GEO after SEO. What can actually be optimised in AI search?

The GEO market often labels every desirable outcome as “optimisation”: source selection, citation, brand inclusion, favourable framing and even recommendation. This is a category error.

A brand can change its website, access conditions, information architecture, content, claims and some of the data submitted to external platforms. It cannot instruct a retrieval system to select a particular document. It cannot force a citation or determine how a model will synthesise its sources. Nor does it control the final recommendation.

A mature approach to Generative Engine Optimisation (GEO) therefore requires three distinct modes of action:

• directly optimising controlled assets and conditions;

• indirectly influencing how information is selected and used;

• monitoring outcomes that remain outside the brand’s control.

Put more simply:

GEO is not the optimisation of a model’s answer. It is the optimisation of controlled conditions, an attempt to influence intermediate processes and the measurement of a representation the brand does not control.

“After SEO” does not mean “without SEO”

The word “after” does not suggest that GEO replaces SEO. It refers to the later stages of the information flow – the point at which an available document may be retrieved, selected, used and transformed into an answer.

For Google Search’s generative features, the fundamental requirements of SEO remain an entry condition. Google states that AI Overviews and AI Mode use its core Search quality and ranking systems, Retrieval-Augmented Generation (RAG) and query fan-out. A page must be indexed and eligible to appear with a snippet, but meeting these conditions does not guarantee crawling, indexing or presentation.

Not every system reaches information through the same route.

OpenAI distinguishes the automated OAI-SearchBot from ChatGPT-User, which may visit a page as a result of a user action. Anthropic assigns separate roles to Claude-SearchBot and Claude-User. Perplexity similarly distinguishes PerplexityBot from Perplexity-User, with user-initiated retrieval potentially operating under different rules from automated indexing.

The scope of this article

Throughout this article, generative search system is used as an umbrella term. It does not imply that every product has the same architecture.

The sequence:

access → retrieval → source selection → citation → absorption → synthesis → representation → recommendation

is an analytical model. In a particular system, some stages may be iterative, parallel, hidden or absent.

What should GEO mean?

Three competing definitions can be identified across academic research and industry discourse.

The original study, “GEO: Generative Engine Optimization”, defined GEO as a black-box optimisation framework designed to increase content visibility within generative-engine responses. The researchers modified documents and measured changes in their exposure. The reported result of “up to 40%” applied to a specific benchmark, its own visibility metrics and a controlled research environment – it is not a benchmark for current production systems.

From Google’s perspective, Answer Engine Optimisation (AEO) and GEO are industry terms relating to visibility in AI search experiences, but optimisation for AI Overviews and AI Mode remains part of Google Search optimisation. Google does not require llms.txt, specialist markup, artificial chunking or a separate writing style for AI systems.

Both definitions are useful, but neither provides a sufficiently neutral definition of the discipline.

An operational definition

Generative Engine Optimisation is the evidence-led practice of modifying controllable information assets and access conditions to create better conditions for generative search systems to find, select, use and accurately represent information – while measuring outcomes that remain outside the brand’s direct control.

This definition does not include:

• directly controlling the answer;

• training or fine-tuning a provider’s model;

• every SEO or PR activity;

• monitoring alone;

• guaranteeing a mention, citation or recommendation.

GEO, AEO, LLMO and AI visibility

These terms describe overlapping but distinct objects.

AI visibility is a measurement domain, not a synonym for GEO.

A brand may be mentioned frequently but described inaccurately. It may receive citations without being recommended. It may appear consistently as a source of information but not as a provider of the relevant solution.

The GEO control surface

The GEO control surface, proposed by Brand Semantics, organises activities according to the level of agency available to the brand.

The model does not suggest that every element belongs exclusively to one category. Citation, for example, is a process that a brand may attempt to influence and an outcome that it must observe.

The classification describes the appropriate mode of management, not merely the element’s position within a pipeline.

What brands can optimise directly

Direct optimisation is possible where a brand controls the object of the intervention and can verify that the change has been implemented.

Technical access is controllable; inclusion is not

In Google Search, a brand can control whether its pages are technically available for crawling, indexing and snippet presentation. It cannot guarantee that Google will index a page or display it within a generative feature.

The same distinction applies to other platforms:

• OAI-SearchBot supports the inclusion of pages in ChatGPT Search features, while GPTBot relates to content that may be used in model development. The controls are independent.

• Claude-SearchBot supports indexing intended to improve the quality, relevance and accuracy of Claude’s search results, while Claude-User handles retrieval initiated by users.

