Generative Engine Optimization Past the Hype: What GEO Actually Is

TL;DR: Generative engine optimization is the working name for a set of practices aimed at increasing the rate at which large-language-model-driven search interfaces cite a publisher's URL inside a generated answer. The measurable surface is narrow but real: AI Overviews citation patterns in Google, Perplexity source lists, ChatGPT search source attributions, and a handful of similar interfaces. The early academic and practitioner work suggests that a small set of content properties (entity clarity, citation density, structured statements, alignment with the question's literal phrasing) correlate with elevated citation rates, but the studies are small, the systems are non-stationary, and the magnitudes are modest compared to the variance in classical SEO. What practitioners can productively do today is a constrained subset of what the marketing literature claims.

A note on the named sources. Aleyda Solis, Lily Ray, Kevin Indig, Cyrus Shepard, and the BrightEdge team appear throughout as the most-cited public practitioners writing on this question through 2024 and into 2025. Reference is made to the published "GEO" paper (Aggarwal et al., 2023), the academic IR literature on retrieval-augmented generation, and the citation-pattern research from BrightEdge, seoClarity, and Authoritas. Quantitative claims framed as advisory observation come from anonymized partner publishers and brand operators, not from the named tools or named retailers.

---

The Working Definition and What It Is Not

The term "generative engine optimization" entered the public conversation in 2023, both through a paper by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande titled "GEO: Generative Engine Optimization" (arXiv 2311.09735), and through the parallel commercial discourse around AI Overviews, Perplexity, and ChatGPT's search-augmented interfaces. The Aggarwal paper proposed a controlled benchmark for measuring how content modifications affect visibility inside generative answers, and the paper's "visibility" metric (a combination of citation presence and citation position-weighted impression share) has become a reference point for subsequent work.

The working operator definition, distilled from the published practitioner writing and the empirical studies, is that GEO is the practice of producing content properties that increase the rate at which LLM-driven search interfaces include the publisher's URL inside a generated answer's citation list. The unit of success is not a ranked position in a SERP; it is an inclusion event in a generated response.

What GEO is not, in the operator-relevant sense, is a replacement for classical SEO. The generative interfaces that cite URLs still depend on a retrieval step that draws from a search index, and the index is largely the same index that produces classical search results. Pages that do not rank in the underlying retrieval do not appear in the candidate set for citation; pages that rank well in the underlying retrieval are oversampled in the citation pool. The classical SEO discipline (technical health, content quality, link graph, structured data) remains the upstream gate. GEO sits downstream, on the question of which of the already-retrievable URLs gets cited.

The Measurable Surface Area

The list of generative-search interfaces with meaningful query volume is shorter than the trade discourse suggests, and the measurement infrastructure for each is uneven. The honest map of where GEO can be measured at the start of 2025:

The honest implication is that the bulk of measurable GEO work centers on AI Overviews and Perplexity, with ChatGPT search and Copilot as secondary surfaces, and that the tracking is genuinely immature compared to classical SEO tracking. The Search Console interface (the only place where Google reports clickthrough data from AI Overviews to the publisher) does not separate AI Overviews from other organic clickthroughs in its default reporting, and the third-party tooling (Authoritas, Semrush, BrightEdge, seoClarity, Ahrefs) that attempts to track AI Overviews citation does so by scraping the SERP and parsing the AI Overviews block, which is a fragile dependency.

What the Empirical Studies Suggest

The Aggarwal et al. paper (2023) ran a controlled experiment on a benchmark of queries fed through several open-source generative search pipelines. They tested nine content modifications (citation insertion, quotation insertion, statistic addition, fluency optimization, easier-to-understand phrasing, authoritative language, simple words, technical terms, keyword stuffing) against the baseline of unmodified content. The reported headline finding was that the citation insertion and quotation insertion modifications produced the largest visibility lifts (roughly 30 to 40 percent in the paper's metric on average across query types), and that the fluency and easier-to-understand modifications produced smaller but consistent lifts. The keyword-stuffing modification produced a small negative effect. The result is widely cited in the practitioner literature, sometimes uncritically.

