WorksComputer Science
Humanity's Last Choice: A Common Evaluation Standard for Generative Engine Optimization
Abstract
Humanity's Last Choice describes the idea that the last choice people need to make is which AI system they entrust with the decisions that follow. These systems rely on external sources that content providers can edit through generative engine optimization (GEO) to increase source visibility and influence recommendations. Choosing among these systems therefore requires comparable evidence of how well they carry out user requests when exposed to GEO. Benchmarks have made progress in standardizing tasks, models, and metrics, but they often define their own spaces of GEO strategies and use different editing constraints. These differences can conflate gains from better optimization algorithms with those from greater freedom to edit sources, while making evaluation coverage difficult to compare across benchmarks. We therefore call for a common evaluation standard and propose a meta-taxonomy that describes where sources are edited, how they are edited, and what signals those edits target. Drawing on representative literature, we map existing strategies into a searchable space and outline a protocol for controlled comparisons with explicit coverage reporting. This provides a foundation for common leaderboards that can inform which systems people choose to rely on.
§ github.com/humanitylastchoice humanitylastchoice.org # [email protected]
Introduction
Conversational agents such as ChatGPT and personal agents such as OpenAI’s dots, Meta’s Muse, and Google’s Gemini Spark are increasingly capable of gathering information, comparing options, and carrying out tasks on a user’s behalf ([21, 18, 9]). As their capabilities grow, people can delegate more of the decisions involved in those tasks to them. Humanity’s Last Choice describes the idea that the last choice people need to make is which AI system to entrust with the decisions that follow.1 Such a system evaluates information and recommends candidates using sources that content providers can edit.
Generative engine optimization (GEO) studies how such edits affect source visibility in generated answers ([1, 13, 23]). Choosing which system to rely on therefore requires comparable evidence of how well it carries out user requests when those sources are optimized to influence its judgments. GEO strategies have grown beyond a fixed list of writing tips. They include authoritative phrasing, added citations, appended token sequences, and document-level feature changes. AutoGEO learns engine preference rules, while AgenticGEO evolves a reusable strategy archive ([32, 37]).
Existing benchmarks, including C-SEO Bench, E-GEO, and the ranking-manipulation GEO-Bench, have made progress in standardizing tasks, models, and metrics [25, 2, 20]. However, shared tasks, models, and metrics do not by themselves define the strategies that methods are allowed to apply. A method limited to appending a short sequence and another allowed to rewrite an entire page can be evaluated by the same ranker but operate under different editing permissions. Comparing their scores without accounting for these differences can conflate gains from better optimization algorithms with those from greater freedom to edit sources. These differences also affect system selection because stronger task performance may reflect a less demanding set of edits. A common evaluation standard therefore needs to specify both the strategy space being evaluated and the conditions under which methods explore it.
Our position is that a common evaluation standard for GEO should be grounded in a shared specification of where sources are edited, how they are edited, and what signals the edits target. We propose a meta-taxonomy with three dimensions. Optimization Layer identifies the source layer being modified, such as wording, information content, or document structure. Optimization Action specifies the operation applied, such as inserting, deleting, or reordering. Optimization Signal identifies the cue through which the edit aims to influence the model’s judgment, such as authority or salience (Figure 1).
Figure 1. From a GEO strategy to a shared specification. Moving the existing evidence unit to the beginning changes document structure (Optimization Layer) through reordering (Optimization Action) and targets salience (Optimization Signal). The specification identifies the unit and its destination and requires all units, including , , and , to remain intact. The displayed layers and signals illustrate part of the taxonomy.
Drawing on representative methods, we map their strategies into the space. We discuss how the space can support strategy construction, search, and evaluation under shared settings. These uses motivate a community agenda for extending the space as new strategies emerge and for developing GEO leaderboards across methods, models, and domains. Such comparisons can provide a clearer basis for choosing AI systems for the tasks people intend to delegate to them.
Background and Related Work
LLMs and AI Systems
Large language models (LLMs) learn from large text corpora and can perform diverse tasks through prompting [3]. Training with human feedback further improves their ability to follow user instructions [22]. To answer questions with external evidence, retrieval-augmented systems pair language models with retrieved passages [14, 11] (Guu et al. 2020). Other systems use LLMs to rerank passages [28, 24], support conversational recommendations [8], or reason with results from external tools [34, 26]. These connections make external sources part of how AI systems answer questions and guide choices.
Generative Engine Optimization
GEO studies how editing these sources can increase their visibility in generated answers. The original GEO work introduces strategies such as authoritative phrasing, citation addition, and fluency optimization, and evaluates combinations of these strategies [1]. Related work extends source optimization to recommendations and search rankings. Strategic text sequences (STS) append optimized text to product descriptions, and conversational-search ranking manipulation searches for text embedded in source pages [13, 23]. RAF searches for short ranking-promoting text with an additional readability objective [33].
Optimization also targets editing instructions and structures. AutoGEO uses learned preference rules to guide or train a rewriter [32]. FeatGEO searches document features before generating text [16], and AgenticGEO evolves reusable strategies [37]. Structured editing choices support composition in SafeGEO [30] and constrain iterative editing in SIREN [4]. Recent work extends GEO to competing optimizers [27], detection of optimized content [6], defenses [38, 15], cross-engine audits [29], surveys [17], and governance [31].
GEO Benchmarks
Benchmarks provide shared settings for evaluating GEO methods. C-SEO Bench evaluates tasks, domains, and adoption levels using citation rank [25]. E-GEO compares heuristics, rewriters, and generative engines in a common product-ranking environment [2]. SafeGEO fixes product attributes and reference utility to measure target promotion alongside recommendation performance [30]. The ranking-manipulation GEO-Bench compares prompt-based, gradient-based, and C-SEO methods under a fixed ranker using effectiveness and detectability measures [20].
