GEO-First Framework Debunked as Google Calls AI Search SEO

A New Line of Attack on GEO

Critics of generative engine optimization (GEO) have mostly argued against it in blog posts and conference talks. Now one of them has gone through the footnotes. A long critique published by Hobo Web picks apart a January 2026 review paper titled "A GEO-First Framework" by Uri Samet, which appeared in the open-access journal WJARR. The critique argues that the paper's central ideas come from marketing blogs and agency pages rather than research.21 Its broader claim is the one the search industry has been arguing about for two years: that GEO is not a new discipline but search engine optimization under another name.21

The timing helps the critics. On May 15, 2026, Google Search Central published documentation on optimizing for generative AI features. It says that, from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, "and thus still SEO."1 Taken together, the source audit and Google's guidance make the strongest case yet against selling GEO as a separate service. The coverage also shows where that case is weakest.

What GEO Is Supposed to Be

The term comes from a 2023 preprint by researchers at Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi, which was presented at KDD 2024.28 The paper described GEO as a creator-focused framework for improving how visible content is in AI-generated answers. It reported that tactics such as citing sources, adding quotations and including statistics could raise visibility by up to 40%.2527 It also introduced GEO-bench, a benchmark of 10,000 queries.19

Marketers seized on one finding in particular. Lower-ranked pages appeared to benefit most: a page ranked fifth saw a 115.1% visibility gain from the "Cite Sources" method, while the top-ranked page lost ground on average.28 That result became the basis of the pitch that GEO "democratizes" search for small publishers who lack links and authority.

Skeptics raised doubts from the start. When Search Engine Land covered the paper in 2024, it noted that these were not real-world results. Only a subset was tested on Perplexity, and Perplexity is not Google.28 The paper's own limitations section concedes that its methods resemble SEO and that their effect could shrink as models take in more sources.27

The Case Against the 'GEO-First' Paper

The Hobo Web critique attacks Samet's paper on its sources. By its account:

• The user-journey table and schema advice trace back to a blog post about answer engine optimization (AEO), not GEO.
• The technical requirements table comes from an e-commerce marketing listicle.
• Material on information gain and server-side rendering comes from a UK agency's service page.
• A closing reference points to a post on the author's own reputation-management agency site.21

The critique also says the paper's working mechanisms, including concepts it calls "Inference Optimization" and the "Fortress of Facts," were taken from Hobo Web's own anti-GEO frameworks without crediting their author, Shaun Anderson.21

One technical claim draws particular scorn. According to the critique, the paper treats Interaction to Next Paint (INP) as a measure of responsiveness for "agentic scraping." Hobo Web points out that INP measures how quickly a page visibly responds to a human click or keypress, so it has nothing to do with headless crawlers that parse raw HTML.21 That error, if accurately described, says a lot about how carefully the paper was put together.

The most damaging evidence comes from the paper itself. Hobo Web quotes its limitations section, which says the GEO-First framework is a conceptual model built from existing literature, has not been tested in controlled experiments, and rests on correlational evidence, so any causal claims remain provisional.21 The critique adds that the paper was accepted 18 days after submission, over the holidays.21

Hobo Web is not a neutral party. It has argued against GEO since at least August 2025, and it says the paper borrowed its own work, so it has a direct stake in the outcome.21 Still, its specific points about where the citations lead can be checked against the bibliography, and they hold together. The fairest reading is that this paper is weak support for GEO as a field, whatever one thinks of the field.

Questions About the Original Study

The critique also goes after the 2024 study that started it all. It argues that the KDD paper took Google's top five results as its pool of sources, so the experiment depended on SEO rankings before any GEO tactic was applied.21 It also claims that some of the biggest gains came from GPT-3.5 inventing citations and statistics.21

Hobo Web cites later academic work as well. By its summary, the NeurIPS 2025 C-SEO Bench study found that the original rewriting tactics often stopped working when tested on other models or when several competitors used them at once. A 2026 benchmark called SAGEO Arena reportedly found that rewriting only the body text could hurt retrieval, while titles and headings drove visibility.21 These descriptions are Hobo Web's, and readers should check the original papers before relying on them. Even so, they point the same way: tactics measured on a fixed, preselected set of documents do not reliably carry over to live search systems.