The honest reading of the Aggarwal paper is more constrained. The benchmark queries were synthetic and may not represent the distribution of real search queries; the generative pipelines were open-source models (and at the time of the paper, models in the GPT-3.5 generation rather than the production search-augmented systems running in AI Overviews or Perplexity); and the visibility metric was a researcher-defined construct rather than a real publisher's clickthrough share. The directional findings are plausible and consistent with subsequent practitioner reports, but the magnitudes should not be projected to real-world production systems.

The subsequent practitioner work has been more directly grounded in production systems. The BrightEdge AI Overviews research, published in segments through 2023 and 2024, analyzed which URLs appeared in AI Overviews citations across thousands of tracked queries and reported persistent patterns: a heavy concentration of citations in domains already ranking in the top three of classical Google results (the "AI Overviews favor incumbents" finding), a category-skew in which informational and how-to queries triggered AI Overviews more often than commercial-intent queries (the "AI Overviews avoid commercial verticals" finding), and a slowly evolving share of brand-name domain citations as the system tuned through 2024 (the "AI Overviews stabilization" finding). The work is informative but limits its inference because it is observational rather than experimental.

The Kevin Indig writing on the same surface, published through Growth Memo and adjacent venues, has emphasized the operator-relevant patterns: that the citation rate inside AI Overviews is non-stationary (Google has visibly tuned the system across multiple iterations through 2024), that the citation behavior differs by query intent in stable ways (informational queries cite more frequently than transactional queries), and that the format of the cited content (whether the source has a clearly stated answer in the first portion of the page) correlates with citation rate.

The pattern in the chart is stable across the published tracking studies and matches the directional reports from Authoritas and seoClarity through 2024. The implication for operating priorities is that GEO investments concentrate on content that addresses informational and definitional queries, where AI Overviews appear and where citation is the achievable outcome, rather than on commercial-intent content where AI Overviews do not displace classical results.

The Content Properties That Correlate with Citation

The cross-cutting finding across the Aggarwal paper, the BrightEdge tracking, and the practitioner case studies from Lily Ray, Aleyda Solis, and Cyrus Shepard is that a small set of content properties correlate with citation. The list is short and contains items that good editorial practice would recommend independently of any AI-driven surface.

The first property is a clearly stated answer in the opening of the page. The generative systems retrieve a passage from the source and assemble the answer; passages that contain a complete, attributable statement of the answer (rather than a meandering setup) are favored. Pages that bury the lede are systematically under-cited.

The second property is entity clarity. Pages that name the subjects and objects of their claims explicitly (the actual people, products, places, terms involved), rather than using pronouns or vague references, are easier for the retrieval and reasoning layer to use. This is the same pattern that knowledge-graph extraction has favored for a decade.

The third property is citation density. Pages that cite their own sources visibly (with named studies, named authors, named datasets) are favored over pages that make claims without sourcing. The pattern is not subtle, and it aligns with the Aggarwal paper's strongest reported finding.

The fourth property is structural cleanliness. Pages with clear headings, parseable lists, defined terms, and tables that mark their rows and columns are easier to retrieve and quote. The retrieval layer can pull a clean paragraph more easily than a passage embedded in a long unstructured prose block.

The fifth property is alignment with the literal phrasing of the question. Pages whose headings and opening sentences match the words and phrasing that a real user would use to ask the question are favored over pages that use jargon or branded language. This effect is observable in the citation patterns reported by Lily Ray and Aleyda Solis in their public analyses.

The honest reading is that the content properties that correlate with citation are the properties that good operator-facing editorial practice would aim at anyway. There is no separate "GEO trick" that elevates a page in the AI Overviews citation pool without making it a better page in absolute terms. The advice is therefore congruent with the rest of the SEO discipline, and the question of whether to "do GEO" reduces to whether to apply discipline to the editorial properties that the modern retrieval systems happen to favor.

