GEO, Agent Intent Optimisation (AIO), and the Difference Between Being Cited and Being Selected
Revision note, September 2026: this dated piece revises the article first published on March 7, 2026 and the 2025 wording of the definition it carried, on the evidence published since; the 2025 paper that introduced the discipline (SSRN 5511758) stays as written, and dated pieces like this one do the revising.
The marketing profession has a long tradition of adopting new acronyms faster than it develops new thinking. GEO, generative engine optimisation, is the latest addition to that tradition. It describes something real and commercially important. It also creates a conceptual ceiling that, if left unexamined, will leave organisations answering the wrong question.
GEO and Agent Intent Optimisation (AIO) are not rivals, and this is not a case against either. It is a case for knowing exactly what each discipline can and cannot do.
Update. Since this article was published in March 2026, Google has released its guide 'Optimizing your website for generative AI features on Google Search' (published 15 May 2026, last updated 10 July 2026), which states that optimizing for generative AI search is 'optimizing for the search experience, and thus still SEO', and its spam policies (last updated 28 August 2026) now name 'attempting to manipulate generative AI responses in Google Search' as spam. A July 2026 critical survey of 45 studies (arXiv 2607.14035) found that no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behaviour, and the C-SEO Bench benchmark of conversational SEO methods (NeurIPS 2025 Datasets and Benchmarks, arXiv 2506.11097) found most of them largely ineffective and frequently negative for document ranking. A separate controlled study (arXiv 2404.07981), which is not a study of generative engine optimisation, showed that a crafted text sequence added to a product's information page can raise its likelihood of being listed as an LLM's top recommendation; it measures a recommendation list, not a purchase. Those findings concern, in turn, search on one platform, organic discoverability, document ranking and a recommendation list in a controlled setting; each is read on its own construct, and none of them measures what a buying agent does in a purchase. The distinction drawn below, between being cited and being selected, stands. The definition of Agent Intent Optimisation below has been rewritten in the form the practice now uses, which says what the discipline is for and names its test; whether and how today's buying agents select that way is the question a pre-registered experiment is measuring now; nothing here reports its result.
What GEO Is, Precisely
GEO was introduced in a 2023 academic paper from Princeton University and Georgia Tech as a framework for optimising content visibility in generative engine responses. The commercial application is clear: as AI-generated answers become a primary way people retrieve information, firms want their content to appear inside those answers rather than simply ranking beneath them.
The strategic logic is not empty. The most critical reading of that literature, the July 2026 survey of 45 studies, itself concedes that already-retrieved content can causally alter its citation or use inside a generated answer. On the published record, structure earns its keep after retrieval; what no reviewed technique has shown is a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behaviour.
But examine the architecture of GEO and one assumption becomes clear: at the end of the optimisation chain, there is a human reading the AI's response. A human who sees your brand cited, follows a link or makes a purchasing decision based on what they read.
GEO is a content strategy discipline, an extension of SEO for an AI-mediated information environment; Google's own guide says that 'optimizing for generative AI search is optimizing for the search experience, and thus still SEO'. In the Agent Decision Preference Stack it operates at the second layer, inferred knowledge, shaping what an agent has learned about a brand before it decides.
It does not address what happens when no human is reading.
The distinction this article drew in March is between being cited and being selected. They are different conversations, and only the second is the subject of agentic commerce.
The Structural Shift GEO Does Not Capture
Agentic commerce is the commercial model in which purchasing is delegated to autonomous AI agents acting on behalf of human principals, without human review at the point of transaction. The consumer delegates. What the agent then does with that delegation is the entire question.
The March version of this article carried adoption figures at this point. This revision does not. No figure for agent-driven sales sits in the evidence this piece allows itself, and a discipline whose test is verification should not lean on a number the reader cannot check. What can be said, and dated, is that the question is now testable: a pre-registered experiment on four language-model buying agents, covering admission, verification and schedules, is running now (OSF, DOI 10.17605/OSF.IO/GPBUK), and nothing here reports its result.
The March version also told you what happens when an autonomous agent evaluates a product: no content read, no citation followed, a transaction executed or abandoned on computational criteria. Written as a discipline should write it, that is a hypothesis, not a report. A buying agent, as the discipline models it, queries product data, weighs interface reliability, pricing consistency and the operational record, and acts on what survives its checks. Whether today's buying agents do select that way, and on which of those signals, is under pre-registered test.
None of those signals is addressed by GEO. That is not an accusation. It is a boundary.
