DavidGadyan

GEO: 11 Techniques to Get Cited by AI

Generative Engine Optimization: eleven ways to get cited by ChatGPT, Perplexity, Claude and Gemini, from standalone passages and named entities to off-site authority and a citation eval run like LLMOps.

  1. Foundation

    Standalone sections

    The copy-paste test

    When it helps
    Sections start with "as we saw above" or only make sense after the previous one, and ranking on Google is not translating into AI mentions.
    What to do
    Paste any single section into a blank document. If it does not make sense and answer one clear question on its own, rewrite it until it does.
    Why it works
    Models chunk a page into passages and score each one alone; a passage that needs its neighbours is skipped or loses to a competitor's that stands alone. Google ranking is no proxy: 2025 research cited by KIME found only about 12 % of ChatGPT citations match a Google first-page URL.
    Example
    Before: "As shown above, this makes the second step cheaper." After: "Prompt caching makes the second request cheaper, because the repeated system prompt is read from cache instead of being billed as new input." Pasted alone, the sentence still says what, how and why.

  2. Foundation

    Name the entity, not "it"

    No pronouns in quotable lines

    When it helps
    The sentences most likely to be quoted say "it", "this" or "our platform" instead of who or what they are about.
    What to do
    Rewrite every citation-worthy sentence with its explicit subject: "KIME tracks nine AI engines", "David Gadyan builds AI agents", never "It tracks…" or "We build…".
    Why it works
    Passages are extracted verbatim or close to it; out of context a pronoun has no referent, so the citation loses its subject.
    Example
    Before: "It integrates with the charting tool and keeps the trader on plan." After: "The AI Trading Assistant integrates with the trader's charting tool and keeps the trader on the plan." Quoted on its own, the second version still names the product.

  3. Foundation

    Fact-dense sentences

    Number, source, date

    When it helps
    Claims read like "many clients see improvement" or "a strong track record".
    What to do
    Give every claim that could be fact-checked its number, its source and its date, in the same sentence: "190+ projects since 2018", not "lots of experience".
    Why it works
    Models favour passages with specific numbers, named entities, dates and sources, and skip the vague version of the same idea.
    Example
    Before: "The assistant answers fast and customers love it." After: "The assistant answers in under 2 seconds at the 95th percentile (load test of 10 March 2026, 500 concurrent users) and scored 4.7/5 in the April customer survey." Number, date and source in the sentence itself.

  4. Foundation

    Evidence beside the claim

    Citations, stats, quotes

    When it helps
    The statistics live on an "about" page or in a footnote, far from the sentence they support.
    What to do
    Place the statistic, its source and any expert quotation directly next to the claim they support, as visible text, so the evidence block is extracted together with the claim.
    Why it works
    In the Princeton-led GEO study, adding source citations, statistics and quotations were the three edits that raised visibility in generative-engine answers the most.
    Example
    Before: "Structured content gets cited more." (source in the footer) After: "Pages with comparison tables are cited 4.2× more often than prose, according to KIME's 2026 guide, which reports a 2025 analysis of 10,000 AI citations." Claim, number, source and date in one sentence.
    Reported lift
    +40 % visibility with citations, +37 % with statistics, +30 % with quotations (published study)

  5. Foundation

    Consistent terms, repeated plainly

    One name per concept

    When it helps
    The same service is called three different things across the site, and the core message appears once, buried in the middle.
    What to do
    Pick one term per concept and use it in headings, body, anchor text and schema; restate the core idea in plain words in more than one section of the page.
    Why it works
    Repeating the core idea across sections raises a model's retrieval confidence and reduces partial or inaccurate summaries; consistent names make entity relationships explicit.
    Example
    Pick "AI agent" and keep it: in the H1 ("AI agents for operations teams"), in the body, in the anchor text ("see the AI agent case studies") and in the schema (about: "AI agents"). Not "agentic workflow", "autonomous assistant" and "copilot" on the same page.

  6. Foundation

    Crawlable, server-rendered content

    robots.txt, HTML, llms.txt

    When it helps
    Key text only appears after a click, a tab or a JavaScript render, or robots.txt blocks the AI crawlers you want.
    What to do
    Allow the AI crawlers you want in robots.txt, render the important text in the HTML, confirm headings, links, tables and schema are visible without interaction, and publish an llms.txt index of your key pages (this site does).
    Why it works
    Orphaned, duplicated or JavaScript-hidden content is less reliable for retrieval and citation; crawler access is part of the GEO foundation layer, and nothing else works without it.
    Example
    robots.txt: User-agent: GPTBot, Allow: /, and the same for ClaudeBot and PerplexityBot. An llms.txt at the site root listing the key pages with one line each (this site: /llms.txt). The FAQ answers are in the HTML, not loaded on click.