The Retrieval-Augmented Generation Stack and Why It Matters

Understanding the systems underneath the citation-rate question helps frame what is and is not achievable. The architecture that drives AI Overviews, Perplexity, and ChatGPT search is a variant of retrieval-augmented generation (RAG), the term Patrick Lewis and colleagues at Facebook AI Research introduced in their 2020 paper "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." The basic pattern is to combine a retrieval module (which surfaces candidate passages from an external corpus given a query) with a generative module (which composes an answer conditioned on the retrieved passages). Modern production systems extend the basic pattern with multiple retrieval rounds, query rewriting, reranking, and grounding constraints.

The implication for GEO is that the generative system does not citation-pick out of a vacuum; it cites the passages that the retrieval module surfaced. The retrieval module is itself a search system, and the search system has its own ranking logic that overlaps with classical SEO. The candidate set is therefore upstream-determined by the search ranking; the citation decision is downstream and concerns which of the already-retrieved passages get included in the answer. The two stages have different drivers, and "GEO" properly refers to optimization for the second stage given that the first stage has succeeded.

The reranking step in production systems is where some of the harder-to-predict citation behavior lives. The reranker is a model that takes the candidate passages and scores them for their fit to the query, and the scoring criteria include relevance, source authority, factuality signals, and a set of properties that are not publicly documented. The reranker introduces some of the volatility in citation patterns that practitioners track: a page that ranks well in the underlying index but does not rerank well will not be cited; a page that reranks well will be cited even at modest underlying ranks. The reranker is opaque, and the GEO discipline operates without direct knowledge of its scoring function.

A useful operator model is to treat the candidate-set problem as the classical SEO problem (rankings, technical health, link graph) and the reranker problem as the GEO problem (content properties that the reranker appears to favor). The candidate set is the upstream gate; the reranker is the downstream filter. Both stages must succeed for a citation to occur. Investments in only one stage produce ceiling effects.

Why the Magnitudes Are Modest

The marketing discourse on GEO often suggests dramatic lifts; the empirical record is more modest. The reasons are worth dwelling on because they affect what publishers should expect.

First, the citation event itself is a low-frequency outcome relative to organic clickthroughs. A page that appears in 30 percent of relevant AI Overviews citations is in the citation pool for those queries; the underlying click rate from the AI Overviews surface to the publisher is small because the generated answer often satisfies the user's question without the click. Publishers tracking the downstream effect of GEO often find that the elevation in citation rate is real but the downstream click effect is much smaller than the citation lift would suggest.

Second, the systems are non-stationary. AI Overviews has been visibly tuned through 2023 and 2024; the citation behavior in a particular query window is not necessarily the citation behavior six months later. The half-life of a GEO test result, in our reading of the published tracking, is approximately one major model update at the underlying generative engine. Investments in measurement tooling that assume the surface is stable will struggle.

Third, the retrieval layer is biased toward already-ranking sources. A site that improves its content properties for GEO but does not rank well in the underlying retrieval will see a small citation lift; a site that ranks well already and improves its content properties will see a larger lift because the retrieval gate is open to it. The interaction effect favors operators who have done classical SEO work already.

Fourth, the citation list is short. AI Overviews typically cite three to five sources per generated answer; Perplexity and ChatGPT search cite five to ten. The total citation slots for a given query are scarce, and competition for the slots is intense in queries with many candidate sources. The competitive distribution is heavy-tailed: a handful of sources capture most of the slots in a given vertical.

The advisory framing we have settled on is that GEO is a defensive practice for operators who already do classical SEO well, rather than a growth lever in its own right. The defensive case is straightforward: if a competitor's content is being cited and the operator's is not, the share-of-voice on the generative surface is being lost. The growth case is weak because the click economics from the generative surface to the publisher are systematically less favorable than from classical SERPs.

