acceptodds
Under review as a conference paper at ICLR 2027

Who Gets Cited? A Game-Theoretic Approach to Generative Engine Optimization

Abstract

Generative search gives publishers an incentive to optimize their sources for visibility, and a source can become more visible by changing how it looks rather than by improving what it says. Existing defenses against generative engine optimization (GEO) are static: publishers can arbitrage them once the policy is set. We introduce Nash-GEO, a game-theoretic method for designing platform policies that reward useful supporting material over presentation manipulation. The platform searches for publisher edits that gain citation exposure under its policy and uses those responses to refine the policy. We prove that a presentation edit offers no citation advantage on any engine if the platform removes it before generation. If such edits cost anything, no publisher benefits from making them, regardless of others' actions. Once citation positions are full, GEO redistributes exposure rather than creating it: even if every publisher optimizes, total exposure cannot increase. Across multiple generative engines, the policy reduces manipulation's advantage over adding supporting material from 4.99% to 0.51% for two search-generated strategies, and to 2.32% when additional strategies are tested. Adding supporting material outperforms presentation editing alone by 7.0%, with an advantage on every engine. These findings show how generative search platforms can shape competition among publishers by rewarding content improvements over presentation manipulation.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
generative engine optimization, generative search, citation allocation, game theory, Stackelberg commitment, adaptive evaluation, adversarial robustness, incentive design, retrieval-augmented generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_JhPPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Nash-GEO, a game-theoretic framework that helps platforms to set policies to reduce the gain from presentation GEO edits from the publishers. It formulates the interaction between the publisher and the platform as a Stackelberg game, where the platform designs defense policies, and publishers modify content to increase citation exposure based on the platform's policies. Nash-GEO aims to discourage presentation edits while rewarding useful supporting material through adaptive strategy evaluation and policy refinement.

Critical strengths
  1. The paper adapts the game theory strategy in the GEO framework. This application could potentially solve the arm-race GEO game between the platform and the publisher. This Stackelberg formulation opens up a new direction in the GEO regulation field.
Critical weaknesses
  1. The assumption is unrealistic. Nash-GEO assumes it has access to the original document before and after edits. In reality, the platform would not know whether the current version of the website has been edited for GEO or not, which keywords were added recently, or which content belongs to presentation manipulation. So the screening stage will fail in the real world.

  2. There is no threat model section to clearly identify the platform and publishers. What are their capabilities and goals? Are Internet users involved in the framework?

  3. In the introduction Section, it is better to give examples of presentation edits using common keywords. The main result presented in the introduction section is not briefly explained. Are the metrics higher the better or lower the better? This helps improve readability, even though there is a more detailed explanation in the experiment section, because this metric is not that common.

  4. Some sentences are too short and not self-explanatory. For example: "The platform moves first, the Stackelberg setting (Conitzer & Sandholm, 2006)", "strategic classification studies rules that subjects adapt to (Hardt et al., 2016; Milli et al., 2019)". The related work section or background section is designed to help readers understand the basic knowledge that is connected to this paper, but not just listing the academic papers' names.

  5. "Earlier citations receive more exposure." This sentence needs a citation since the framework uses ranking-related metrics below.

  6. 3 metrics are introduced here: normalized reciprocal rank, discounted cumulative gain, and rank-biased precision, but only Unnormalized Reciprocal Rank is used. This is inconsistent.

  7. Some mathematical notations have no explanation, which will confuse readers. What is the value of viv_i, the editing cost ci(ai)c_i(a_i), zz revealed by the platform, AivalidA_i^{\mathrm{valid}}, Valid⁡\operatorname{Valid},engine probability, d−i\textbf{d}_{-i}, Z→B\mathcal{Z} \to B, H(W)=ΦLK(W,aH)H(W) = \Phi_L^K(W, a_H), M(W)M(W) etc.? Every mathematical notation in the paper needs to have a formal definition. The paper introduces too many mathematical notations. It would be better to simplify it and help improve readability.

Questions for authors

See the weaknesses;

  1. The whitespace-only normalization also has a low manipulation advantage. Does Nash-GEO also include the whitespace-only normalization operation? It seems there is no benefit from Screening. Can you clarify what additional benefits Nash-GEO provides beyond simple text preprocessing?

  2. Why do you only compare to SafeGEO-L1? SafeGEO has a complete framework from L1 to L5. If you choose SafeGEO as your main benchmark, it is better to have a complete comparison rather than just comparing with the basic SafeGEO-L1.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JhPP (Weakness)

We thank the reviewer for raising the distinction between the policy's inputs and what we need to measure an edit's effect. We respectfully clarify that the screen, but not the normalizer, needs a reference copy. Differences below use the paper's absolute percentage scale.