A Meta-Taxonomy of GEO Strategies
The meta-taxonomy connects the edits made to a source with the cues they aim to change. Optimization Layer and Optimization Action describe where and how the source is edited, and Optimization Signal describes what the edit aims to affect, such as the source’s apparent authority or the prominence of its evidence. The three dimensions provide a shared vocabulary for describing existing strategies and constructing a searchable space of optimization choices. Table 1 summarizes the categories with descriptions and literature mappings, and Appendix § 8 gives definitions and annotation rules.
| Description | Literature | |
| Optimization Layer | Where is the source edited? | |
| Surface form | Wording, tone, syntax, or surface tokens. | GEO (2024) |
| Information content | Claims, caveats, examples, or evidence units. | GEO (2024); SafeGEO (2026a) |
| Document structure | Order, grouping, and nesting of source units. | FeatGEO (2026) |
| Attribution and source signals | Fields linking claims to sources or credentials. | GEO (2024); SafeGEO (2026a) |
| Model-facing content | Instructions or control syntax inside a source. | Rank. manip. (2024) |
| Multimodal representation | Visual, audio, or cross-modal source elements. | MGEO (2026) |
| Optimization Action | How is it edited? | |
| Insert | Add a new unit or a copy of an existing unit. | GEO (2024); STS (2024) |
| Delete | Remove an identified unit or span. | SafeGEO (2026a) |
| Substitute | Replace a bounded object with another value. | SIREN (2026) |
| Rephrase | Re-express a span under preservation rules. | AutoGEO (2026) |
| Reorder | Change unit order without changing content. | E-GEO (2026) |
| Compress | Shorten content under a stated retention policy. | E-GEO (2026) |
| Restructure | Change grouping or nesting among source units. | E-GEO (2026); FeatGEO (2026) |
| Optimization Signal | What signal does the change target? | |
| Query matching | Correspondence to request terms or needs. | GEO (2024) |
| Evidence support | Apparent support for a claim or conclusion. | GEO (2024); SafeGEO (2026a) |
| Authority and credibility | Expertise or trustworthiness conveyed by a source. | GEO (2024); SafeGEO (2026a) |
| Salience and extractability | Prominence and ease of finding or reusing content. | FeatGEO (2026); SafeGEO (2026a) |
| Comparative framing | How a candidate is presented relative to others. | SIREN (2026) |
| Persuasion | Appeals intended to increase candidate preference. | E-GEO (2026) |
| Model control | Directions about how the model should respond. | Rank. manip. (2024); SIREN (2026) |
Table 1. The GEO meta-taxonomy and representative literature mappings. The three groups describe where sources are edited, how they are edited, and what signals the edits target. Mappings interpret reported strategies, prompts, or features, not entire optimizers or established causal mechanisms.
Optimization Layer
Optimization Layer identifies the part of a source that a strategy changes. Changes to wording, tone, syntax, or individual tokens belong to surface form, and changes to claims, caveats, or evidence belong to information content. Document structure covers how this information is ordered and grouped. Attribution and source signals covers the references and credentials that connect claims to their stated sources. Model-facing content covers instructions or control syntax embedded in a source and addressed to the model [23]. Multimodal representation includes visual and audio content and their relationship with text, with joint image and text optimization providing an example [7]. The layers describe the source itself and apply whether it is retrieved or supplied directly to a model.
An edit can affect more than one layer. We identify the smallest part of the source that captures an edit and record any other layers affected by it. For example, making a sentence sound more confident changes its surface form. If the revised sentence also makes a stronger factual claim, the edit affects information content as well. The mapping therefore considers the actual source changes and checks what the strategy had to preserve.
Optimization Action
Optimization Action describes the operation performed on the selected part of a source. The actions are insert, delete, substitute, rephrase, reorder, compress, and restructure. They may apply to a text span, an information unit, a source field, or the organization of a document. Each action specifies what it changes and what must remain intact.
For example, Reorder changes the positions of existing units without changing their contents. Rephrase changes the wording of selected text within the stated limits on meaning. Compress produces a shorter version that retains the required information, and Restructure changes how information is grouped or organized. The action describes the concrete source change, and Optimization Signal identifies the cue that the change aims to affect.
Optimization Signal
Optimization Signal describes the cue through which a strategy aims to influence how the model judges or uses a source. For example, authority and credibility concern how trustworthy or expert
| Paper | Surface form | Information content | Document structure | Attribution and source signals | Model-facing content | Multimodal representation | Insert | Delete | Substitute | Rephrase | Reorder | Compress | Restructure | Query matching | Evidence support | Authority and credibility | Salience and extractability | Comparative framing | Persuasion | Model control |
| GEO (2024) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||
| STS (2024) | ✓ | ✓ | ||||||||||||||||||
| Rank. manip. (2024) | ✓ | ✓ | ✓ | |||||||||||||||||
| RAF (2026) | ✓ | ✓ | ||||||||||||||||||
| E-GEO (2026) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |||||||||
| AutoGEO (2026) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||
| FeatGEO (2026) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||
| AgenticGEO (2026) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||
| SafeGEO (2026a) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | |||||||
| SIREN (2026) | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ||||||||||||||
| MGEO (2026) | ✓ | ✓ | ✓ |
caption: tbl:2. Representative literature mapped into the GEO strategy space. A checkmark denotes a value supported by at least one selected strategy, prompt, feature, or edit. Blank cells are unconfirmed in this mapping, not evidence of absent capability. Checks within a row do not establish that all corresponding combinations were used.
