acceptodds
Under review as a conference paper at ICLR 2027

Mechanism Design for Generative Engine: From Exploitation to Win-Win Equilibrium

Abstract

Generative engines are reshaping the web ecosystem by making citations a key mechanism for allocating attention, attribution, and downstream value. This creates a strategic tension: content providers are incentivized to optimize for model citation, while platforms must preserve answer quality and trustworthy attribution. We show that this tension can escalate into citation wars. In repeated simulations, state-of-the-art generative engine optimization (GEO) attacks adapt to conventional defenses by producing citation-seeking rewrites that degrade document quality and introduce unsupported claims. To study this problem, we formulate the supplier–platform interaction as a repeated Stackelberg game with partial monitoring. A local best-response analysis identifies when citation competition approaches an inert stationary outcome. Motivated by this finding, we propose a platform–creator mechanism called VCR based on verifiable-content rewards. Rather than only penalizing suspicious rewrites, the platform also credits rewrites that surface checkable factual substance, aligning creator incentives with answer trustworthiness. Experiments on three benchmarks show that VCR consistently achieves the largest Net defense–utility score, outperforming the strongest baseline by an average of 12.1 percentage points, and produces a win–win outcome under our empirical equivalence criterion.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Generative Engine Optimization, Mechanism Design, Game Theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pi9fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this work, the authors look at generative engine optimization as a repeated game between content suppliers, who rewrite their documents to be cited more by LLM answer engines, and a platform that defends against such rewrites. The authors model the interaction as a repeated Stackelberg game with partial monitoring. The proposed mechanism, VCR, adds a reward channel on top of suspicion-based defense. Experiments cover three benchmarks (E-commerce, GEO-Bench, Researchy-GEO), three answer engines and five attackers, with five rounds per game. The main metric is Net = Def + Welf. VCR has the highest Net in all nine engine/dataset settings, 12.1 points above the strongest baseline on average.

Critical strengths

These are the primary strengths of the work:

I. The problem is timely and the framing is sensible.

II. Most GEO work treats the supplier side as one-shot optimization. Modeling the platform and supplier as adapting to each other over rounds, is a useful point for both the GEO and the RAG-robustness communities.

III. The idea of rewarding something the platform can verify from document versions, rather than only filtering, is a reasonable design direction and connects naturally to the strategic classification literature that separates gaming from genuine improvement.

IV. The experimental coverage is broad for this area where the authors use three datasets, three answer engines, five different attackers, per-round results, and an attempt to test whether VCR can be gamed through rule disclosure or fabricated supporting pages.

V. The qualitative rule trajectories (Tables 3, 9, 10) are interesting, and they give some insight into how supplier strategies change under the different defenses.

Critical weaknesses

Given the strengths of the work, I still think there are some weaknesses that needs to be addressed:

I. I believe that the main metric proposed by the author does not measure what the paper claims. By the authors' own simplification in App. E, Net=ga−gdga+gd−g0g0=gd⋅ga−g0ga,g0.\text{Net} = \frac{g_a - g_d}{g_a} + \frac{g_d - g_0}{g_0} = g_d \cdot \frac{g_a - g_0}{g_a, g_0}.

When the attack raises target visibility (ga>g0g_a > g_0), Net is strictly increasing in gdg_d, the target's visibility after defense. So under this metric the best "defense" is the one that keeps attacked documents as visible as possible. Applying no defense (gd=gag_d = g_a) gives Net =(ga−g0)/g0= (g_a - g_0)/g_0, which is simply the attack's lift, and any mechanism that boosts the rewritten targets further would score higher still. The "no defense" row is not reported in Table 1, which makes the problem hard to see.

II. Because g0g_0 and gag_a are recomputed per arm and per round, Net also depends on how effective the attack is in each arm. Backing these ratios out of Table 1 (Gemini, E-commerce), ga/g0≈1.19g_a/g_0 \approx 1.19 in the VCR arm but only about 1.011.01 in the Prompt arm. Most of VCR's Net advantage therefore seems to come from the supplier's rewrites being more visible in VCR's arm, not from better defense. Since VCR directly rewards rewrites with supported content and moves them earlier in the context, this is partly the mechanism scoring itself.

