What LLM SEO Really Means, and What the Evidence Says Works

One discipline, several names

Search the phrase "LLM SEO" and you will find a dozen guides insisting it is a brand-new field. It is not new so much as newly named. Practitioners use LLM SEO, GEO (generative engine optimization), AEO (answer engine optimization), and AI SEO almost interchangeably, and every one of those labels points at the same job: getting a large language model to mention, quote, or cite your business when it writes an answer for a potential customer.

At Formative Digital we file all of it under AI search engine optimization, the umbrella term for the full discipline across every engine and surface. That hub page covers the strategy layer. This page goes one level down, into the machinery the "LLM" in LLM SEO refers to: how a language model comes to know anything about your business at all, which optimization methods have actual experimental evidence behind them, and how you tell whether the work paid off.

The distinction matters because most advice in this space is written engine-first: tips for ChatGPT, tips for Perplexity, tips for AI Overviews. The engines change quarterly. The model mechanics underneath them change slowly, and once you understand those mechanics the engine-specific tips mostly write themselves.

The two paths your content takes into a model

A language model can only tell a customer about your business if information about your business reached the model. That happens through exactly two channels, and they behave so differently that treating them as one thing is the root error behind most failed AI-visibility efforts.

Path one is training. Before a model ships, it is trained on an enormous snapshot of text: web crawls, licensed archives, reference corpora. Whatever the crawl contained about your business as of the training cutoff is baked into the model's parameters. This knowledge is durable and always available, even offline, but it is frozen. You cannot edit it, and it will not refresh until the provider trains again.

Path two is retrieval. When an engine like Perplexity or ChatGPT Search handles a question, it fetches live web pages at answer time and writes from them, attaching citations to the pages it used. This knowledge is current and editable, because the pages being read are yours to change, but it only applies when the engine decides the question warrants a lookup. We walk the retrieval pipeline stage by stage in how AI search works.

Both paths matter, and they matter differently. Training presence shapes what a model says about you when it does not search: whether it recognizes your brand name, associates it with the right category, and repeats accurate facts. Retrieval presence decides whether you get cited, with a visible link, in the answers that carry buying intent. A business with only training presence gets vaguely remembered; a business with only retrieval presence gets cited on some prompts and is a stranger on the rest. The work plan has to feed both.

What you can realistically do about training data

The training path frustrates marketers because there is no submission form. You cannot upload your brand into a model. What you can do is make sure the corpora that providers crawl contain accurate, consistent, well-distributed information about you before the next training run happens.

In practice that means three workstreams. First, entity consistency: your business name, category, location, and core claims should read identically across your site, your directory listings, your social profiles, and every third-party page that describes you, because a model learns an entity by seeing the same facts repeated across independent documents. Second, breadth of footprint: mentions in industry publications, local news, supplier pages, and community sites all become training text, and a brand that exists in fifty independent documents is learned far more reliably than one that exists only on its own domain. Third, patience with the calendar: model releases arrive on the provider's schedule, so training-side work compounds over quarters, not weeks.

This is Vector 2 in our methodology, Anchor: establish the entity so thoroughly and consistently that any system reading the web at scale arrives at the same picture of who you are.

What retrieval-time optimization actually is

The retrieval path is where day-to-day LLM SEO lives, because it responds to changes within a crawl cycle. When an engine searches on a user's behalf, your page is competing in three sequential contests: getting fetched into the candidate pool, getting selected as a source, and getting quoted in the finished answer.

Each contest has its own levers. Fetching depends on plain technical access: server-rendered HTML, no crawler blocks against the AI user agents you want, presence in the search indexes some engines use as their candidate supply. Selection depends on relevance and trust: does a specific passage on your page answer the question directly, and do your entity signals check out against the rest of the web. Quoting depends on extractability: a paragraph that states its point in the first sentence, carries a concrete figure or date, and stands alone without needing the surrounding page is far easier for a model to lift accurately than a meandering one.

Notice what is absent from that list: tricks. There is no header value, no hidden text pattern, no magic ranking dust for language models. The levers are all versions of the same instruction: publish passages worth quoting and make them effortless to find and attribute.

The tested methods: what the Princeton experiments found

Almost every claim in the LLM SEO conversation is vibes. The useful exception is the 2023 benchmark study by Aggarwal and colleagues at Princeton and IIT Delhi (arXiv:2311.09735), the paper that coined the term GEO. The authors built a 10,000-query benchmark, applied nine different content modifications to real web sources, and measured how each changed the source's share of the generated answer. They framed the problem honestly: site owners have "little to no control over when and how their content is displayed" by generative engines, so the only rational move is to test what shifts the odds.

Three methods clearly won. Adding quotations from credible sources, adding statistics in place of vague claims, and adding citations to authoritative references each lifted a source's visibility in answers substantially, with the paper reporting gains of up to 40 percent on its position-adjusted word count metric. Fluency and readability edits produced smaller but real gains, which tells you the reading model rewards clean prose, not just facts.