:::
a source appears, and salience and extractability concern how prominent its information is and how easily the model can find or reuse it. The other signals are query matching, evidence support, comparative framing, persuasion, and model control.
The same editing action can target different signals. Inserting a citation can aim to strengthen authority or evidence support, and inserting a copy of an existing evidence unit can aim to increase its salience. Conversely, different actions can target the same signal. Authoritative phrasing changes surface form through rephrasing, and citation addition changes attribution through insertion, but both can aim to make a source appear more authoritative [1]. Signal labels record the cue an edit targets, not its measured effect, which the evaluation tests separately. When the cue cannot be identified, as with opaque optimized token sequences, the signal is recorded as unspecified.
GEO Strategy Space
The meta-taxonomy organizes the optimization choices available to GEO methods. These choices form a strategy space in which an optimizer searches for effective edits. We define this space through the three taxonomy dimensions and use it to describe existing methods.
Defining the Strategy Space
Each optimization choice combines one Optimization Layer, Action, and Signal. Let contain the combinations an evaluation permits and the permitted strategies that use combination . The strategy space is
A combination describes a family of related strategies. For example, document structure / reorder / salience includes strategies that move different evidence units to more prominent positions. Similarly, strategies that append different text sequences can share the same layer and action. A strategy can also combine several choices and therefore belong to multiple sets . The evaluation sets the allowed changes and combinations.
Optimization over the Space
In a GEO task, a target system answers a request from a set of sources, and an optimizer seeks to raise the visibility of the target source . Applying a strategy to yields the optimized source , and the optimizer solves
where measures the source’s expected visibility in the generated response, with other sources and evaluation conditions held fixed. contains the edited sources that satisfy the evaluation’s editing constraints, including limits on editable regions, content changes, or source length.
An optimizer can select strategy families, search for edits within a family, or combine choices. Practical methods seek high-scoring strategies under the stated constraints and available resources. § 5 specifies shared comparison conditions.
Mapping Existing Literature
Table 2 maps representative strategies to the taxonomy values. The mapping shows which parts of the space each method uses and which choices remain open to optimization. Appendix § 9 gives source locations and mapping limits.
The original GEO work uses templates such as authoritative phrasing and citation addition [1]. They instantiate choices such as surface form / rephrase / authority and attribution / insert / evidence support. STS and RAF fix the Optimization Action to Insert and search for the appended text [13, 33].
Other methods search or learn instructions that guide these choices. E-GEO optimizes the rewriter’s prompt [2]. AutoGEO learns engine preference rules for prompted or trained rewriting [32]. FeatGEO searches document feature configurations before generating text [16]. AgenticGEO evolves a strategy archive and selects strategies for multiple rounds of rewriting [37]. The methods use different internal representations to explore choices captured by the meta-taxonomy.
The matrix records values supported by the selected descriptions. A combination requires evidence from the same reported strategy. The mappings support shared evaluation and further search.
A Common Evaluation Protocol
An evaluation must connect GEO strategies to a concrete task, specify the conditions under which a source is edited, and measure the resulting change in source visibility. The protocol uses a common task structure to compare GEO strategies, optimizers, and models, connecting the study of source influence to evidence about the systems people rely on for decisions. Figure 2 shows the workflow.
Figure 2. From a GEO strategy to evaluation. The example moves evidence unit to the beginning and inserts an exact copy. The resulting source is checked against the editing constraints before independent evaluation against the original-source baseline. Source visibility and original task performance are measured separately. The lower panels summarize the conditions for different comparisons. Appendix § 10 specifies both actions.
Evaluation Setting
Each task fixes the request, the sources, and the target source (§ 4.2). The task uses Optimization Layer, Optimization Action, and Optimization Signal to specify allowed GEO strategies, together with editing constraints that state which content may change and what must be preserved. These fields connect the meta-taxonomy to the choices in a task, as the illustrative GEO Task shows.
The example fixes the layer, action, and signal and leaves the ordering of source B to the optimizer. A task can also allow several strategies and state which may be combined.
Illustrative GEO Task
Request. Which tools support local PDF table extraction?
Sources. Descriptions of candidate tools, A, B, and C.
Target source. The description of tool B.
Allowed GEO strategy.
Optimization Layer Document structure.
Optimization Action Reorder.
Optimization Signal Salience and extractability.
Editing constraints. Change only the order of existing units in source B. Preserve every unit and its text. Leave sources A and C unchanged.
The evaluation setting specifies how those edits are produced and tested through the target system , visibility measure , feedback access, and resource budget. It also identifies how sources reach the target system, since supplying edited sources directly and passing them through retrieval create different evaluation conditions [12, 35].
Comparing Strategies, Optimizers, and Models
The protocol supports three types of comparison that examine GEO strategies, optimization algorithms, and target models. Each uses the same task structure but changes a different component, so the conditions held fixed depend on the question. Table 5 in Appendix § 11 summarizes these conditions.
| Comparison | Held fixed | Varied and reported |
| Strategies | Tasks, target system, editing model, validation procedure, and resource limits | Prescribed strategies, parameters, and combinations; edit validity and task outcomes. |
| Optimizers | Strategy space, editing constraints, feedback access, tasks, target system, and resource limits | Search or learned policy; text generation within the optimizer; training and per-task costs. |
| Models, fixed suite | Original and optimized sources, task evidence, context rules, and scoring | Evaluated system and configuration; task performance before and after optimization. |
| Models, adaptive | Optimizer, strategy space, feedback rules, and budgets | System-specific edits; optimization costs, coverage, and outcomes from independent evaluation. |
Table 5. Comparison conditions derived from the strategy space. These are different evaluation questions and should be identifiable in a leaderboard row. Equal layer, action, and signal labels alone do not establish equal conditions.