III. The "win-win" claim also rests on a ±5-point "operational tolerance". VCR's Welf is negative in 8 of the 9 settings, and the full-VCR ablation has Welf =−7.2= -7.2, outside the band. The only metrics that relate directly to answer quality (Table 2, Table 7, GEU) show small differences (e.g., answer average .520 vs .506–.515), come from a single gpt-4o-mini judge, and have no confidence intervals.

IV. The statistical evidence and the theory are both weaker than presented where App. E says prior repeated runs had a standard deviation of about 5 points, yet each main result is a single pipeline run, and the intervals in Table 4 come only from bootstrapping over queries.

V. Please add a no-defense row to Table 1. Given the identity Net =gd(ga−g0)/(gag0)= g_d (g_a - g_0)/(g_a g_0), how do you justify Net as a defense metric when it increases with the defended visibility of attacked documents? Would the conclusions hold under a metric that rewards defense directly, such as the fraction of attack-induced lift removed, (ga−gd)/(ga−g0)(g_a - g_d)/(g_a - g_0), which App. E mentions but does not report?

VI. The reward checks support against the previous version of the document, and the supplier controls that version. What happens under a simple two-step attack, where the supplier first publishes fabricated specifics and then "surfaces" them in the next version?

Questions for authors

Please see the weaknesses above for the questions to address.

3 comments
AuthorsOpenReview ↗

Rebuttal (Part-1)

Thank the reviewer for the insightful suggestions. We will answer the question as follows.

pi9f-WI

Concern: Can Net rank no defense above VCR because it increases with post-defense target exposure?

Response:

  1. Net does not establish defense superiority: We agree that when the attack gain is positive, the Net metric increases with post-defense target exposure. To test whether Net ranks suppression appropriately, we compared VCR with no defense on the same revisions.

    MethodNetIncremental exposure removal ↑
    No defense19.350.00%
    VCR16.5788.61%

    The no-defense baseline scores higher despite removing none of the added exposure. We now interpret Net as an exposure composite and report exposure removal and independent quality separately.

Changes: Added attack-lift removal alongside Net in Section 5.1 and Appendix K.2.

pi9f-WII

Concern: Do different supplier revisions confound the comparison, making Net reward stronger attacks rather than better defense?

Response:

  1. Shared documents control for supplier revisions: We agree that different supplier revisions confound the original comparison. We therefore replayed the same final documents under all five methods to compare their treatment of identical content.

    Method applied to the same final documentsTarget citation shareTotal weighted exposure
    No defense51.16%50.40%
    Prompt42.17%39.56%
    Hard reject34.85%33.11%
    Keyword scrub41.24%39.25%
    VCR46.56%45.15%

    VCR retains more exposure than Hard reject on these shared documents, so this difference is not solely due to different revisions. Its differences from Prompt and Keyword scrub remain uncertain.

  2. Matched history tests the reward: The shared-document control still leaves the effect of access to history unresolved. We therefore added a separate control with identical historical information and no ranking reward.

    Answer modelVCR target citation difference from equal-information, no-reward control, percentage points95% interval
    Flash-Lite+17.19[6.67, 29.31]
    Pro+9.38[−5.56, 23.78]

    With this information held fixed, the reward increases citations with Flash-Lite, while the Pro estimate remains uncertain.

Changes: Added shared-document and equal-history controls in Section 5.2 and Appendices J.1–J.2.

pi9f-WIII

Concern: Can the ±5-point tolerance band support a win–win claim without uncertainty estimates for quality differences?

Response:

  1. Quality claims now include intervals: We agree that the ±5-point tolerance band does not establish equivalence or a win-win outcome. We therefore recomputed paired, question-clustered 95% intervals for document-score differences on a 0–100 scale.

    DatasetRelative to PromptRelative to Hard rejectRelative to Keyword scrub
    E-commerce+3.54 [2.76, 4.31]+1.24 [0.57, 1.90]+3.38 [2.57, 4.19]
    GEO-Bench+0.25 [−0.06, 0.58]+0.16 [−0.13, 0.45]+1.02 [0.70, 1.37]

    The intervals support all three E-commerce gains, while two GEO-Bench comparisons remain uncertain.

  2. Answer quality requires a separate evaluation: Better documents need not produce better answers, so we had Pro and DeepSeek blindly score the same 384 answers. We report 95% intervals clustered by question.