W1. Access to the document before and after editing

Section 4.2 now distinguishes what the two components need. The normalizer reads only the submitted text. The screen also needs a reference copy and the declared supporting sentence. A previous crawl is not necessarily an unmanipulated reference.

Section 5.3 now reports the “Deletion only” result from Table 17 of the submission. This runs the full normalizer without the screen, not just whitespace normalization. On the original 200 questions, four engines, and two searched edits, it lowers manipulation's advantage from 4.99% to 0.51%, the same result as the full policy. Thus, for the headline comparison, a platform without a reference copy can use the normalizer and skip the screen.

Comparing edited with unedited outcomes measures an edit's gain. The normalizer does not read both versions. Proposition 2's recovery benchmark does assume the original source and correct support identification, unlike the normalizer.

W2. Threat model: publishers, platform, and users

We agree that the actors' roles should be stated together. Section 3.2 now opens with a threat-model paragraph that describes the publishers, the platform, and the users.

  • Publishers edit their own sources to gain citation exposure. In the main setting they know the policy but not the answering engine. Admissible edits preserve content and provenance, meet a length budget, and add no unsupported claim. Prompt injection is excluded (Sections 3.2-3.3).
  • The platform commits first and processes sources before generation. It minimizes manipulation return subject to quality and cost constraints, without retraining the engine or changing retrieval.
  • Users submit queries and receive answers. They are not strategic players. The platform's quality constraint represents their interest.

W3-W7. Examples, metrics, related work, and notation

We recognize that these passages were too compressed.

  • W3: examples and metric direction. Section 1 now uses the example from Appendix C.3: a keyword edit adds “champions, league, final:” before an unchanged source. It also explains that lower is better. The strongest tested edit earns 4.99% more citation exposure than adding support without defense, versus 0.51% under our policy.
  • W4: related work. Section 2 now explains that the platform commits before publishers act. It connects this to strategic classification, where people adapt to an announced decision rule.
  • W5: earlier citations. Section 3.1 now cites Aggarwal et al. (2024) for position-adjusted GEO visibility. Giving earlier citations higher weight is part of our measure, not a new finding about user attention.
  • W6: three metrics. Section 3.1 keeps the general family of weights but removes the unused metric names. Sections 3.1 and 5.1 now give the weights used in the main evaluation: 1, 1/2, and 1/3 for the first three distinct cited sources, without rescaling.
  • W7: notation. Sections 3.1 to 3.3 now explain publisher value and cost, what publishers observe, and how the response rule uses that signal. For reference, here are the requested definitions:
SymbolMeaning
viv_iPublisher ii's value per unit of exposure. We report citation exposure, not estimated monetary value.
ci(ai)c_i(a_i)Publisher ii's cost of edit aia_i. The reported citation gains are before editing costs.
zzThe observed platform signal. The policy is public, but the signal gives no information about the engine in the main setting.
Validi\mathrm{Valid}_i, AivalidA_i^{\mathrm{valid}}The admissibility rule and the set of edit procedures that satisfy it.
rhr_hThe probability that engine hh answers a query.
d−i\mathbf{d}_{-i}The sources submitted by publishers other than ii.
g:Z→Bg:\mathcal{Z}\to BA rule that chooses an edit from the tested set based on what the publisher observes.
H(W)H(W)The expected exposure lift of adding support under platform procedure WW, relative to no edit.
M(W)M(W)The largest expected lift of a tested manipulation response rule under WW.
AuthorsOpenReview ↗

Response to Reviewer JhPP (Quesitons)

Q1. Benefit beyond whitespace and text preprocessing

Yes. The full normalizer includes whitespace normalization. On the same two searched edits, whitespace alone leaves a 1.94% advantage and our policy 0.51%. The paired reduction is 1.43% (submitted Appendix D.10, Table 18).

In a separate exploratory analysis on the same 200 questions and four engines, we used engine scores to choose edits against each policy rather than giving them the same edits. The advantage was 3.27% under whitespace and 0.68% under ours. Whitespace explains about two-thirds of the original reduction. The difference on the fixed nine-edit set is unresolved.

The policy is preprocessing by design. Nash-GEO evaluates and selects such policies against publisher responses, not only a fixed edit list. In the submitted label-repair study, renewed search found no established advantage for a promising revision, so it was not adopted (Section 5.2 and Appendix D.4).