Strategies. A strategy comparison evaluates how the prescribed optimization choice affects source visibility. Each strategy is applied to the same original sources through a shared editing model and generation procedure, with the requests, target system, common editing constraints, and resource limits held fixed. Edit validity is reported alongside visibility to show how reliably the editing model follows each strategy.
Optimizers. An optimizer comparison evaluates how effectively different algorithms find edits that increase visibility within the same strategy space. The tasks, target system, allowed strategies, editing constraints, feedback access, and resource limits are therefore shared across methods [36]. Each optimizer controls its own search and text-generation procedure within these conditions, and its reported resource use includes any prior training as well as per-task computation.
Models. A model comparison examines how well different target systems carry out the same requests when their sources are edited. A fixed-suite comparison gives every model the same original and optimized sources, providing a common set of inputs for measuring their responses. An adaptive comparison runs the same optimizer separately against each model under shared editing constraints, feedback rules, and budgets, allowing the optimizer to find edits tailored to that model. In both settings, task performance before and after optimization is reported alongside source visibility, providing evidence of how well each system carries out the request under GEO. These results can inform system choice for the tasks and optimization conditions covered by the evaluation.
Evaluation Metrics
A common GEO evaluation should measure the visibility gained by the target source, the effect on the system’s original task performance, and compliance with the allowed strategy and editing constraints.
Source visibility. Source visibility measures how prominently the target source appears in the system’s output. The evaluation defines through a measure such as citation frequency, attributed answer share, or a ranking-based score that increases as the target moves toward the top. With , , and as in § 4.2, the visibility gain is
Reporting original visibility and optimized visibility alongside the gain distinguishes a source that was already prominent from one whose visibility improved through optimization.
Original task performance. A system entrusted with a task must still satisfy the user’s request when its sources are edited, and greater source visibility may accompany better, unchanged, or worse performance on that task. The evaluation therefore applies the same task-performance measure before and after optimization, such as answer correctness for question answering or relevance and suitability for ranking and recommendation. Reference answers, relevance judgments, or verified candidate information provide evidence independent of the optimized source [32, 30].
Edit validity. Edit validity checks whether the produced source follows the allowed strategy and editing constraints. In the Illustrative GEO Task, an optimizer may change the order of source B, but an edit that also deletes a caveat violates the task constraints even if it achieves a large visibility gain. The evaluation reports the share of attempts that satisfy these rules, together with failed or unresolved cases and their resource costs. A scoring rule declared before evaluation determines how unsuccessful attempts contribute to the overall result.
Common Leaderboards
The protocol above can support common leaderboards where new GEO methods and AI systems are evaluated against an established set of reference results. Existing benchmarks can provide the tasks, with submissions evaluated under the shared conditions in § 5.2 before their scores enter a ranking. A new optimizer would be tested on the same tasks and target systems as existing optimizers, and a new target system would be tested with the optimizers already included in the model comparison. Researchers could therefore place their contributions within an ongoing comparison that builds on earlier evaluations. Each leaderboard also reports its evaluation coverage: the strategies and combinations actually tested, together with allowed choices that remain untested, so that a score is read against the part of the space it covers.
A shared leaderboard would also connect progress in GEO methods to the evidence used for system selection. A new optimizer may reveal weaknesses in task performance that earlier methods did not expose. When it becomes part of the model evaluation suite, all systems in the updated ranking would be evaluated on the expanded suite under common conditions. Earlier rankings would remain available under their original evaluation versions. As methods and models are added, the leaderboard would accumulate comparable evidence of how well systems carry out users’ requests under GEO, helping people choose which systems to entrust with the decisions that follow.
Call for Action
We invite method developers, benchmark maintainers, and the wider research community to develop this proposal into a common GEO evaluation standard. Shared descriptions of GEO strategies and comparison conditions would allow existing methods and benchmarks to contribute comparable evidence about optimization effectiveness and system performance under GEO. This evidence would help researchers compare their methods and give people a clearer basis for choosing the systems to which they delegate decisions.
Method Developers
Method developers can support comparison by releasing representative sources before and after optimization, with editing permissions and the procedure used. Each example should identify the content to preserve and describe the changes through Optimization Layer, Optimization Action, and Optimization Signal. Examples with unclear mappings should remain available to guide taxonomy revisions. For learned optimizers, optimized sources should be released with the prompts, rules, or strategy descriptions that produced them. Reusable implementations and validation examples would help groups examine the roles of strategy search, text generation, and target-system feedback.
Developers can contribute new strategies and combinations beyond an existing evaluation. Contributions should document the source changes and the conditions that permit them. Exploratory results should remain separate from scores under a shared protocol, allowing researchers to study new strategies without changing an existing leaderboard’s comparison conditions.
Benchmark Maintainers
Benchmark maintainers can retain their tasks and metrics and map their evaluated strategies to the proposed taxonomy. This mapping should show which strategies and combinations are tested, what editing constraints apply, and which parts of the space remain untested. Following § 5, maintainers should report the feedback and resources available to each method. Readers can then identify comparable results.
When adding a strategy, maintainers should explain which tasks support it, which edits it permits, how it combines with other strategies, and how the resulting edits will be checked. Each release should have a version number, with its tasks and strategy coverage fixed before evaluation. Previous releases should remain available, and documentation should identify which earlier results remain comparable. This preserves the meaning of earlier scores as the benchmark expands.
