The term Generative Engine Optimization was established in November 2023 by a research group around Pranjal Aggarwal (IIT Delhi) and Karthik Narasimhan (Princeton University) in the paper "GEO: Generative Engine Optimization" (arXiv 2311.09735, published at KDD 2024). The research tested which content manipulations increase a source's visibility in generative answers — and produced the first quantified leverage effects.

In day-to-day SEO practice the term is often used as a buzzword without the technical foundation being understood. That is a problem, because GEO follows a different logic than classical SEO. Treating both as equivalent leads to addressing the wrong levers.

The fundamental difference: retrieval vs. inference

Classical SEO optimizes a process with three steps: crawl → index → rank. Google collects pages, stores them in a structured way, and orders them by relevance and authority for a given query. The output is a list of documents.

Generative engines work differently. Their process is: retrieve → synthesize → generate. The system pulls relevant passages from various sources (RAG — Retrieval Augmented Generation) or draws on knowledge stored in its own training corpus, combines the signals via semantic embeddings, and generates a new answer. The output is text — not a list. For Google there is an important caveat: according to Google’s guide to generative AI features (as of July 10, 2026), AI Overviews and AI Mode retrieve their sources through Search’s core ranking systems. Classical SEO remains the entry ticket there.

That distinction has consequences that are routinely underestimated in SEO thinking:

What did the Princeton study on GEO find?

The Princeton study shows that a source’s visibility in AI answers can be raised deliberately, mainly through evidence-backed content. The team led by Pranjal Aggarwal and Karthik Narasimhan tested nine optimization methods on GEO-bench, a benchmark of 10,000 queries from different domains. The three strongest methods, adding citations (Cite Sources), adding quotations (Quotation Addition) and adding statistics (Statistics Addition), increased a source’s position-adjusted word count in the generated answer by a relative 30 to 40 percent. More fluent and easier-to-understand text added 15 to 30 percent. An authoritative, persuasive tone, by contrast, showed no significant improvement, and keyword stuffing achieved little: in the test on Perplexity it performed about 10 percent worse than the unmodified baseline. The work is a controlled experiment on its own benchmark, not a measurement of production systems such as ChatGPT or AI Overviews (Aggarwal et al., KDD 2024).

+30-40%

from citations, quotations and statistics (position-adjusted word count)

+15-30%

from more fluent, easier-to-read text

±0

authoritative tone without significant effect, keyword stuffing without benefit

The three structural GEO levers

The study measures text changes to individual sources. In consulting practice, three structural levers come on top that go beyond the single text. They are our working model (evidence level D), based on our own prompt tests, not results of the Princeton study:

Lever 1: entity clarity

Internally, LLMs represent entities as vectors in a semantic space. A brand recognized as a clearly defined entity (via Wikidata, Wikipedia, Knowledge Panel, structured data, consistent naming) is systematically included in answers more often. A brand that exists as a diffuse mix of name variants disappears into the noise.

In practice this means:

Lever 2: semantic co-occurrence

LLMs learn associations statistically. When the brand "Murat Ulusoy" appears regularly in high-quality text corpora close to terms like "SEO strategy", "AI visibility", "reputation management" and "international scale", that co-occurrence hardens inside the model. For a relevant prompt query the brand is activated as a candidate with higher probability.

The consequence: the distribution of brand mentions matters more than raw quantity. A steady signal in trade media, academic sources, podcast transcripts and recognized industry portals outweighs 1,000 thematically irrelevant mentions.

Lever 3: trust density

Internally, LLMs weight sources by perceived authority. That weighting rests on several signals:

High trust density emerges when a topic is presented consistently across multiple independent, qualified sources. Contradictory or isolated signals are weighted as uncertain and used in answers less often.

"SEO was a contest for positions. GEO is a contest for memories — in the synthetic memory landscape of models that half the world is asking for answers."