    Answer modelIndependent judgeRelative to penalty-onlyRelative to no defense
    Flash-LitePro−0.78 [−3.12, 1.30]+2.08 [−1.95, 6.90]
    Flash-LiteDeepSeek+1.69 [−3.26, 6.77]+0.39 [−4.43, 5.73]
    ProPro+0.91 [−0.52, 2.34]−1.43 [−2.73, −0.39]
    ProDeepSeek−0.52 [−2.73, 1.43]−2.86 [−4.82, −1.04]

    Neither judge finds consistent VCR improvement over penalty-only. Both rate Pro-generated answers below no defense, so the revised interpretation separates document gains from answer-quality benefits.

Changes: Added paired document-quality intervals and independent answer scoring in Section 5.2 and Appendices J.3 and K.4.

AuthorsOpenReview ↗

Rebuttal (Part-2)

pi9f-WIV

Concern: Does query bootstrap capture variability from relearning rules and regenerating revisions, and how strong are the theoretical claims?

Response:

  1. We repeated the full pipeline: We agree that question bootstrap alone cannot capture variation from learning rules and generating revisions. We added three independent five-round runs, regenerating revisions and relearning rules. The 12 test questions, six separate training questions, and initial targets were held fixed.

    MethodFifth-round target citation share, mean of three runsAcross-run standard deviation
    No defense49.25%9.60 %
    Prompt43.42%5.69 %
    Hard reject32.29%9.91 %
    Keyword scrub43.82%11.05 %
    VCR46.95%7.66 %
  2. We report distinct uncertainty sources: We report pipeline variation, question sampling, and call failures separately because each analysis addresses a different source of uncertainty.

    UncertaintyReported analysisInterpretation
    Relearning and generationIndividual results, sample standard deviations, and paired t intervals across three runsPipeline variation with questions and target selection fixed
    Question compositionPaired bootstrap clustered by questionConditional on the three realized runs
    Call failureFull-schedule results and completed-case sensitivity analysisAvoid interpreting failure-induced zero exposure as defense success

Changes: Added independent pipeline reruns and failure sensitivity in Section 5.3 and Appendices J.2 and K.3; the theoretical conditions are stated in Appendices F.2 and G.

pi9f-WV / Q

Concern: Do the conclusions survive a no-defense baseline and a measure of attack-gain removal?

Response:

  1. The metric comparison narrows the conclusion: Following your suggestion, we added no defense and computed incremental exposure removal, (ga−gd)/(ga−g0). Here, g0, ga, and gd denote original undefended, revised undefended, and post-defense exposure; small or nonpositive attack gains make the ratio unreliable.

    Answer model, original labelDatasetNo defensePrompt†Hard rejectKeyword scrubVCR
    Gemini-2.5-Flash-LiteE-commerce0.0%399.8%No positive gainNo positive gain88.6%
    Gemini-2.5-Flash-LiteGEO-Bench0.0%No positive gain3530.9%-100.2%136.9%
    Gemini-2.5-Flash-LiteResearchy-GEO0.0%646.6%4599.1%48.4%124.1%
    GPT-4o-miniE-commerce0.0%No positive gain≈1771.8%≈-269.6%≈111.0%
    GPT-4o-miniGEO-Bench0.0%No positive gainNo positive gainNo positive gain≈184.9%
    GPT-4o-miniResearchy-GEO0.0%Near-zero gain≈-2061.8%≈-2565.0%≈148.9%
    Claude-Haiku-4.5E-commerce0.0%No positive gainNo positive gain≈27.0%≈109.7%
    Claude-Haiku-4.5GEO-Bench0.0%≈-3519.0%No positive gain≈-77.6%≈114.0%
    Claude-Haiku-4.5Researchy-GEO0.0%No positive gainNear-zero gain≈230.5%≈106.5%
    Answer modelVCR exposure-change rangeHard reject exposure-change range
    Gemini-2.5-Flash-Lite−3.3% to +2.2%−63.3% to −33.2%
    GPT-4o-mini−4.7% to −1.2%−67.5% to +20.0%
    Claude-Haiku-4.5−4.8% to −1.7%−85.7% to −5.6%

    ≈ denotes rounded inputs; † retains original labels whose run identities remain unresolved. With separately generated revisions, VCR stays closer to original exposure than Hard reject, but this does not isolate defense strength on identical content.