The screen makes no difference to the headline. It improves the gap by 0.98% in the separate QAMPARI test of added query terms (Section 5.3). That is why we report the screen separately rather than make it a requirement for using the normalizer.

Q2. Comparison with SafeGEO-L1 to L5

The submission used L1 as a prompt-level baseline, not a test of the whole SafeGEO framework. We now include L2–L5. Each policy saw the same edits on the 200 ASQA questions and four engines.

PolicyManipulation advantage (%)
No defense8.32
SafeGEO-L18.40
SafeGEO-L28.13
SafeGEO-L36.58
SafeGEO-L48.27
SafeGEO-L58.06
Our policy1.06

Thanks for your clarification; here is my response.

W2: Your response did not fully resolve my concerns about the reality assumption. Showing "Deletion only" matches the full policy (0.51%) on the two searched edits does not prove that the screen is optional in reality. In your Appendix D.3, it says every admissible submission passes the screen by construction, and Appendix D.9 constrains searched candidates to survive it. In practice, publishers have no such constraints. So, skipping the screen exposes the platforms to an unconstrained environment. In your experimental setup, the screen adds a +0.98% improvement in Table 8. Or are you trying to claim that "screen" is not that helpful?

Without the original document, both the screen and Proposition 2 become undeployable. Then the only thing left is the rule-based deletion normalizer. In Proposition 2(b) and Proposition 11, it says rule-based methods suffer from fundamental theoretical limitations and can sacrifice the citation reward for useful supporting material.

W3: I can't find the content that you said was added to section 1.

Reviewer_4qzkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In generative search, an LLM decides which sources to cite and how to rank the citations. Therefore, publishers have an incentive to use presentation tricks to improve their rankings so that their sources get more exposure. However, the action we actually want to incentivize is honest action (i.e., adding supporting evidence to improve the source). The paper formalizes this process as a Stackelberg game, where the leader (platform) first commits to a defense policy for how to “clean up” the sources before feeding them into the LLM, and the followers (publishers) then best respond. Under their model, they prove that if a presentation edit can be removed, it has no impact on citation ranking. The paper shows that the proposed defense method reduces the advantage of presentation manipulation compared to honestly adding supporting material.

Critical strengths

[1] This paper writes the motivation part clearly. It motivates the incentive of publishers well, and the Stackelberg game formulation is novel.

[2] The experiment section has a relatively strong robustness check. For example, they show the presentation manipulation works under different model families

Critical weaknesses

I find the paper hard to follow. Many important concepts are not discussed or discussed very briefly. Just to give a few examples of places where it’s hard to follow,

[1] there is no explanation of the description of ASQA when it’s introduced in Section 5. I think a description of the benchmark is very helpful for understanding the experiment setting.

[2] In Proposition 1 (b), what does the word identify mean? Either a mathematical definition or an intuitive explanation should appear in the main text.

[3] Line 162 introduces “empirical scores” without connecting this concept to the exposure measure discussed earlier.

[4] Line 119 is a broken sentence.

Overall, I think the clarity of the writing should be improved a lot. Having examples about presentation tricks and supporting evidence in the main text is also very helpful.

Questions for authors

[1] The paper treats adding supporting evidence as good edits, but what if the supporting evidence is not good quality content?

[2] I am very surprised that larger models are still very easily manipulated. I think larger models should be able to rank supporting evidence as higher compared to presentation tricks. Is there any intuition behind that? Does this phenomenon scale to frontier-sized models (I think API calls is enough to carry out the experiment so hopefully this is not too much to ask for during rebuttal).

[3] Reading Figure 2, I think the the proposed method did not significantly outperform simple methods like whitespace (not statistically significant in any settings). Is this understanding correct?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 4qzk

We thank the reviewer for the careful reading. We recognize that the benchmark, theory, and metrics need clearer explanations. Differences below use the paper's absolute percentage scale.

W1-W4. Presentation and examples

The revision now explains these items where readers first encounter them.

  • W1: ASQA. ASQA contains ambiguous factoid questions that call for answers covering several interpretations (Stelmakh et al., 2022). Each question in our study has five retrieved Wikipedia passages that compete for citations. Section 5.1 now explains this setup.
  • W2: "identify" in Proposition 1(b). The full input must determine what the engine would have received without the edit and, if the inputs differ, which source was edited. Think of it as enough information to tell the engine who changed their source. An engine could then give that source the top citation whenever the edit survives. Part (a)'s guarantee for identical inputs does not need this assumption.
  • W3: "empirical scores." Section 3.1 now connects the definition to the weights used in the main evaluation: 1, 1/2, and 1/3 for the first three distinct cited sources.
  • W4: line 119. The revision replaces the fragment with an explanation: in a Stackelberg game one player commits before the others respond. Here, the platform commits to a policy before publishers edit.
  • Examples. The introduction now gives a short example from Appendix C.3. On a question about the 2018 Champions League final, a presentation edit adds "champions, league, final:" to an unchanged source. The support action adds a retrieved sentence about 2Cellos performing the anthem.

