Is GEO just SEO? Rankings win retrieval, content wins the citation

The claim: "GEO is just SEO. Rank well and AI visibility follows." Circulating this week in a large practitioner thread and standing behind a lot of agency positioning, argued in both directions with equal confidence. Claim Check: Overstated. The evidence supports search rankings as the eligibility layer of AI answers, and it breaks exactly where citation is decided.

This claim crossed my desk three times this week, and both camps sounded certain. We checked it against the library. Google itself states that appearing in AI Overviews requires nothing beyond being indexed and eligible for a standard snippet Google · 2025A. The same evidence base shows that a strong ranking converts into AI answers far less often than the claim implies. The sections below separate the half that holds from the half that does not.

The true half: AI answers still run on search infrastructure

Google has said the quiet part in its own documentation. There are no extra technical requirements for appearing in AI Overviews or AI Mode beyond being indexed and snippet-eligible Google · 2025A. A follow-up guide goes further. Sites do not need llms.txt files, content chunking, AI-specific rewrites, or extra structured data to appear in Google's generative features Google · 2026A. On Google's surfaces, the entry mechanics are inherited from search.

The pattern extends beyond Google. About 83% of ChatGPT's shopping-carousel products matched Google Shopping's top organic positions, with 60% drawn from the top ten Landwehr · Peec AI · 2026B. The study is first-party vendor research with a disclosed method; treat the exact percentages as directional. The direction is hard to argue with: ChatGPT Shopping largely re-ranks what Google Shopping already ranked.

Aggregate scoring points the same way. A meta-analysis of 54 AI-citation studies scored 23 factors. Search rank rated 9.4 out of 10 and fan-out rank 9.3, behind only URL accessibility at 9.5 Shepard · Zyppy · 2026B. Pages ranking for both a main query and at least one fan-out sub-query collected 51% of AI Overview citations. Pages ranking only for the main query collected under 20% Search Engine Land · 2025B. Both are vendor studies; the convergence with Google's own documentation is what earns them a place here.

Rank even outweighs the tactic most often sold as its replacement. Across 730 citations over 75 commercial queries, schema presence did not independently predict AI citation once Google rank was controlled (corrected odds ratio 0.678) Fischman · Growth Marshal · 2026B. The retrieval layer of AI answers was not rebuilt. It was inherited.

The false half: eligibility does not convert into citation

Generative engines weaken the link between ranking and visibility. A benchmark built to measure whether source articles shape AI answers tested the ranking link directly. It found the connection between search position and answer influence materially looser than in classic search arXiv · 2025A. Generative and traditional search return different things for the same query. Generative engines pull from a broader pool of sources and mix in varying amounts of internal model knowledge Kirsten et al. · Ruhr University Bochum / Max Planck Institute for Software Systems · 2025A.

The divergence compounds across engines. Across 127,198 citations and roughly 16,400 commercial-intent answers, 69.6% of cited domains appeared on only one of five engines and just 2.7% on all five Khallad · SurfacedBy · 2026B. A Google ranking feeds Google's surfaces directly. ChatGPT, Perplexity, and Claude run their own retrieval, and they mostly read a different web.

The clearest single number comes from brand discovery. When users named a product directly, ChatGPT recognised it 99.4% of the time. When they asked discovery questions such as "best AI tools launched this year," success collapsed to 3.32% Sharma · arXiv · 2026A. These were products with functioning websites and normal search footprints. Eligibility existed; citation did not follow. The library's discovery-gap article sets out that retrieval half in full.

The citation test is also winnable from below, which a pure-SEO account cannot explain. In the foundational GEO experiment, adding citations, quotations, and statistics raised visibility in AI answers by up to about 41% Aggarwal et al. · Princeton University / Georgia Tech / Allen Institute for AI / IIT Delhi · 2024A. Pages ranked outside the top positions saw the largest gains. If rank alone decided the answer, the biggest wins could not sit at the bottom of the ranking.

Two tests, not one: retrieval decides who competes, the answer decides who is quoted

A critical survey of 45 GEO studies lands on the frame that reconciles both halves. Generative visibility is not one ranking task but a stochastic, multi-stage pipeline. Getting retrieved and getting cited are separate steps with separate levers Martinez · arXiv (cs.IR) · 2026A. The first step behaves like search, because for the most part it is search. The second step does not.

The second step has its own measured drivers. Across 252,000 paired trials spanning 18 content factors and six models, topical relevance and list position inside the retrieved context drove which source was cited first. Formatting-only changes did almost nothing Vishwakarma et al. · Sprinklr · 2026A. What decides the quote is what is inside the candidate content, not the rank that got it there.

One honest boundary on the verdict: the claim is closest to true on Google's own surfaces, where entry mechanics are documented as inherited. It is weakest on ChatGPT, Perplexity, and Claude, where citation overlap with any other engine drops to single digits Khallad · SurfacedBy · 2026B. "GEO is just SEO" is a reasonable approximation of one engine's front door. It is a poor description of the market.

What I'd watch

The number this debate needs is a conversion rate: how often a search position turns into presence in the AI answer. It would have to be reported per engine and per query type. A figure circulating among practitioners puts the conversion from position one at roughly one in three. The methodology behind it has not been published, so the figure stays out of the body of this article. If a disclosed-method study lands near that figure, both camps turn out to be measuring different tests, and both are right about their own test. That is the study I am watching for.