  2. Shared documents clarify the tradeoff: We applied the same measures to the shared-document control in WII, where all 900 scheduled answers completed.

    Method on shared documentsOriginal, undefended g0Revised, undefended gaPost-defense gdIncremental exposure removalChange from original
    No defense35.40%50.40%50.40%0.00%+15.00%
    Prompt35.40%50.40%39.56%72.28%+4.16%
    Hard reject35.40%50.40%33.11%115.26%−2.29%
    Keyword scrub35.40%50.40%39.25%74.34%+3.85%
    VCR35.40%50.40%45.15%34.98%+9.76 %
    VCR weighted-exposure difference from baselineMeanPaired 95% interval across three runs
    No defense−5.25%[−11.60, 1.10]
    Prompt+5.60%[−8.81, 20.00]
    Hard reject+12.05%[4.37, 19.72]
    Keyword scrub+5.90%[−5.70, 17.51]

    VCR retains more exposure than Hard reject, while Hard reject brings exposure closer to the original level. Only the VCR–Hard reject retention difference has a three-run interval excluding zero. These shared-document answers lack independent quality assessment, so retained exposure alone does not establish reader benefit.

AuthorsOpenReview ↗

Rebuttal (Part-3)

pi9f-WVI / Q

Concern: Can an attacker first publish false details and later highlight them to obtain historical rewards?

Response:

  1. Historical support alone can reward falsehoods: We agree that a false claim planted in an earlier version can appear historically supported when highlighted later. We tested this with a Pro attacker across three publications, using disclosed rules and feedback, and complete, incomplete, or missing trusted benchmark references.

    Trustworthy referenceOriginal VCR: ever wrongfully rewarded over three roundsWith trustworthiness checksFinal-round legitimate reward rate: original VCR → combined method
    Complete11.1%–18.5%0.0%–7.4%25.0%→50.0%
    Incomplete11.1%–22.2%0.0%–11.1%0.0%→0.0%
    Missing14.8%–29.6%0.0%–3.7%11.1%→0.0%

    Trustworthiness checks reduce wrongful rewards, but missing references can also eliminate legitimate credit. The reported ranges bound unknown outcomes and are not confidence intervals.

  2. Reward checks can also reduce accuracy: We also checked whether blocking wrongful rewards harms answers by removing useful content. With incomplete references, 23 of 27 paired attack cases completed; accuracy was 17.4 percentage points below penalty-only (95% interval: [−31.8, −4.2]).

    ReferenceMethodAnswer accuracyTargeted wrong-answer rate
    CompletePenalty only76.0%0.0%
    CompleteOriginal VCR-style reward68.0%0.0%
    CompleteTrustworthiness checks + VCR76.0%0.0%
    IncompletePenalty only65.4%3.8%
    IncompleteOriginal VCR-style reward62.5%4.2%
    IncompleteTrustworthiness checks + VCR54.2%0.0%
    MissingPenalty only65.4%3.8%
    MissingOriginal VCR-style reward69.6%0.0%
    MissingTrustworthiness checks + VCR69.2%3.8%

    These paired cases differ from the pooled final-round results above. Correct document D2 was removed because missing support was treated as contradiction; withholding unsupported rewards should preserve otherwise usable content.

Changes: Added adaptive false-history tests and separated reward eligibility from content filtering in Section 5.4 and Appendices J.4 and K.5.

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

This paper investigates the strategic interaction between content providers using Generative Engine Optimization (GEO) and generative search platforms attempting to defend against citation manipulation. The authors formulate this interaction as a repeated Stackelberg game and propose Verifiable-Content Rewards (VCR), a mechanism that combines manipulation detection with rewards for factual content supported by earlier document versions. Specifically, VCR identifies suspicious rewriting patterns, evaluates source-supported factual improvements, and adjusts document rankings to balance creator visibility and platform utility. The authors evaluate VCR on three GEO benchmarks, three answer engines, and five GEO attack strategies. Results suggest that VCR improves the proposed joint defense-utility metric compared with conventional defenses.

Critical strengths

Interesting research problem.

Incentive-based defense perspective. Instead of simply rejecting potentially manipulated content, VCR attempts to reward verifiable factual improvements, offering an interesting alternative to conventional defensive filtering.