What GEO is not

A few popular misconceptions that cause damage in practice:

GEO ≠ keyword stuffing with AI terms

Content that forcibly mentions "ChatGPT", "AI" and "LLM" is not preferred by models. The opposite: overfitting on buzzwords degrades coherence and reduces citability.

GEO ≠ implementing llms.txt

The "llms.txt" file is a proposal for a Markdown overview of a website’s key content. In its guide of July 10, 2026, Google made clear that Google Search does not use such files: they neither harm nor help. There is no official confirmation so far that OpenAI or Anthropic use llms.txt for source selection. Sources are evaluated through content and entity signals, not via a meta file.

GEO ≠ SEO in new clothing

Classical on-page tactics (title tag optimization, meta description, heading structure) remain relevant — but they are not GEO. GEO addresses a different layer: how knowledge is represented inside a model.

Operator insight

The measurement category that changes everything

The most important shift for marketing organizations is the introduction of prompt audits: systematic tests with 300-1,000 relevant prompts against ChatGPT, Claude, Perplexity and Gemini. The measurement: whether your brand appears as an answer, how it is presented and which competitors are named instead. For GEO this data is what ranking reports were to classical SEO — but orders of magnitude more meaningful as a signal of market relevance.

A practical GEO assessment in five steps

  1. Entity audit: Is there a Wikidata entry? Is the Knowledge Panel maintained? Is the schema markup correct?
  2. Prompt test: Run 100 relevant prompts against four LLMs and document how the brand is treated.
  3. Co-occurrence mapping: Which terms appear in high-quality sources alongside the brand? Which do not — even though they should?
  4. Trust source gap: Which authoritative sources (trade media, studies, Wikipedia, industry reports) do not mention the brand? Which competitors dominate them?
  5. Content passage readiness: Are the top content passages structured as citable units (clear definitions, numbers, short paragraphs)?

The mathematics of inference: how LLMs choose sources

Anyone running GEO strategically must understand how a model evaluates sources internally. As a thinking model (simplified, explicitly not a documented vendor mechanism, but operationally useful), the process can be expressed as a retrieval-generation scoring formula:

P(source | query) = softmax( α · sim(q, p) + β · auth(d) + γ · ent(e, q) + δ · fresh(t) )

with:
sim(q, p)  = cosine similarity between query embedding and passage embedding
auth(d)    = domain authority weight (training-based, not equal to PageRank)
ent(e, q)  = entity alignment between page entities and query entities
fresh(t)   = freshness decay function (often exp(−λ·Δt))
α, β, γ, δ = model-specific hyperparameters

The production pipelines of ChatGPT, Claude, Gemini and Perplexity are proprietary and not open to inspection. The formula is the working model behind our measurement methodology, not a documented Google or OpenAI mechanism. Practically: four levers must be served simultaneously. Optimize only one (for example, sim through keyword density) and the softmax barely moves — because the score is beaten by other sources on other dimensions.

The embedding space: where brands really compete

An LLM represents every brand as a vector in an embedding space. Vendors’ internal representations are not visible; in our methodology we approximate the position with open embedding models (such as Sentence-BERT, typically 768 to 4,096 dimensions). In that space, every brand has a neighbourhood cluster: semantically related terms that sit close to it. According to our working model, that cluster decides whether the brand is activated for a query.

A practical example: for a B2B SaaS client we analysed the embedding position of the brand in Sentence-BERT. The nearest neighbours were "CRM software", "customer management", "sales pipeline". On prompts about "automation" or "AI integration" the brand never surfaced — although it offers both capabilities strongly. The reason: the training corpus had not seen the brand in those contexts.

The solution was not new SEO. It was corpus work: 14 trade articles in external publications featuring the brand in co-occurrence with the missing terms. Three months later the brand appeared on automation prompts in 37 % of tests. Before: 2 %. This is a single project case from our own data (evidence level C), measured with the methodology of our LLM citation benchmark, not a general effect.