Community Contributors
Community contributors can assess whether the taxonomy supports consistent descriptions of source changes. Independent annotators can examine the same original and optimized sources using fixed annotation guidelines. Agreement should be reported for each dimension, including disagreements about edit boundaries and cases with unclear signals. Annotators should not see rankings or optimization scores, so judgments remain based on source changes. Appendix § 8.§ 5 outlines such a protocol. Evaluation on methods and domains excluded from guideline development would help identify definitions needing clarification or extension.
Researchers can test whether a shared strategy description helps groups agree on the permitted edits, the content to preserve, and which outputs violate the rules. The mappings in § 4 provide examples for these checks. Groups can independently implement the same strategies and document disagreements that affect task eligibility, edit validity, or result interpretation. The studies would show whether the descriptions allow researchers to reproduce an evaluation’s editing conditions without rebuilding the original optimizer.
Conclusion
Humanity’s Last Choice highlights the importance of choosing the system that will make subsequent decisions on a person’s behalf. Personal agents now make this choice concrete, and the sources they consult are already being optimized to influence them. A common GEO evaluation standard can inform that choice by comparing how well systems perform under such optimization. Our meta-taxonomy describes where sources are edited, how they are edited, and what signals the edits target. Literature mappings place existing strategies in this space, and the proposed protocol supports comparisons of strategies, optimizers, and models under explicit conditions. Versioned releases let the taxonomy grow with new strategies while earlier scores keep their meaning. We invite method developers, benchmark maintainers, and the wider community to describe their strategies in this vocabulary and report the conditions and coverage behind their scores. Shared benchmarks and independent validation can turn this foundation into evidence that helps people choose systems suited to the tasks they entrust to them.
Limitations
This paper does not empirically establish whether adopting the proposed meta-taxonomy and shared evaluation standard improves optimization effectiveness or original task performance. Edits across source layers may also be coupled, as changing document structure can require changes to wording or information content. Such dependencies require case-specific analysis when defining permitted edits and interpreting the contribution of individual layers.
Alternative view. An alternative view favors domain-specific evaluation and open-ended strategy discovery, since a fixed common strategy space may overlook differences in available evidence and exclude new strategies beyond its current categories. The proposed standard therefore retains domain-specific tasks, metrics, and evidence requirements and allows the taxonomy to evolve as new strategies emerge, with each evaluation release defining the scope within which results can be compared.
Ethical Considerations
GEO is dual-use: a shared strategy space can help researchers evaluate systems exposed to optimized sources, but it can also help content providers steer those systems. We propose no new optimization method and release no optimizer; the mapped strategies are already published. The protocol reports original task performance and edit validity alongside visibility, so gains at the user’s expense are exposed rather than rewarded. Editing a source is not harmful by definition: clarity can help users, whereas apparent authority or unsupported claims can mislead. Versioned releases and held-out tests limit narrow optimization for a leaderboard, and contributors should document source access restrictions and the basis for claims about candidate quality. The paper involves no human subjects, new data, or model experiments.
Annotation Codebook
Annotation Units and Evidence
Annotate a concrete change between a source snapshot and its edited output. When only a strategy description is available, mark the record as a family-level mapping; it cannot certify what a particular implementation changed. Store the paper version and a section, table, prompt, or artifact locator. Do not assign one layer, action, and signal to an entire optimizer when it can generate different edits.
For an instance, retain the source representation, parser version, unit identifiers, and exact diff. Segment a composite change into actions with independently describable scopes and preservation conditions. If segmentation is ambiguous, preserve the alternative decompositions. Do not count their union as additional experimental coverage. Specify whether edit addresses refer to the original source or to each intermediate output.
Optimization Layer Definitions
Surface form. Expression within an identified span: lexical choice, syntax, tone, typography, or an uninterpreted token span. A claim-preserving rephrasing belongs here only when its declared preservation condition is satisfied. Opaque text is not automatically a model-control instruction.
Information content. Claims, caveats, examples, or evidence units. Adding a claim, removing a qualification, or repeating a complete evidence unit changes content at this level. Repetition can preserve the set of claims while changing their multiplicity.
Document structure. Order, grouping, nesting, and presentation relations among units. Moving an unchanged paragraph is reorder; changing prose into a list is restructure only if the unit mapping and permitted linguistic changes are explicit.
Attribution and source signals. A citation field, authorship or credential field, source identifier, or relation between a claim and its attributed support. Record where the field is physically stored and whether the supplied evidence supports the attribution. An authority-like style alone is a surface-form edit.
Model-facing content. Source-embedded instruction spans, role-like delimiters, or control syntax. The locator identifies the span and its body, metadata, or attribute container. The model’s actual system prompt remains outside the source-edit permission.
Multimodal representation. Image regions, rendered badges, graphical text, audio segments, or cross-modal relations. Require coordinates or time bounds, representation and rendering versions, and modality-specific preservation conditions. This layer requires modality-specific implementations.
Select a primary layer using the smallest addressable object that captures the action. Record affected secondary layers and check their preservation conditions. This is an annotation convention, not a claim that source representations have mutually exclusive physical regions.
Optimization Action Definitions
Insert. Add a specified unit or span at a declared position. repeat is an insertion parameter that references an existing unit and records its multiplicity; it is not a separate action.
Delete. Remove identified content. A validator checks the deletion mask and protected units. Removing a caveat must be declared as deletion or content change even if the remaining text is fluent.
Substitute. Replace a bounded object with a supplied value. Record both values. Replacing an attribution and replacing a claim share an action but differ in layer and preservation conditions.
Rephrase. Re-express a specified span under explicit semantic and structural preservation conditions. Store the rephrasing prompt or policy and validator evidence. A lexical similarity score alone does not establish claim preservation.
Reorder. Change the order of stable units. Verify the exact multiset of unit contents and the declared ordering relation. Byte offsets alone require rebasing after earlier edits.