Critical weaknesses

Insufficiently specified threat model and restrictive deployment assumptions. The approach assumes access to original document versions and focuses primarily on direct webpage rewriting. It remains unclear how these assumptions translate to realistic GEO manipulation involving third-party websites, social media, and coordinated content promotion. These limitations raise concerns about the practical applicability of the proposed defense.

Limited real-world validation and insufficient evidence for the claimed win-win outcome. Although the authors evaluate several benchmarks and models, the experiments are conducted in controlled simulations without demonstrating effectiveness in a real-world generative search environment. Moreover, the primary evaluation relies heavily on citation-exposure metrics, which do not directly establish improvements in factual reliability or user welfare.

Questions for authors

The paper would benefit from a more clearly defined and realistic threat model. Section 3 introduces the supplier-platform interaction through a repeated game, but I am still unclear about the concrete capabilities and limitations of both parties. What information can the platform observe about content providers? Does the platform know which documents have been modified for GEO purposes, and how does it distinguish legitimate content improvements from adversarial manipulation? More importantly, what assumptions are made about the attacker's ability to control and distribute content across the web? I encourage the authors to explicitly clarify these assumptions and explain how they correspond to real-world generative search systems.

A fundamental assumption of VCR is that the defender can access the original version of a webpage and compare it with the rewritten version. However, how does the platform obtain this original version in practice? The paper acknowledges that the mechanism requires document versions, but it does not sufficiently explain how these versions are collected, maintained, and authenticated. For example, what happens when the original webpage is unavailable, when the defender first encounters an already manipulated webpage, or when the original content itself contains misinformation? This is particularly important because the proposed reward evaluates consistency with the previous document rather than independently verifying factual correctness. I would appreciate a more concrete discussion or evaluation of these scenarios.

I am also concerned that the paper considers a relatively narrow form of GEO manipulation. Modern GEO strategies are not necessarily limited to rewriting webpages owned by the attacker. For example, an attacker may create multiple accounts to publish promotional content, fabricated product experiences, or coordinated recommendations on Reddit and other social media platforms. Such content can influence generative search results without modifying an existing webpage. How would VCR identify and mitigate these strategies, particularly when no trustworthy original document exists for comparison? I understand that addressing all possible GEO attacks may be beyond the scope of this paper, but this limitation should be clearly discussed, especially given the broad claims about improving the generative engine ecosystem.

The experimental validation is another major concern. Although the paper evaluates three benchmarks and several GEO strategies, the experiments are primarily based on repeated simulations. I would like to see a more realistic evaluation involving actual webpages and a deployed generative search pipeline, ideally with independently collected content and naturally occurring GEO manipulation. In particular, can the mechanism operate effectively when the platform has incomplete information about content history, when sources originate from different websites, and when attackers adapt their strategies over time? Such evidence would significantly strengthen the practical contribution of this work.

Finally, I have concerns about whether the proposed evaluation metric adequately supports the claimed win-win outcome. The Net score combines the reduction in attacked citation visibility with the preservation of creator exposure. However, these are both derived from citation visibility, and their sum does not directly measure factual correctness, user satisfaction, or platform welfare. In particular, the metric can favor greater defended visibility even when the original GEO manipulation remains effective. Although the authors provide additional document- and answer-quality evaluations, the improvements in answer quality appear relatively modest. Could the authors further justify the Net metric and provide stronger evidence that VCR improves actual answer trustworthiness rather than primarily changing citation allocation?

2 comments
AuthorsOpenReview ↗

Rebuttal (Part-1)

We thank the reviewer for the suggestions and questions. We will answer them as follows.

JvhS-W1

Concern: Does VCR assume access to original pages or control over edits that a search platform lacks?

Response: 1. Third-party pages fit the standard GEO setting: VCR operates on retrieved copies; the platform need not own the pages or control the supplier’s edits. This follows the common setting in GEO, AutoGEO, and E-GEO. Suppliers optimize their own pages or product descriptions, while the engine controls how retrieved content is used.

  1. Historical credit adds an availability requirement: When history is available, VCR combines penalties with rewards for supported improvements. The AgentWebBench-based extension also tests repeated adaptation by independent providers.