  7. Foundation

    Visible freshness

    Published, updated, dated

    When it helps
    Pages carry no date, or 2026 advice sits under a 2023 date.
    What to do
    Show published and updated dates, refresh the facts and the date together when something changes, and keep dateModified in the schema in sync.
    Why it works
    Freshness is part of the GEO foundation layer. Perplexity reflects changes fastest because it pulls live from the web; ChatGPT follows more slowly through training snapshots and index refreshes.
    Example
    Above the article: "Published 21 Apr 2026 · Updated 2 Nov 2026". Schema: datePublished 2026-04-21, dateModified 2026-11-02. When a price or a figure changes, both dates change the same day.

  8. Foundation

    One canonical page per topic

    Hub + descriptive anchors

    When it helps
    Three pages half-answer the same question, and internal links say "read more".
    What to do
    Give each topic one canonical page, link the supporting pages to it with anchor text that names the topic, and group related pages into clusters around a hub.
    Why it works
    Models learn which page is the canonical source for a concept from the link graph and the anchors; duplicated pages split the signal and make the citation ambiguous.
    Example
    Hub: /materials/geo/ is the one page about GEO. Supporting posts link to it as "the GEO technique map", and the hub links back to each post by its topic. One canonical URL, no second "AI SEO" page saying the same thing.

  9. Amplification

    Off-site presence

    Directories, lists, reviews, media

    When it helps
    Your own pages are extractable, but assistants still name competitors when asked who to hire.
    What to do
    Get listed in the directories and "best of" lists for your category, collect public reviews (Upwork, Google), and earn mentions in the media the models already cite.
    Why it works
    Amplification is GEO layer 2: it makes the brand trusted. It only pays off once layer 1 gives the model a passage to pull from your own site.
    Example
    For a freelance AI engineer: an Upwork profile with public reviews, a listing in a developer directory, a place in a "best AI developers in Barcelona" list, a conference talk page naming the same services. Each one is a third-party page a model can cite about the same entity.

  10. Measurement

    Measure citations like an eval

    Fixed prompts, weekly, versioned

    When it helps
    Pages were restructured, but nobody knows whether ChatGPT, Perplexity or Gemini cite them now.
    What to do
    Keep a fixed set of buyer prompts, run it weekly against each engine, and record which URL is cited, the share of voice and the sentiment. Treat it as an LLMOps eval set: same prompts, versioned content, regressions noted.
    Why it works
    Measurement is GEO layer 3; without it you cannot tell which restructuring worked. Most pages show a measurable shift within 30–45 days.
    Example
    A sheet with twenty prompts ("Who can build an AI agent for a trading desk?", "Expert-Vetted AI engineer in Barcelona"…), one row per engine per week: cited URL, position, sentiment. A restructured page still uncited after six weeks goes back to the audit.
    Reported lift
    Measurable change in 30–45 days (estimate)

  11. Measurement

    Start with ten pages

    One page a day

    When it helps
    The site has 300 pages and the plan is to rewrite all of them before measuring anything.
    What to do
    Pick the ten pages with the highest strategic value (pricing, comparisons, category pages, top-of-funnel research), restructure one a day for two weeks with the 7-step audit, measure, then expand to the next tier.
    Why it works
    Most pages do not need a rewrite, only an audit: question per H2, answer-first paragraph, one-idea paragraphs, tables, numbered steps, named entities, valid HTML.
    Example
    Week 1: the pricing page, the two comparison pages, the two top case studies. Week 2: the service page, two category pages, two research posts. One page a day with the 7-step audit, measured from day one.
    Reported lift
    30–60 minutes per page (estimate)

Pick a technique on the map to see when it helps, what to do and why it works.

Sources: KIME, "How to Structure Content for LLM Extraction: A GEO Guide for 2026" (Apr 2026) for the three-layer framework, the 30–45-day window and the 12 % figure; Victoria Olsina, "Content Structure for SEO and LLMs (2026)" (updated Oct 2026) for the Princeton-led GEO study figures and the 8-step checklist. Running the measurement as an LLMOps eval is David Gadyan's own practice.

How do LLMs rank content?

Large language models cite some formats far more than others. Relative citation rate by content format, from the 2025 analysis of 10,000 AI citations reported in KIME's 2026 GEO guide; an unstructured prose paragraph is the 1.0× baseline.

Content formatRelative citation rateBest for
Table4.2×Multi-attribute comparisons, pricing, feature matrices
Answer-first paragraph3.1×Definitional and conceptual questions
Numbered list2.7×Step-by-step processes, ranked recommendations
Bullet list1.8×Feature enumeration, unordered options
Prose paragraph (unstructured)1.0× baselineNarrative or editorial framing

Source: KIME, "How to Structure Content for LLM Extraction: A GEO Guide for 2026" (Apr 2026), reporting a 2025 citation-pattern analysis. Multipliers are relative to unstructured prose on the same data.