Q1. What if the supporting material is low quality?

We recognize that retrieval alone does not guarantee useful support. A fixed rule retrieves the passage from outside the five shown sources. The publisher does not freely write it. The submission states that its support for the source's claims is not independently verified (Section 5.1).

To see whether the reduction also appears when the added passage is judged to support the source, we restricted the original two-edit comparison to the 387 sources across 144 questions that a model judge labeled this way. On that subset, our policy reduced manipulation's advantage by 4.62% relative to no defense. Neither the passage label nor the policy independently verifies that the passage is true or useful.

Q2. Larger and frontier models, and why presentation changes matter

Here we changed the answering model and kept the 200 questions and two original searched edits fixed. The table shows how much more the strongest tested edit earned than adding a supporting passage. We report each added model separately.

Answering modelManipulation advantage without defenseUnder our policy
Original four 7B-8B engines, pooled4.99%0.51%
Gemma-2-9B5.89%0.87%
Qwen2.5-14B4.69%0.16%
Qwen3-14B6.46%1.55%
Qwen2.5-32B4.32%0.21%

Without defense, the added models through 32B still reward the tested edits by 4.32% to 6.46% more than the passage. The advantage is lower under our policy in each reported row.

We also tested GPT-5 mini on the same ASQA questions and two original edits. Manipulation's advantage was 0.14% without defense and 1.74% under our policy, so this test did not show a reduction. In a separate study with the same API model, two simulated publishers competed for citations on four synthetic tasks. On 64 new questions, the verification policies chosen for each task achieved 98.05% correctness with a valid citation and verified the sources on 30.47% of answer draws. This shows how the framework can choose verification policies for an API engine. It does not show that the original text normalizer transfers.

A possible explanation is that, among passages retrieved for the same question, the engine may use wording and placement as cues to relevance. Adding a keyword label puts the question's terms first without adding information. Our tests measure the change in citations, not why the engine makes that change.

Q3. Does the policy significantly improve on whitespace?

The interpretation is partly right: whitespace explains about two-thirds of the reduction on the original two edits, and the difference on the fixed nine-edit set (2.76% versus 2.32%) is unresolved. But overlapping error bars in Figure 2 do not test the difference between policies. Pairing results by question on the two original edits gives 1.94% under whitespace and 0.51% under our policy, a reduction of 1.43% (Appendix D.10, Table 18).

In the separate exploratory comparison, we used engine scores to choose edits against each policy. Across the same 200 questions and four engines, the advantage was 3.27% under whitespace versus 0.68% under our policy, a reduction of 2.59%. The reductions may not hold for each engine.

Reviewer_4qzkOpenReview ↗

Rebuttal Acknowledgment

Thanks the author for the rebuttal. The new result shows that the defense's effectiveness doesn't always transfer to larger models like GPT5-mini. It would be helpful to show effectiveness of the method under more realistic settings. I hope the author can keep improving the writing.

I plan to maintain my original score.

Reviewer_kUriOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Nash-GEO, a game-theoretic framework to design platform policies that are robust to techniques like GEO. The key idea is to have the platform proactively search for publisher edits that might gain citation exposure under its current policy and update the policy accordingly.

Critical strengths
  • The problem is an important one to tackle as publishers increasingly use GEO and related techniques to influence how content is presented to humans and agents on the web.
  • The method is well motivated (Stackelberg game)
  • Evaluating defenses against an adaptive attacker is a good idea
Critical weaknesses
  • The empirical evaluation is weak. The policy evaluated is manually constructed, not game-theoretic and the models used are fairly far from the frontier, so combined with the results (max drop of 4.8% to 1.03%), it's difficult to say how the method will fare if one uses a frontier model (like fable or astra) for GEO. Also, from the current results, it looks like removing whitespace gives a major portion of the improvements, so it's difficult to say whether the game theory formulation is helping at all.
Questions for authors

I would encourage the authors to

  1. run the actual game-theoretic optimization and demonstrate improvements (since that is the central claim)
  2. use a larger space of edits
  3. if possible test on actual search engines (e.g. chatgpt) on more realistic queries (e.g. retail)

in total, I think the paper has promise and can be influential if proven right - just that the experimental results are not really backing up the ideas currently and it's unclear if they will hold in the real world.