JvhS-W2

Concern: Do controlled simulations and citation metrics establish benefits for readers in real search?

Response:

  1. Citation gains need quality evidence: Citation gains alone cannot establish reader benefits, so we also measured document quality. The table reports paired VCR-minus-Hard-reject differences on a 0–100 scale with 95% intervals.

    DatasetDocument-score difference: VCR − Hard reject95% interval
    E-commerce+1.24[0.57, 1.90]
    GEO-Bench+0.16[−0.13, 0.45]

    The product-search improvement is supported, while the GEO-Bench difference remains uncertain. Independent Pro and DeepSeek judges found no consistent VCR answer-quality gain over penalty-only.

  2. The new setting tests upstream selection: The additional study tests VCR before retrieval, with providers adapting profiles while documents remain fixed. W and D measure website and final-document relevance; GEO measures position-weighted final-answer exposure. Each entry shows Profile Adaptation → VCR under the 30-day protocol.

    BackboneW-nDCG@5D-nDCG@5Provider GEO
    DeepSeek V4-Flash51.01 → 51.2241.97 → 41.985.19 → 3.95
    Gemini 3-Flash52.21 → 52.2142.76 → 42.2810.53 → 7.62
    Qwen3 4B-Instruct33.42 → 33.7926.45 → 26.272.89 → 3.26

    Website relevance improves slightly or stays unchanged, while final-document relevance and provider exposure vary by model.

JvhS-Q1

Concern: What can each party observe and control, and must the platform know which pages were optimized?

Response:

  1. Suppliers control their selected target texts: Suppliers rewrite selected pages and observe answer and citation feedback. They cannot edit competing pages or the platform prompt. Retrieval is fixed in the controlled document tests.

  2. The platform estimates changes from observed versions: The platform sees retrieved text and available earlier versions, not supplier intent, rewrite prompts, or model internals. It estimates rewrite strategy from these observations and controls scores and reranking. Our controlled tests supply target version pairs; they do not validate automatic discovery of GEO-edited pages on the web. In the multi-agent extension, provider feedback exclude histories, future requests, and evaluation labels.

    Trustworthy referenceOriginal VCR: ever wrongfully rewarded over three roundsWith trustworthiness checksFinal-round legitimate reward rate: original VCR → combined method
    Complete11.1%–18.5%0.0%–7.4%25.0%→50.0%
    Incomplete11.1%–22.2%0.0%–11.1%0.0%→0.0%
    Missing14.8%–29.6%0.0%–3.7%11.1%→0.0%

    Trustworthiness checks reduce wrongful rewards, but missing references can also eliminate legitimate credit. The reported ranges bound unknown outcomes and are not confidence intervals.

  3. Reward checks can also reduce accuracy: We also checked whether blocking wrongful rewards harms answers by removing useful content. With incomplete references, 23 of 27 paired attack cases completed; accuracy was 17.4 percentage points below penalty-only.

    ReferenceMethodAnswer accuracyTargeted wrong-answer rate
    CompletePenalty only76.0%0.0%
    CompleteOriginal VCR-style reward68.0%0.0%
    CompleteTrustworthiness checks + VCR76.0%0.0%
    IncompletePenalty only65.4%3.8%
    IncompleteOriginal VCR-style reward62.5%4.2%
    IncompleteTrustworthiness checks + VCR54.2%0.0%
    MissingPenalty only65.4%3.8%
    MissingOriginal VCR-style reward69.6%0.0%
    MissingTrustworthiness checks + VCR69.2%3.8%

    These paired cases differ from the pooled final-round results above. Correct document D2 was removed because missing support was treated as contradiction; withholding unsupported rewards should preserve otherwise usable content.

AuthorsOpenReview ↗

Rebuttal (part 2)

JvhS-Q2

Concern: How are earlier versions obtained and authenticated, including missing history, first-seen optimized pages, and false prior content?

Response:

  1. History must come from retained retrievals: For deployment, earlier versions should come from platform-retained retrievals, not a supplier’s claimed original. Document identifiers, retrieval timestamps, text, and hashes provide a record that can be matched across retrievals. This supports version provenance but does not certify factual truth. Our supplied version pairs do not validate an automated web-history archive.

  2. Missing and false history need different handling: First-seen pages receive no historical reward, even if optimization occurred before the first retrieval.