Compress. Shorten specified content with a declared retention policy. Record removed units or clauses and check protected caveats. A shorter document does not establish a valid compression.
Restructure. Change grouping or nesting relations among units under a declared mapping. If wording changes, record the associated rephrasing as another step or declare a combined implementation with both checks.
Optimization Signal Definitions
Signals are query matching, evidence support, authority and credibility, salience and extractability, comparative framing, persuasion, and model control. A benchmark may subdivide them. Store both declared intent and the annotation supported by the edit description or instance. Use multiple labels when supported; identify their evidence independently. Query matching concerns alignment with request terms or requirements. Evidence support concerns apparent grounding; authority concerns the status or credibility attributed to a source. Salience concerns prominence or ease of extraction. Comparative framing changes how alternatives are presented relative to one another; persuasion changes appeals for preference. Model control presents instructions about system behavior within source material.
Use unspecified when the signal cannot be justified. It is exclusive of positive signal assignments in an annotation. Record uncertainty as metadata, not as an extra taxonomy dimension. Never infer a signal from a successful outcome alone. An evidence-support label records a targeted cue; evidence validity is a separate condition and may fail.
Proposed Annotation Protocol
Freeze the codebook before a held-out annotation round. At least two annotators independently receive the original source, edited source, permitted scope, and any declared plan; they do not see the model’s ranking or success label. Collect primary layer, secondary layers, action sequence, declared and inferred signals, uncertainty, and evidence spans. Report per-dimension agreement, unit-segmentation disagreements, and unspecified rates before adjudication. For multiple signal tags, report set agreement as well as per-label statistics. Preserve the initial annotations, adjudication changes, and codebook version. No independent annotation results are claimed in this paper.
Literature Mapping
Table 3 and 4 distinguish our proposed mappings from facts in the sources. We selected work that covers named strategies, token optimization, learned rules, feature configurations, evolving strategies, and shared evaluators. The mapping is scoped to the listed strategies. Blank or uncertain fields remain unresolved rather than being filled from reported success.
| Source and locator | Strategy described | Layer | Action | Signal and mapping limit |
| GEO ([1]), §2.2.2, App. B.4–B.5 | Authoritative phrasing | Surface form | Rephrase | Authority; check whether the rephrased text also strengthens claims. |
| GEO, same locations | Citation addition | Attribution | Insert | Authority / evidence support; added evidence content requires another operation. |
| STS ([13]), §2 | Optimized appended sequence | Surface form, appended span | Insert | Unspecified unless the actual sequence supports a signal label. Gradient search is method metadata. |
| Conversational ranking ([23]), method and prompt examples | Page-embedded ranking instruction | Model-facing content | Insert | Model control for the instruction example; other generated text needs its own mapping. |
| RAF ([33]), §3 | Ranking-promoting text with readability objective | Surface form, appended span | Insert | Unspecified at the family level; readability constrains the generated text. |
| AutoGEO ([32]), method and learned-rule prompts | Rewriting under learned rules | Per actual edit | Rephrase or composition | Signal depends on the selected rule and diff. Rule extraction and training do not determine the layer, action, or signal. |
| FeatGEO ([16]), Table 1, §3.3 | Increase source-citation feature | Attribution; possible content change | Insert / rephrase | Evidence / authority, inferred from the feature. A configuration alone does not specify an exact edit. |
| FeatGEO, same locations | Introductory summary and list features | Document structure; content | Restructure / insert | Salience / extractability. Verify retention of units in the generated page. |
| AgenticGEO ([37]), §4.3, §4.5, App. A.1 | Archived strategy applied through multiple rewriting rounds | Per round | Ordered composition | Per-step signal or unspecified. The archive contains method-level policies, not one fixed edit form. |
| SafeGEO ([30]), §3.2, tbl:2 | Model-directed instruction | Model-facing content | Insert | Model control for the source-embedded directive; map its other edits separately. |
| SIREN ([4]), §5–6, App. C | Declarative ranking claim in a metadata line | Information content; metadata container | Insert | Comparative framing / persuasion. A ranking claim need not be a model-control instruction. |
| SIREN, same locations | Directive in a link or image attribute | Model-facing content; attribute container | Substitute | Model control. Carrier and deterministic encoding remain execution parameters. |
Table 3. Representative strategies from the literature mapped into the meta-taxonomy, giving each strategy’s Optimization Layer, Action, and Signal. The mappings are our interpretations of the listed strategy descriptions and apply at the family level.
Symbolic Strategy Specifications
The following example describes the semantics of optimization actions using abstract source units. It represents no recommendation case, dataset item, or measured outcome. Let be a source with four addressable units whose contents remain fixed. Figure 1 and Figure 2 label their contents , , , and . A reorder specification moves to the beginning:
Its Optimization Layer is document structure, its Optimization Action is reorder, and its Optimization Signal is salience and extractability. Parameters identify and its destination. A valid edit preserves
| Closest standardization work | What the source contributes | Connection proposed here |
| Original GEO-bench | Shared queries, sources, visibility metrics, and strategy combinations | Record the actual edits and their preservation conditions. |
| C-SEO Bench | Shared tasks, domains, citation-rank measurement, and adopter settings | Declare the allowed strategies separately from task and competition settings. |
| E-GEO | Common product-ranking environment, rewriter comparisons, and prompt optimization | Record both permitted edits and the edits produced under learned instructions. |
| FeatGEO | Interpretable feature configurations separated from text generation | Specify observable edit semantics and validation across methods. |
| SafeGEO | Controlled source edits, composed strategies, and utility-grounded recommendation evaluation | Map its individual and composed strategies to shared specifications and task adapters. |
| SIREN | Technique specification, constrained iterative editing, and fixed-source replay | Preserve its carrier and encoding detail while adding composition across strategies and coverage reporting. |
| Ranking-manipulation GEO-Bench | Common datasets, ranker, metrics, and access-based method comparison | Add a declared strategy space and coverage reporting to that evaluator. |
| TextAttack (Morris et al., [19]); search-method benchmarking (Yoo et al., [36]) | Reusable transformations and constraints; control of space, algorithm, and budget | Apply the decomposition to GEO source representations and task adapters. |
caption: tbl:4. The proposed increment is an interface connecting existing forms of standardization. This table does not claim that the cited systems have implemented the interface or that their broader functionality is exhausted by the listed contributions.