2 comments
AuthorsOpenReview ↗

W1 + Q1. Policy selection and benefit beyond whitespace

We thank the reviewer for asking what policy selection adds beyond preprocessing and whether the results hold when stronger models write the edits or choose citations. Differences in citation exposure use the absolute percentage scale.

Does anticipating publisher responses improve policy choice? We tested this with GPT-5 mini and two simulated publishers across four synthetic tasks. The three selection methods used the same five verification policies and task information. We chose the policies on 48 questions and froze the choices before testing them on 64 new questions, with four answers per question. Both methods that held publisher behavior fixed during selection chose no verification, so they share the same 59.38% correctness result. With the method that anticipated publisher responses, 251 of 256 answers (98.05%) were correct and cited a source that the model received. Net utility is correctness as a fraction minus the expected verification cost. It was 0.918, compared with 0.594 for no verification, 0.850 for one policy selected across all tasks, and 0.800 for always verifying.

Policy choice on ASQA. The four candidates in the original setting were designed by hand: no processing, whitespace alone, the full normalizer without the screen, and the full policy. The program in Section 4.3 chooses among them. Using the two original edits, it looks for the policy that leaves the least room to gain over an unchanged source. We chose the policy on 100 of the existing ASQA questions and evaluated that choice on the other 100, keeping it fixed. The program retained the full policy in the paper. On the evaluation questions, that policy reduced the largest gain over no edit by 2.23% relative to whitespace alone. This ASQA test compares the selected policy with whitespace. The verification study above compares how the policies were chosen.

Why search against each policy? In the submission, every policy sees the same edits. Whitespace explains about two-thirds of the reduction on the original two edits, and its difference from our policy on the nine-edit set is unresolved. In a separate exploratory test, we let the search use engine scores to choose different edits for each policy. An edit is chosen using one order of the sources and scored using the other. On the same 200 questions and four answering engines:

PolicyManipulation advantage over adding support (%)
No defense7.38
Whitespace3.27
Our policy0.68

The paired reduction beyond whitespace is 2.59%. These levels differ from Table 1 because the search is different. The result shows why a fixed list can give a different picture from testing edits chosen for each policy.

The submitted paper also reports a decision informed by renewed search. A proposed label repair appeared promising against fixed edits but showed no established advantage after another search, so it was not adopted (Section 5.2 and Appendix D.4).

W1 (model scale). Larger answering models and stronger editing models

We tested larger models in both roles. In the first new test, a 32B planner wrote the edits, but the original four engines still answered. In the second, we kept the original two edits and changed the answering engine to Qwen2.5-32B. Both tests used the same 200 ASQA questions.

EvaluationManipulation advantage without defenseUnder our policy
Original four 7B-8B answering engines, pooled4.99%0.51%
32B editing planner with two selected edits and the original four answering engines8.10%0.84%
Qwen2.5-32B answering engine with the two original searched edits4.32%0.21%

These are separate comparisons, not a trend with model size. The 4.8% to 1.03% quoted in the review is the Llama panel of Figure 2. The 4.99% to 0.51% result pools the original four engines.

With GPT-5 mini as the answering engine on the same ASQA questions and two original edits, manipulation's advantage was 0.14% without defense and 1.74% under our policy. We did not find a reduction on that measure. The synthetic study above tests a different question: whether the framework can select a better verification policy with an API engine.

AuthorsOpenReview ↗

Q2 & Q3: A larger space of edits, and Live engines and retail queries

Q2. A larger space of edits

We tested 14 more edit variants on the same ASQA questions. They cover paraphrases, query terms inserted into existing sentences, and sentences composed from the source. When the search chose among these edits using engine scores, our policy reduced their advantage over adding support by 3.37% relative to no defense. This measures what those edits earn, not whether the policy deletes them.

In a separate test on 50 ASQA questions, we let the search make up to 64 proposals per document for each policy. The advantage was 8.67% without defense and 0.39% under our policy. Proposals could repeat, and neither test rules out stronger edits.

Q3. Live engines and retail queries

We agree that a live-search test on retail queries would be useful. We have not run that test. Our policy changes the retrieved text before the engine answers, a step we cannot change in ChatGPT Search. In the paper, each policy receives the same sources so that changes in citations can be attributed to the policy rather than to different search results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.