768-4,096

Dimensions of the open proxy embedding models in our analysis

3-8

Neighbourhood clusters a strong brand covers

14

Qualified co-occurrence publications needed to shift the signal

Tutorial: a six-step entity consolidation workflow

Entity clarity cannot be produced on the side. It is a dedicated workflow with a strict order of operations. Order matters — starting with schema before Wikidata is clean produces inconsistent signals.

Step 1 — create a name authority file

A single canonical internal document containing the brand, its official spellings, legal form, founding date, founders, headquarters, sub-brands, key products and board members. Everything downstream references it.

Step 2 — Wikidata entry

A Wikidata item with at least 15 valid properties: P31 (instance of), P571 (inception), P159 (headquarters), P452 (industry), P856 (official website), P1448 (official name), P2002 (X username) and so on. Back every claim with an independent source, otherwise the item risks deletion.

Step 3 — Wikipedia article (when notability is reachable)

Not every brand qualifies. What counts are Wikipedia’s notability guidelines. Our rule of thumb: at least three substantive reports in independent trade media. The article must meet Neutral Point of View, verifiable sources and encyclopedic relevance. Not a PR piece.

Step 4 — consolidate Schema.org

On the homepage, Organization with @id, url, logo, sameAs array (Wikipedia, Wikidata, LinkedIn, Crunchbase, Twitter). @id must be identical across every page — the critical mistake most brands make.

Step 5 — author entities

For every published author of the brand, a Person schema with its own @id, sameAs to official profiles and knowsAbout with 5-8 expertise entities. Link to the organization entity via worksFor.

Step 6 — distribution

Distribute the entity across 20+ independent profiles: LinkedIn, Crunchbase, Bloomberg, Glassdoor, industry registers, professional-association member lists. Every profile must contain the canonical name and website exactly (NAP consistency for entities).

Co-occurrence engineering: the underrated discipline

Co-occurrence does not happen by chance — it is built. The methodical approach has four phases:

1. Gap analysis. We use a custom tool that queries the training-corpus approximation via Common Crawl plus news APIs and quantifies co-occurrence frequencies between the brand and a target-term list. Result: a heatmap with a target/actual comparison per concept.

2. Define target publications. Only media with high "trust weight" in LLM training data count. Our internal whitelist covers 340 German-language publications. In our internal source weighting, a piece in a tier-1 outlet such as the Süddeutsche Zeitung counts many times more than a piece in an unknown industry blog.

3. Write co-occurrence briefings. Every guest post, press release and podcast appearance is accompanied by a briefing that names the co-occurrence targets. Concretely: the brand must appear in the same paragraph as the target terms, not scattered across the text.

4. Reinforcement cycle. Every 90 days, retest prompt visibility. Concepts with stagnant presence receive a second wave of publications. The cycle runs at least four quarters until the models are updated (training cutoff + refresh).

Measuring trust density: the multi-source agreement score

Trust density is not a gut feeling. It can be quantified through the MSA (multi-source agreement score), our internal measurement convention whose thresholds come from our own prompt tests, not from vendor documentation:

MSA = (n_consistent_sources / n_total_sources) × log(n_total_sources) × w_quality

with:
n_consistent_sources  = number of sources with a consistent representation of the fact
n_total_sources       = total number of sources mentioning the fact
w_quality             = weighted quality factor of the sources (0.1-1.0)

In our prompt tests, an MSA > 3.5 signals that LLMs reproduce the fact with high confidence. An MSA < 1.5 means there that the model "is not sure" and will use the fact less often in answers, or with hedging ("according to company information…").

Anti-patterns: what regularly destroys GEO strategies

Conclusion

GEO is not an extension of SEO. It is a different discipline with different mathematics, different signals and different success metrics. Organizations that master both worlds in parallel — classical SEO for classical Google search, GEO for generative engines — will dominate the visibility economy over the next five years.

The others become sources that are still crawled — but no longer selected.

Sources