GEO take to deliver results compared with traditional SEO?

Classic SEO delivers stable results in 6 to 12 months. GEO can bring its first citations 3 to 8 weeks after you publish well-structured content, because LLMs reindex faster and don't depend on accumulated backlinks. It's also more volatile: every model update (GPT-4o to GPT-5, Claude 3.5 to 4, Gemini 2.0 to 2.5) can reshuffle the citations. It's an ongoing discipline, not a one-off campaign.

Closing

GEO doesn't replace classic SEO. It complements it and, in many verticals, is starting to carry more commercial weight. What sets it apart from the 2015 wave is who decides which content has value. It used to be an algorithm with 200 known factors. Now it's a generative model whose internal criteria are opaque but whose working preferences can already be measured: well-named entities, valid schema, primary sources, literal questions.

The good news is that almost everything you need to do can be applied in an afternoon per post. The bad news is that if your content architecture was designed around keywords and the 2018 SERP, the move to GEO isn't cosmetic. You'll need to revisit templates, blog components, schema generation, your sourcing policy and how you present identified authors.

It's not complicated. It's discipline.

If you want an operational review of your site, with an entity audit, a map of the missing schema and a ranking of the 10 pages with the most potential for generative citations, you can book a 30-minute session. If what you're after is building the GEO layer into your own system that runs content, clients and automations from one place, see how we approach your own vertical software for consultancies and professional firms that no longer want to run the business on top of three disconnected SaaS tools.

That's it.

Sources and references

• Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 (AI Act). EUR-Lex
• Gartner, "Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents", February 2024. Gartner Newsroom
• Semrush, "AI Traffic Study 2024", August 2024. semrush.com
• BrightEdge, "Generative AI Search Impact Report", January 2025. brightedge.com
• Schema.org, vocabulary reference. schema.org
• Google Search Central, "Structured data general guidelines". developers.google.com/search
• Perplexity, "How Perplexity works", public documentation, November 2024. perplexity.ai
• Law 18/2022 (Crea y Crece, Spain's business growth act), article 12 on mandatory electronic invoicing. BOE