:::
every unit’s identifier and text, as well as the source metadata. Rephrasing would violate this specification even if it produced the same order.
schema_version: geo-strategy/0.1-draft
strategy:
layer: document_structure
action: reorder
signal:
- salience_extractability
parameters:
unit: u_e
destination: first
preserve:
unit_text: exact
unit_multiset: exact
source_metadata: exact
validation: exact_units_and_order
A second specification, Copy, inserts one exact copy of immediately after the current first unit. The copy receives a distinct identifier, ; Figure 2 labels both occurrences because their contents are identical. Its Optimization Layer is information content and its Optimization Action is insert; it also targets salience. The preservation rule allows exactly one additional copy and requires every other unit to remain unchanged.
Order affects the resulting source:
Here Move addresses the original through its stable identifier. The second operation interprets its position parameter against the intermediate source. Thus even a composition of deterministic actions needs explicit address and order semantics. If a later operation removes a protected caveat, the complete composition fails its global preservation condition despite earlier steps being valid.
Free-form rephrasing needs a different validator. Exact-unit checks cannot establish preservation after paraphrasing. Its specification must name the claims and caveats that should survive, the permitted expressive changes, and the procedure for resolving uncertain semantic judgments. These are requirements for an implementation; the symbolic example supplies no evidence about an editing model’s compliance.
Comparison and Reporting Requirements
Table 5 lists the conditions held fixed and varied in each comparison of § 5.2.
The shared record has four parts, only the first of which uses the taxonomy: (i) the strategy’s layer, action, and signal; (ii) execution specification and artifacts; (iii) comparison conditions; and (iv) results. The comparison conditions include suite and schema versions, source and parser hashes, task and split identifiers, permitted combinations, source scope, how sources reach the target system, target-system configuration, editing procedure, feedback permissions, resource accounting, validation procedure, metric definitions, repetitions, and aggregation weights. Results preserve original model outputs and parser failures. Edit validity and task evidence are stored independently of the optimized source.
A benchmark release declares its permitted combinations before evaluation. Report attempted and validated coverage with separate denominators, and list combinations that are inapplicable, unimplemented, untested, failed, or unresolved. Inapplicability must be decided from the original task and specification, not from a method’s failure. A benchmark cannot improve its apparent coverage by removing failed combinations after a run. For open-ended parameter ranges, predeclare discretization or use an explicit distributional summary instead of reporting a finite completeness percentage.
Freeze the scoring rule for invalid outputs and exhausted budgets. For optimizer comparison, report both valid-output effectiveness and an unconditional task-level estimate with invalid attempts assigned the preregistered outcome. If unresolved validation prevents an unconditional estimate, report bounds or withhold that aggregate. For model comparisons, report task performance on the original sources beside task performance on the optimized sources. The protocol must not reward a system that limits visibility gains merely by declining all useful recommendations.
References
- [1]Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande. 2024. GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5–16.arxiv.org/abs/2311.09735
- [2]Puneet S. Bagga, Vivek F. Farias, Tamar Korkotashvili, Tianyi Peng, and Yuhang Wu. 2026. E-GEO: A testbed for generative engine optimization in e-commerce. Preprint, arXiv:2511.20867. Version 2, July 2026.arxiv.org/abs/2511.20867
- [3]Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhar iwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, and 12 others. 2020. Language models are few-shot learners. In Advances in Neural Information Processing Systems, volume 33.arxiv.org/abs/2005.14165
- [4]Evan Gaville, Spencer Kayser, Siamak Layeghy, Billy Sung, Sara Dolnicar, and Marius Portmann. 2026. SIREN (unrin LLMs onto the rocks): Pair-driven preference manipulation in web-RAG recommenders. Preprint, arXiv:2607.21951. Version 2, August 2026.
- [6]Junjie Chu, Ye Leng, Mingjie Li, Yun Shen, Xinyue Shen, and Yang Zhang. 2026. GEO-Flag: Detecting and measuring GEO-optimized web content. Preprint, arXiv:2608.16824.arxiv.org/abs/2608.16824
- [7]Yixuan Du, Chenxiao Yu, Haoyan Xu, Ziyi Wang, Yue Zhao, and Xiyang Hu. 2026. Multimodal generative engine optimization: Rank manipulation for vision-language model rankers. Preprint, arXiv:2601.12263. Version 2.arxiv.org/abs/2601.12263
- [8]Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. 2023. Chat-REC: Towards interactive and explainable LLMs-augmented recommender system. Preprint, arXiv:2303.14524.arxiv.org/abs/2303.14524
- [9]Google. 2026. The Gemini app becomes more agentic, delivering proactive, 24/7 help. Google blog, May 19, 2026.
- [11]Gautier Izacard and Edouard Grave. 2021. Leveraging passage retrieval with generative models for open domain question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 874–880. Association for Computational Linguistics.arxiv.org/abs/2007.01282
- [12]Sunghwan Kim, Wooseok Jeong, Serin Kim, Sangam Lee, and Dongha Lee. 2026. SAGEO Arena: A realistic environment for evaluating search-augmented generative engine optimization. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2, pages 2342–2353.