The Click Economics of the Generative Surface

The single most important constraint on GEO as a growth lever is the click economics. The classical SERP delivers a click to the publisher with probability that decays with rank position, and the underlying click rate at position one for an informational query is well-documented to be in the 25 to 40 percent range depending on SERP features and query type. The generative surface delivers a click to the publisher with probability that is materially lower because the generated answer often satisfies the user's question without requiring the click.

The published click-rate data from the generative surface is thinner than for classical SERPs, but the directional finding from the available studies (the Authoritas reports, the seoClarity tracking, and a small number of published publisher case studies) is that the click rate from an AI Overviews citation to the publisher is in the low single digits for informational queries. The publisher is cited (which is a surface presence event) but the click is taken much less frequently than from a classical position-one ranking. The implication is that the population of click traffic shifts from organic SERP clicks toward AI Overviews appearances without clicks, with the net effect being negative for the publisher's total clicks per query in the segments where AI Overviews appear.

The downstream consequence for the publisher's analytics is that Search Console total clicks decline for queries that have triggered AI Overviews, even when the publisher's citation rate inside AI Overviews is high. The interpretation that practitioners often reach for ("AI Overviews are stealing our traffic") is partially correct, but the more precise framing is that AI Overviews have changed the contract: from a click-based contract (impression to click) to a citation-based contract (impression to citation, with click as a low-frequency outcome). The contract change favors users (faster answers) and the platform (engagement retention) at the expense of the publisher's per-query click count.

The defensive case for GEO sits inside this contract change. If the publisher's content is going to be summarized either way (whether or not the publisher invests in GEO), the choice is between being cited inside the summary or being unmentioned. Being cited preserves at least the citation traffic (the low-single-digit click rate, plus the brand-impression effect of being a named source) where being unmentioned forfeits both. The defensive logic does not require the click rate to be high; it requires the citation rate to be positive and competitive.

Structured Data and Knowledge Graph Adjacency

A consistent strand in the practitioner writing on GEO emphasizes structured data (schema.org markup, especially Article, FAQPage, HowTo, and Dataset) and knowledge-graph alignment as content properties that correlate with citation. The mechanism is plausible: the retrieval layer behind the generative interface uses the same indexing pipeline that uses structured data to surface rich results, and content with clean structured data is more easily extracted into the generated answer.

The honest caveat is that the published empirical evidence on structured data as a GEO lever is thinner than the practitioner consensus suggests. The Aggarwal paper does not directly test structured data as a treatment; the BrightEdge tracking has not isolated structured-data effects from the confounded content-quality effects; and the practitioner case studies that report a structured-data lift have not, in our reading, controlled for the other content changes that typically accompany a structured-data audit.

The operating recommendation, conditional on the limited evidence, is that structured data is worth maintaining for the standard SEO reasons (rich results, knowledge-graph clarity, search-feature eligibility) and that any incremental citation effect is a bonus rather than a primary justification. The operator effort of maintaining structured data is small enough that the case for doing it does not depend on the GEO uplift being large.

The knowledge-graph angle is more direct. The retrieval pipeline behind the generative interface is grounded in the same entity model that supports the knowledge graph; pages that map cleanly to entities (the company has a Wikipedia page, the author has a verified profile, the product has a clean entity record, the brand has a Knowledge Panel) are easier to assemble into answers. Entity hygiene is a recommendable investment for the structured-search surface and pays a parallel dividend on the generative surface.

Practical GEO Disciplines an Operator Can Actually Run

The constrained operator-facing recommendation, given the empirical record, is to run a small set of disciplines that produce content properties known to correlate with citation, while keeping classical SEO investment as the dominant work.

The first discipline is the "answer in the opening paragraph" pattern. For each informational page on the site, the opening paragraph should state the answer to the page's primary question in one to three sentences, with the relevant entities named and the supporting evidence summarized. The rest of the page can elaborate. The pattern is the inverse of the "build to the conclusion" structure favored in conventional editorial writing, and it requires editorial discipline to maintain.