• PerplexityBot supports Search surfaces, while Perplexity-User may visit a page in response to a user request. Perplexity states that the latter generally ignores robots.txt because the retrieval is user-initiated.

There is therefore no single “allow AI” or “block AI” decision. Automated indexing, retrieval on demand, model development and Web Application Firewall (WAF) rules must be considered separately.

Content, claims and structured data

A brand can improve:

• the precision of its definitions;

• the transparency of its methodology;

• the quality of its data;

• the clarity of its sources;

• the structure of its argument;

• the recency of its information;

• the distinction between facts, interpretations and limitations;

• the consistency of names, products and categories.

This does not imply the existence of a universal writing style that guarantees citation.

Google recommends useful, distinctive and non-commoditised content, but rejects the need for a specialist writing style for generative Search, an ideal document length or the artificial division of content into short fragments.

Structured data is also a controllable element. It can help Google Search understand visible content and determine eligibility for particular rich results, but technically correct implementation does not guarantee that those results will be shown. Structured data must reflect information that is available to the user.

The detailed design of entity, claim and source maps is covered in Brand Semantics Infrastructure: how to make AI Search understand your brand correctly. Here, these elements matter as controllable inputs – not as a guarantee of the final representation.

What can only be influenced

A brand can create better conditions for retrieval, source selection and accurate synthesis, but it does not control those decisions.

Retrieval and source selection

Potential interventions include:

• technical accessibility;

• semantic alignment between the document and the query;

• clear terminology;

• the presence of relevant claims;

• current data;

• availability in the user’s language;

• external sources that corroborate important information.

However, the full candidate set, all auxiliary queries and the weights applied by the platform remain unknown. Retrieval and source selection are therefore areas of influence and partially unobservable processes.

The absence of a visible citation does not establish that a source played no role in retrieval or generation. Without access to the platform’s internal logs, part of the process remains unobservable. A specific invisible source should not be credited with shaping an answer without additional evidence.

Citation and absorption are not the same

A brand can improve a document, but it cannot implement a citation rate on the page.

Citation rate is an outcome, not an object of optimisation.

The study “From Citation Selection to Citation Absorption” distinguishes between:

• citation selection – the selection and presentation of a source;

• citation absorption – the influence of the cited page on the language, facts, evidence or structure of the answer.

Within the dataset analysed, citation breadth and depth of influence were not equivalent. The study also found that pages with greater observed influence were more likely to be well structured, semantically aligned and rich in extractable evidence. These are descriptive relationships, not proof that a single structural change will cause higher absorption. The publication is a preprint.

A full methodology for distinguishing citation from absorption requires a separate analysis. For the purposes of this model, the relevant distinction is:

• structure and evidence are controllable inputs;

• citation selection and absorption are areas of influence;

• citation rate and claim absorption are observed outcomes.

External sources

Digital PR, publisher relationships and the correction of external sources may support GEO, but not every mention is a GEO intervention.

An external publication becomes part of a GEO programme when it:

1. supports a defined claim or entity relationship;

2. addresses a specific source gap;

3. is connected to an explicit influence hypothesis;

4. is subsequently monitored for selection, citation or representation.

The brand controls its own data, research and outreach. It does not control the publisher’s editorial decision or the system’s subsequent selection of the publication.

A full analysis of owned, earned and third-party sources belongs elsewhere in the content cluster. This article only classifies them according to the level of control available.

Framing and recommendation

A brand can:

• define its category clearly;

• explain relevant use cases and limitations;

• publish comparisons based on explicit criteria;

• correct inaccurate information;

• build consistency between its offer, audience and the problem it solves.

It cannot determine whether a system presents it as the first recommendation, one option among several, a niche solution or a brand that is irrelevant to the scenario.

Recommendation rate, answer salience and framing are observed outcomes. Claiming that a brand can “optimise recommendations” assigns it a level of control that it does not possess.

Example: one asset, four levels of agency

Suppose a business-to-business software provider publishes its own report comparing methods for measuring process efficiency.

The example demonstrates why an increase in citation rate should not be described as “implementing citations”. The brand implemented a report, improved its source environment or increased availability. Citation is a downstream outcome.

What should primarily be monitored

AI visibility is not a single metric.

A single prompt does not provide a stable measurement of a platform. Answers may vary between runs, prompt variants and dates. Recent preprints recommend treating visibility as a distribution of outcomes and reporting uncertainty rather than presenting individual results with false precision.

This article does not resolve the full problem of experimental GEO measurement. That requires a separate methodology covering scenario libraries, repeated trials, platform-change controls, response classification and statistical uncertainty. It will be developed in a separate article: Measuring AI visibility reliably: prompts, variance and experimental design.