Google's Guidance Changes the Default

The academic argument matters less than what Google has now put in writing. Its May 2026 guide says AI Overviews and AI Mode are built on its core ranking and quality systems. They use retrieval-augmented generation (pulling passages from indexed pages to ground an answer) and query fan-out (breaking one question into several searches) to draw content from the regular Search index.1 A "Mythbusting generative AI search" section says several tactics are unnecessary for Google: llms.txt files, content chunking, AI-specific rewrites, special schema, and inauthentic third-party mentions.12

Coverage of the guide largely agrees on what it means. The Decoder called it a direct shot at consultants selling specialized GEO strategies.6 Search Engine Journal said the mythbusting section carries the most weight because it lists tactics an entire service category had been promoting.1 Several reports noted that Google staff, including Gary Illyes and Cherry Prommawin, had said the same thing at Search Central Live before it appeared in the documentation.110

Some practitioners have changed their advice. One German SEO publicly corrected an earlier article that had called schema the "entry ticket" to AI Overviews. He pointed to the guide and to an Ahrefs study of 1,885 pages that found no gain in AI Overview citations after JSON-LD structured data was added.2 What Google recommends instead is ordinary SEO work: original "non-commodity" content, clean indexing and crawl access, less duplicate content, and Merchant Center or Business Profile data where they apply.10

One version of events needs correcting. Some outlets dated the guidance to a June 5 update.37 Most reports put the original publication on May 15.19

Where the Coverage Splits: Engines Other Than Google

Much of the coverage agrees on one limit: Google spoke only for Google. Search Engine Journal noted that the guide does not settle the question for ChatGPT or Perplexity, which may weight signals differently.1 Other analysts described "still SEO" as half the picture, because those systems have their own retrieval sources and habits.45

The data on third-party engines is where the reporting disagrees most. One roundup of five 2026 studies reported a weak correlation (r=0.34) between organic rank and AI citation, and found that 46% to 69% of citations came from outside the top 10.12 The same roundup found that a top-three ranking made a page 7.82 times more likely to be cited than a page ranked 11 to 30, and that Google AI Overviews followed rankings most closely while Perplexity followed them least.12 Another analysis found that only 6.82% of ChatGPT results overlapped with Google's top 10.15 Yet a GEO vendor's own guide cites figures saying 93.67% of AI Overview citations link to at least one top-10 result, and that 87% of ChatGPT citations match Bing's top 10.17 Another statistics roundup claims 83% of AI Overview citations come from outside the organic top 10.11

Those numbers cannot all be true at once. Most come from vendors that sell AI-visibility tracking, and the studies define "citation" and "overlap" differently. Some of the observed shift is real: brand mentions showed a stronger correlation with AI visibility than backlinks in an Ahrefs study of 75,000 brands.11 But brand reputation and earned mentions have long been part of SEO and digital PR. That is Hobo Web's point when it describes "share of voice" and "sentiment" dashboards as old reputation metrics in new packaging.21

Our Reading

The evidence points to one conclusion. For Google's AI Overviews and AI Mode, GEO as a separate discipline is finished. Google has said so in writing and listed the tactics it ignores.19 The "GEO-First Framework" adds little to the debate. Going by its critics' citation audit and its own admitted lack of testing, it is a literature review assembled from marketing material, not evidence of a new field.21

The remaining case for GEO depends on non-Google engines and on measurement. Even that case mostly calls for familiar work: crawlable pages, original data, clear writing and real brand authority. Some dissenters go further. One argues that LLMs work as systems separate from search and need a new discipline, which is itself being sold under another acronym.23 Others, including a vendor that tracks citations across four platforms, say Google-only optimization leaves visibility on the table.7

Those arguments deserve testing, not dismissal. But the bar has moved. Anyone selling GEO as distinct from SEO now has to show, with controlled experiments, a tactic that works outside Google's index and that good technical SEO and content work would not already deliver. Until then, the label describes a way of measuring AI visibility more than a new way of earning it.