The second discipline is the entity-naming audit. For each informational page, the pronouns and vague references should be resolved to named entities. "It" becomes the specific tool name; "the company" becomes the named company; "they" becomes the named research team or author. The exercise can be done on a large corpus mechanically with a careful edit pass.

The third discipline is citation insertion. For each meaningful claim, the page should cite the source visibly (an author, a study, a dataset, a public report). The citations should be linked where possible. The discipline aligns with the strongest finding from the Aggarwal paper and with the editorial standards that good practitioner publications maintain anyway.

The fourth discipline is the FAQ block at the bottom of informational pages. The block answers the three to five most-likely follow-up questions in a question-and-answer format that maps cleanly to FAQ schema markup. The pattern surfaces the page on multiple related queries and provides additional citation-eligible passages.

The fifth discipline is the dataset block where the page makes a quantitative claim. A small table with the data, with named sources and a methodology note, gives the retrieval layer a clean unit to pull and the operator a Dataset schema markup opportunity.

The sixth discipline is the periodic refresh. The non-stationary nature of the generative systems means that the citation patterns on a query shift over time, and the date freshness signal is observable in the citation patterns. Pages on time-sensitive topics benefit from a documented refresh cadence.

The chart shows the qualitative pattern we have observed in the limited dataset we have collected: a single-intervention (answer block in the opening) produces a modest lift; the full set of disciplines produces a larger lift; the no-intervention control drifts slightly down as competitors apply their own GEO discipline. The absolute levels are page- and vertical-specific, but the directional pattern has been consistent across the small number of cases where we have tracked it.

What Aleyda Solis, Lily Ray, and Cyrus Shepard Have Each Emphasized

The practitioner literature on GEO through 2024 converged around a small group of authors whose framings are worth distinguishing because they emphasize different aspects of the same underlying surface.

Aleyda Solis's writing on AI search and GEO, distributed through the SEOFOMO newsletter and a series of talks at conferences through 2024, has emphasized the international dimension of the question. The generative surfaces behave differently across languages and markets because the underlying training data is uneven; English-language queries are saturated with sources, while Spanish, German, Japanese, and Arabic queries draw from a much narrower candidate pool. The implication Solis has drawn is that the citation competition is materially lower in non-English markets, that brand-recognized publishers in those markets have an outsized opportunity to capture citation share, and that the international SEO discipline (hreflang, language-specific content, regional authority) translates directly into GEO advantage.

Lily Ray's writing, distributed through Amsive Digital and her Search Engine Land columns, has emphasized the E-E-A-T overlap with GEO. The pattern she has tracked is that the generative systems lean heavily on signals that the classical Google ranking system uses to assess expertise, authoritativeness, and trustworthiness: named expert authors, transparent methodology, dated content, citation density. The same pages that win the E-E-A-T-weighted classical results also win the citation slots in AI Overviews. Ray's framing has been that GEO is the same E-E-A-T discipline applied with greater rigor, not a separate practice.

Cyrus Shepard's writing on the question, distributed through Zyppy and his ongoing experimentation log, has emphasized the format and structure dimension. His tracking work has focused on the page-level features that correlate with citation: heading hierarchy, table presence, bulleted list density, FAQ block presence, image alt-text completeness. Shepard has tested the components individually and reported small but consistent effects from format cleanups alone, holding content quality constant. The practical recommendation has been format-first: do the structural work, then the content work, then track.

Kevin Indig's writing through Growth Memo has emphasized the strategic-portfolio framing. The question Indig has posed is which content investments to prioritize given that the generative surface is taking share from the click-based monetization of informational queries. The answer he has developed across multiple columns is that the publisher should bifurcate: continue to invest in informational content as a defensive citation play (with full GEO discipline applied), while shifting the growth investment toward content categories where the generative surface does not displace the click (commercial-intent content, product-specific content, deep technical comparisons, original research). The portfolio reallocation is the larger move; GEO is the tactic inside the defensive segment.