- [13]Aounon Kumar and Himabindu Lakkaraju. 2024. Manipulating large language models to increase product visibility. Preprint, arXiv:2404.07981.arxiv.org/abs/2404.07981
- [14]Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, volume 33.arxiv.org/abs/2005.11401
- [15]Haozhang Li, Yangguang Shao, Xinjie Lin, Zhong Guan, Mi Zhou, and Junzheng Shi. 2026. When optimization becomes manipulation: Defending generative search against malicious generative engine optimization. Preprint, arXiv:2609.02964.arxiv.org/abs/2609.02964
- [16]Zikang Liu and Peilan Xu. 2026. Think before writing: Feature-level multi-objective optimization for generative citation visibility. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 20290–20303.
- [17]Olivier Martinez. 2026. Optimizing visibility in generative engines: A critical survey of generative engine optimization (2023–2026). Preprint, arXiv:2607.14035.arxiv.org/abs/2607.14035
- [18]Meta. 2026. Introducing Muse: The world’s first personal AI agent built for everyone. Meta Newsroom, September 8, 2026.
- [19]John Morris, Eli Lifland, Jin Yong Yoo, Jake Grigsby, Di Jin, and Yanjun Qi. 2020. TextAttack: A framework for adversarial attacks, data augmentation, and adversarial training in NLP. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 119–126.arxiv.org/abs/2005.05909
- [20]Ojas Nimase, Zhe Chen, Genpei Qi, Yue Zhao, and Xiyang Hu. 2026. GEO-Bench: Benchmarking ranking manipulation in generative engine optimization. Preprint, arXiv:2605.29107.arxiv.org/abs/2605.29107
- [21]OpenAI. 2026. Introducing dots. OpenAI, September 29, 2026.
- [22]Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In Advances in Neural Information Processing Systems, volume 35.arxiv.org/abs/2203.02155
- [23]Samuel Pfrommer, Yatong Bai, Tanmay Gautam, and Somayeh Sojoudi. 2024. Ranking manipulation for conversational search engines. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 9523–9552.
- [24]Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023. RankZephyr: Effective and robust zero-shot listwise reranking is a breeze! Preprint, arXiv:2312.02724.arxiv.org/abs/2312.02724
- [25]Haritz Puerto, Martin Gubri, Tommaso Green, Seong Joon Oh, and Sangdoo Yun. 2025. C-SEO Bench: Does conversational SEO work? In Advances in Neural Information Processing Systems 38, Datasets and Benchmarks Track.
- [26]Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, volume 36.arxiv.org/abs/2302.04761
- [27]Vaibhav Sourirajan, Yao Zhang, Himanshu Kumar, Sahil Wadhwa, Mann Patel, and Amirfarrokh IraniTalab. 2026. Beyond the vacuum: Combinatorial strategy selection for competitor-aware generative engine optimization. Preprint, arXiv:2608.27631.arxiv.org/abs/2608.27631
- [28]Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. Is ChatGPT good at search? Investigating large language models as re-ranking agents. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 14918–14937.arxiv.org/abs/2304.09542
- [29]Benjamin Tannenbaum. 2026. Scoring with the engine: Retrieval exposure, cross-engine divergence, and the limits of engine-agnostic GEO scores. Preprint, arXiv:2609.22655.arxiv.org/abs/2609.22655
- [30]Qianfeng Wen, Yifan Simon Liu, Xin Liu, Difan Jiao, Blair Yang, Junda Wu, and Zhenwei Tang. 2026a. SafeGEO: Understanding generative engine optimization risks in recommendation agents. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing.arxiv.org/abs/2606.28356
- [31]Yizhu Wen, Nan Zhang, Haohan Yuan, Xun Chen, Haopeng Zhang, and Hanqing Guo. 2026b. Position: Generative engine optimization creates underexamined risks, governance must target concentration, disclosure, and academic blind spots. In Proceedings of the 43rd International Conference on Machine Learning, Position Paper Track.arxiv.org/abs/2606.12439
- [32]Yujiang Wu, Shanshan Zhong, Yubin Kim, and Chenyan Xiong. 2026. What generative search engines like and how to optimize web content cooperatively. In International Conference on Learning Representations.arxiv.org/abs/2510.11438
- [33]Tiancheng Xing, Jerry Li, Yixuan Du, and Xiyang Hu. 2026. Are LLMs reliable rankers? Rank manipulation via two-stage token optimization. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9120–9132.arxiv.org/abs/2510.06732
- [34]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations.arxiv.org/abs/2210.03629
- [35]Yu Yin, Shuai Wang, Bevan Koopman, and Guido Zuccon. 2026. Can it reach the generator? Investigating the survival of prompt-injection attacks in realistic RAG settings. Preprint, arXiv:2605.28017.arxiv.org/abs/2605.28017
- [36]Jin Yong Yoo, John Morris, Eli Lifland, and Yanjun Qi. 2020. Searching for a search method: Benchmarking search algorithms for generating NLP adversarial examples. In Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 323–332.arxiv.org/abs/2009.06368
- [37]Jiaqi Yuan, Jialu Wang, Zihan Wang, Qingyun Sun, Ruijie Wang, and Jianxin Li. 2026. AgenticGEO: A self-evolving agentic system for generative engine optimization. Preprint, arXiv:2603.20213.arxiv.org/abs/2603.20213
- [38]Bing Zheng, Zongyao Zhao, and Wenming Yang. 2026. Counter-GEO-Bench: Evaluating defenses against information-distorting generative engine optimization. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing.arxiv.org/abs/2609.02316