Agent Intent Optimisation: A Different Discipline, Built for a Different Test
Agent Intent Optimisation (AIO) is the discipline that succeeds SEO when the buyer is a machine: engineering a product's data, claims, and commercial signals so that an AI agent can retrieve them, read them, and verify them independently. Where SEO competes for human attention and generative engine optimisation (GEO) competes for citation inside a generated answer, Agent Intent Optimisation (AIO) competes for the selection itself, and an agent selects on what it can check rather than on what it is told. Whether and how today's buying agents select that way is the question a pre-registered experiment is measuring now; nothing here reports its result.
The discipline addresses three layers of algorithmic commerce readiness on the seller's side, distinct from the three layers of the agent's Preference Stack. The informational layer, accurate and machine-readable product data, overlaps with GEO. The evaluative layer, specifications, pricing and operational data prepared to the precision an agent's decision logic is built to apply, goes beyond it. The transactional layer, making it possible for an agent to execute a purchase through your systems without friction, is entirely outside GEO's scope.
The Preference Stack itself, introduced in the 2025 paper, is the three-layer model of how an AI buying agent selects: foundational directives, the hard constraints an offer must satisfy or be eliminated; inferred knowledge, the learned heuristics such as brand reputation, expert citations and review patterns that order the surviving options, where generative engine optimisation primarily operates; and real-time optimisation, current price, availability, delivery, transaction ease and service reliability, where the final selection is made and where Agent Intent Optimisation operates. The Preference Stack describes how an agent decides; Agent Intent Optimisation describes what a seller makes checkable so that the decision can rest on it. Where a buying agent makes its selection, and on what, is the model's proposition; it is under pre-registered test, and nothing here reports a result.
The strategic implication is not that GEO is wrong about what it measures. It is that GEO is incomplete as a framework for the commercial challenge ahead: the published evidence for generative engine optimisation concerns visibility and ranking in generated answers, and whether a buying agent selects on what it can check is the question now under pre-registered test. Whether a firm with strong GEO performance can still fail when an autonomous agent attempts to purchase from it is exactly the question that test is built to answer.
Why Terminology Has Commercial Consequences
The practitioner vocabulary adopted now will shape the strategies, budgets and organisational structures built over the next three to five years. If GEO becomes the dominant frame for AI commerce readiness, firms will plan for content optimisation and measure success by citation rates: valuable work that addresses one layer of a three-layer problem.
A note on the letters, because the trade has made them ambiguous. In the search trade, the three letters in this article's title are read as Google's AI Overviews or as artificial intelligence optimisation, which the English Wikipedia article on generative engine optimisation lists among other terms for the same concept. That collision is why the name is spelled out in full here: Agent Intent Optimisation is not a synonym for GEO, and the argument between the disciplines is about the test, not the letters.
The questions the discipline exists to make answerable read differently from citation metrics. Is my product data machine-readable to the precision an autonomous agent requires? Can an agent complete a purchase through my systems without manual steps? What does my operational reliability look like to an algorithm that has never heard of my brand?
And one discipline rule, stated plainly: until the pre-registered question is answered, the practice makes claims checkable and reports what the platform and the studies say, and recommends nothing about spend.
The difference between the two disciplines is not technical sophistication. It is commercial scope. GEO solves for discoverability. Agent Intent Optimisation (AIO) is built for selectability: it makes a product's claims checkable so that an agent that selects on what it can check can select it. In commerce, those are not equivalent aims.
GEO tells the algorithm who you are. Agent Intent Optimisation (AIO) tells the algorithm what it can check about you. The second conversation is the one built for a buying agent to act on, and whether today's agents do is under pre-registered test.
References
- Princeton/Georgia Tech GEO paper (arxiv.org/abs/2311.09735)
- SSRN Abstract 5511758
- California Management Review Insights
- Google, 'Optimizing your website for generative AI features on Google Search', published 15 May 2026, last updated 10 July 2026 (developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- Google, 'Spam policies for Google web search', last updated 28 August 2026 (developers.google.com/search/docs/essentials/spam-policies)
- Martinez, critical survey of 45 studies, July 2026 (arxiv.org/abs/2607.14035)
- C-SEO Bench, NeurIPS 2025 Datasets and Benchmarks (arxiv.org/abs/2506.11097)
- Kumar and Lakkaraju, 'Manipulating Large Language Models to Increase Product Visibility' (arxiv.org/abs/2404.07981)
Agent Intent Optimisation (AIO) is a registered trademark in the UK and the EU.
© 2026 Paul F. Accornero. All rights reserved. · paulaccornero.com/legal-and-ip