The four framings (international opportunity, E-E-A-T overlap, format-first, portfolio reallocation) are mutually compatible and emphasize different aspects of the same surface. Reading any one of them in isolation produces an incomplete operator picture; reading all four together produces the working synthesis that the rest of this essay has drawn on.

The Open Research Questions

The GEO discourse is at the stage where the practical work is reasonably stable and the research questions are still wide open. Five open questions are worth flagging for operators who are deciding how much to invest in GEO measurement.

The first question is whether the AI Overviews citation share materially affects downstream brand outcomes (awareness, search demand for the brand) at the population level, independent of the per-query click effect. The evidence is too thin to answer; the marketing-mix-modeling apparatus has not yet incorporated AI Overviews citation as a tracked input, and the brand-lift measurement infrastructure does not yet isolate generative-surface impressions from organic impressions.

The second question is whether the citation behavior of Perplexity and ChatGPT search is converging on or diverging from the AI Overviews citation behavior. The early tracking suggests partial convergence (the same domain authority patterns dominate) and some divergence (Perplexity appears to favor more sources per answer, ChatGPT search appears to favor authoritative encyclopedic sources more heavily). The behavior is changing fast and the comparative tracking is sparse.

The third question is whether structured data has an isolatable causal effect on citation rate, controlling for content quality. A controlled experiment that holds content quality constant and varies only structured-data markup would be the right design; we are not aware of one having been published.

The fourth question is whether the citation behavior of the generative systems is robust to adversarial inputs (content explicitly optimized to be cited) or whether the systems will incorporate adversarial filtering that suppresses heavily-optimized pages. The classical SEO history suggests the latter is likely over a multi-year window; the GEO discipline as currently practiced may have a limited shelf life as an unambiguously safe optimization.

The fifth question is what the equilibrium publisher behavior looks like once GEO is widely practiced. If every informational page on the open web starts with a one-paragraph answer, an entity-named opening, and a citation density that matches the Aggarwal recommendations, the differentiation effect of these patterns will erode and the systems will need different signals. The current advantage is partly a first-mover effect on a still-uncommon discipline.

Key Takeaways

1. Generative engine optimization is the working name for the practice of increasing citation rate in LLM-driven search interfaces; the measurable surface today is AI Overviews, Perplexity, ChatGPT search, and Copilot, with uneven tracking maturity.
2. The empirical record (Aggarwal et al. 2023, BrightEdge tracking 2024, practitioner case studies) supports a short list of content properties that correlate with citation: clear opening answer, entity clarity, citation density, structural cleanliness, alignment with literal question phrasing.
3. The citation pool overlaps heavily with the top of the classical SERP. GEO is a refinement on top of classical SEO, not a replacement; sites that abandon classical SEO in favor of "AI-first" practice tend to lose both surfaces.
4. Magnitudes are modest. Citation lifts of 30 to 70 percent in relative terms are achievable on tracked queries with full discipline; downstream click effects are smaller than the citation lifts suggest because the generated answer often satisfies the user's question.
5. The systems are non-stationary. Citation behavior is observably tuned by the operators of the generative engines, and the half-life of a GEO test result is approximately one major model update.
6. Structured data is worth maintaining for the standard SEO reasons; the incremental causal contribution to GEO citation has not been cleanly isolated in the published evidence.
7. The practical disciplines an operator can run today are the answer-in-opening pattern, entity naming, citation insertion, FAQ blocks, dataset blocks, and periodic refresh; all of these are recommendable on conventional editorial grounds independent of GEO.
8. GEO is best treated as a defensive practice for operators with strong classical SEO already, rather than as a primary growth lever; the click economics of the generative surface are systematically less favorable than the classical SERP for the publisher.