Where GEO ends

GEO intersects with several other disciplines, but it should not replace them.

Monitoring is a necessary part of a GEO programme, but monitoring alone is not optimisation.

How to assess evidence for GEO techniques

Not every recommendation has the same evidential status.

Every recommendation should identify:

• the object being changed;

• the level of control;

• the proposed mechanism;

• the type of evidence;

• the outcome to be measured;

• alternative explanations.

What this does not mean

GEO does not mean SEO has become obsolete

For systems that depend on a Search index, technical accessibility and resource quality remain entry conditions.

GEO does not mean every answer can be optimised

A brand optimises inputs. The final answer is an output of the system.

GEO does not guarantee citation

A correctly implemented resource may still not be retrieved, selected or displayed.

Citation does not mean absorption

A visible link and the source’s actual influence on the answer are separate outcomes.

More mentions do not necessarily improve representation

Mentions may be inconsistent, outdated or irrelevant to the scenario.

An external publication does not change model memory on demand

Live retrieval, a search index and future training data are different mechanisms.

Schema.org does not guarantee generative visibility

Structured data may support machine readability and eligibility for Search features, but it does not force source selection.

One prompt does not represent a platform

A single answer is an observation, not a stable brand position.

Visibility does not mean business impact

Mention rate, citation rate and recommendation rate are intermediate outcomes. Traffic, conversion and business value require separate measurement.

The C–I–O–X framework

Before selecting a GEO tactic, work through five steps.

1. Identify the object

Do not say “improve AI visibility”. Specify the intervention or measurement:

• allow OAI-SearchBot;

• update a claim;

• publish a methodology;

• measure citation rate;

• classify framing.

2. Assign a class

• C – Control: the brand can implement the change;

• I – Influence: the brand can improve the conditions;

• O – Observation: the element is an outcome;

• X – Exogenous condition: the variable originates in the surrounding system.

3. State the mechanism

Not:

Adding a table will increase citations.

Instead:

A table may make data easier to interpret and comparisons easier to extract, which justifies testing its effect on selection, citation or absorption.

4. Define the evidence level

Does the recommendation come from:

• provider documentation;

• research;

• correlation;

• interpretation;

• hypothesis?

5. Match the metric and report the limitations

The intervention should be evaluated against an outcome that corresponds to its proposed mechanism, while accounting for variability and alternative explanations.

The most practical principle of GEO is therefore:

Optimise the assets and conditions you control. Influence the source environment where possible. Measure the representation you do not control.

If a team cannot distinguish between these three layers, its next step should not be another GEO tactic. It should be a diagnosis of its GEO control surface.

Brand Semantics applies this distinction through AI Strategic Consulting – identifying which problems require technical or content intervention, which require work on the source environment and which require reliable monitoring. To discuss a specific case, contact Brand Semantics.

Sources and methodological notes

Official provider documentation

1. Google Search Central, “Optimizing your website for generative AI features on Google Search”, updated 10 July 2026. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

2. Google Search Central, “AI Features and Your Website”. https://developers.google.com/search/docs/appearance/ai-features

3. Google Search Central, “General structured data guidelines”, updated 10 July 2026. https://developers.google.com/search/docs/appearance/structured-data/sd-policies

4. OpenAI, “Overview of OpenAI Crawlers”. https://platform.openai.com/docs/bots

5. Anthropic Privacy Center, documentation covering ClaudeBot, Claude-SearchBot and Claude-User. https://privacy.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler

6. Perplexity, “Perplexity Crawlers”. https://docs.perplexity.ai/docs/resources/perplexity-crawlers

Research

7. Aggarwal, P. et al., “GEO: Generative Engine Optimization”. https://arxiv.org/abs/2311.09735

8. Kai, Z., He, X. and Yao, J., “From Citation Selection to Citation Absorption”, arXiv:2604.25707, version 2, 29 April 2026. https://arxiv.org/abs/2604.25707

9. Schulte, J., Bleeker, M. and Kaufmann, P., “Don’t Measure Once: Measuring Visibility in AI Search (GEO)”, arXiv:2604.07585. https://arxiv.org/abs/2604.07585

10. Sielinski, R., “Quantifying Uncertainty in AI Visibility”, arXiv:2603.08924, version 2, 9 June 2026. https://arxiv.org/abs/2603.08924

Brand Semantics contributions

The operational definition of GEO, The GEO control surface and the C–I–O–X classification are methodological proposals developed by Brand Semantics. They are not official provider terminology or a generally accepted academic standard.