Two findings from the same experiments deserve more attention than they get. Keyword stuffing, the classic SEO reflex, did nothing and sometimes hurt: the reading model is not counting keyword occurrences. And the gains were not distributed evenly. Sources ranked fifth in the underlying search results gained over 100 percent visibility from citation-style edits while the top-ranked source actually lost share, meaning the techniques disproportionately help the smaller sites that classic search buries. Our full breakdown of the study, including its limits, is in the GEO paper explainer; the surrounding concept gets a plain-language treatment in what is GEO.

The myths: there is no LLM meta tag

Because the field is young, folklore fills the gaps. Three claims come up in almost every sales call, and all three are wrong.

Myth one: a meta tag can request AI citations. No provider, not OpenAI, not Google, not Anthropic, not Perplexity, reads any tag that asks a model to include or favour your content. The only markup-level controls that exist are exclusionary: robots.txt rules and meta directives that keep crawlers out. You can opt out of AI systems. You cannot opt in by declaration.

Myth two: llms.txt is the new sitemap. The proposed llms.txt standard, a markdown index of your site for AI agents, has genuine uses for developer documentation, where tools such as coding assistants fetch it deliberately. As a visibility tactic it is inert. Google's John Mueller publicly compared it to the keywords meta tag, Google has said it does not use the file, and a 2026 analysis of large-scale bot traffic found the major AI crawlers overwhelmingly skip it and read HTML directly (Search Engine Land, February 2026; Search Engine Journal, 2026).

Myth three: AI visibility is a paid placement away. No major assistant currently sells citation slots in organic answers. Where ads exist, they are labelled and separate. Anyone selling guaranteed placement inside model answers is selling something they do not control.

The honest position, and the one we build on, is that influence over model answers is probabilistic. Truth, not tricks: the levers that exist are the ones the experiments above validated, and they raise odds rather than purchase outcomes.

Structure the machines can parse: a working example

One lever sits between the myth pile and the content advice: structured data. Schema markup does not command a model to cite you, but it removes ambiguity about who is speaking, which matters at the source-selection stage where engines cross-check entities. This is Vector 6, Structure, applied to the LLM problem. A minimal, correct block for a service business looks like this:

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "ProfessionalService",
  "name": "Example Consulting",
  "url": "https://example.ca",
  "areaServed": {
    "@type": "AdministrativeArea",
    "name": "Ontario"
  },
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Brantford",
    "addressRegion": "ON",
    "addressCountry": "CA"
  },
  "sameAs": [
    "https://www.linkedin.com/company/example-consulting"
  ]
}
</script>

Twenty lines that settle name, location, service area, and canonical identity in a format every crawler parses the same way. Pair it with matching facts on your profiles and the entity question stops costing you selections. If you want a second opinion on whether your current markup validates, send us the URL and we will look.

Measuring LLM SEO without fooling yourself

Measurement is where LLM SEO differs most from the classic discipline. There is no rank tracker of record, because there is no rank: the same prompt can produce different answers across engines, sessions, and days. What you can build instead is a three-layer measurement stack.

Layer one: prompt sampling. Write down the 20 to 40 questions a real buyer would ask an assistant about your category. Run them through ChatGPT, Perplexity, Gemini, and Google AI Overviews on a fixed schedule. Log mentions, citations, and which competitor got the slot when you did not. Repeated sampling smooths out run-to-run variance and turns a noisy signal into a trend line. This is Vector 11, Measure, and it is the single highest-value habit in the discipline.

Layer two: referral analytics. Traffic from chatgpt.com, perplexity.ai, gemini.google.com, and copilot.microsoft.com shows up in analytics with identifiable referrers. The volumes look small next to organic search, so judge them on conversion rate instead; a visitor arriving from an answer that already recommended you is deep in the funnel.

Layer three: attribution by asking. A large share of AI influence never produces a click, because the user reads the answer and later searches your name or calls directly. The correction is unglamorous: ask every new lead where they first heard of you, and log it. Businesses that skip this layer systematically underestimate what the machines are sending them.

Where LLM SEO fits in the rest of the plan

None of this replaces conventional search work, and the dependency runs deeper than "do both." Several engines source their candidate pages from the same indexes classic SEO fills, so a site that cannot rank also struggles to get retrieved. Meanwhile the durable assets, consistent entities, cited claims, and passage-level answers, improve both surfaces at once. The practical sequencing we use with clients: fix technical access first, then rebuild the highest-intent pages for extractability, then broaden the entity footprint, then measure and iterate against the prompt sample. The deeper technical material behind each step lives in the research library.

Scale expectations honestly. This is a probabilistic channel measured in share of answers, not a switch you flip, and outcomes vary with industry, competition, and the state of your existing web presence. What the evidence supports is narrower and more useful: specific, tested edits change how often models quote you, the changes favour smaller sites, and the businesses tracking their prompt sample are the only ones who actually know their position.