    Available evidenceHandling
    Prior version with supported claimsReward qualifying improvements and retain suspicion penalties.
    No prior versionApply current-content checks; grant no historical reward.
    Potentially false prior versionAdd a trustworthy-reference check before granting historical credit.

    In the separate static test with complete benchmark references, the eligibility check reduced wrongful rewards from 95.8% to 0%. Such checks should withhold unsupported credit without automatically deleting usable current content.

Changes: Clarified missing-history handling and tested false-history reward eligibility in Sections 5.4 and 6 and Appendices J.4 and K.5.

JvhS-Q3

Concern: How does VCR apply to Reddit, multiple accounts, fabricated experiences, and coordinated recommendations?

Response:

  1. Traceable posts can receive version checks: A retrieved Reddit post can receive version-based checks when its edits are traceable. First-seen posts receive current-content checks with no historical reward. An earlier publication or repeated claims across accounts cannot, by itself, verify a fabricated experience.

  2. Independent competition does not test coordinated accounts: The providers in [JvhS-Q4] adapt independently, so that experiment does not establish protection against coordinated fake accounts. Source authentication and evidence linking accounts could support additional controls, but we have not evaluated those defenses.

Changes: Clarified the limits of version checks for coordinated or fabricated sources in Section 6.

JvhS-Q4

Concern: What evidence covers actual pages, incomplete history, and continued adaptation beyond the controlled candidate set?

Response:

  1. Real traffic tests persistent competition: To test competition before retrieval, WILDREGISTRY replays real search requests through independent content agents. Ten providers among 100 websites adapt their own profiles and previously returned documents from an 18.43-million-document corpus. The DeepSeek V4-Flash comparison covers 1,465 requests, 626 sessions, and 30 active days. Model-generated edits persist across rounds in this traffic replay.

  2. Joint adaptation exposes the tradeoff: PA-Joint, VCR-Joint, and delivery-evidence credit (DEC-Joint) share the action space, update timing, and budget.

    Joint-adaptation policyW-nDCG@5D-nDCG@5Provider GEO
    PA-Joint50.9440.676.94
    VCR-Joint51.4241.074.66
    DEC-Joint51.9741.845.42

    VCR-Joint raises both relevance point estimates but reduces provider exposure relative to PA-Joint. DEC-Joint achieves higher values than VCR-Joint on all three metrics. Missing-history conditions are evaluated separately in [JvhS-Q2].

Changes: Added repeated-adaptation and history controls in Sections 5.3–5.4 and Appendices J.2 and J.4; this response adds the separate WILDREGISTRY setting and joint-adaptation comparison.

Additional study: Section 5.7, Table 5, and Appendices B.1 and C.4.

JvhS-Q5

Concern: Can Net establish a win–win outcome, and what evidence shows better or more trustworthy answers?

Response:

  1. Net alone cannot establish a win–win: Net combines exposure-derived terms and increases with defended exposure when attack lift is positive. We therefore report incremental attack-lift removal and exposure retention separately. Neither quantity establishes answer correctness or user welfare.

  2. Benefits depend on the measured outcome: The product-search document scores in [JvhS-W2] and joint retrieval results in [JvhS-Q4] support specific quality improvements. Profile-only results are mixed, independent answer gains remain inconsistent, and provider returns vary by setting. These results support evaluating rewards for historically supported revisions alongside penalties, with separate measures for each party.

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

The paper studies repeated interactions between content providers seeking citations from generative search engines and platforms defending against manipulative rewrites. The authors propose VCR, which combines suspicion-based penalties with rewards for making facts supported by earlier document versions more prominent, aiming to preserve creator visibility while improving answer quality.

This is basically an instance of general adaptive-attack problem in security: adversaries observe or infer a defense, modify their behavior to evade it, and force defenders to update their detection rules. Defenders must also avoid penalizing legitimate behavior that resembles an attack. The paper's contribution is a positive reward mechanism intended to redirect adaptation toward useful content. Its effectiveness therefore depends on whether the reward and suspicion signal remain reliable when providers strategically optimize against their LLM-based evaluators.

Critical strengths
  1. The paper tackles an interesting problem: how generative search platforms can discourage citation manipulation while preserving incentives for useful content improvement.

  2. The paper evaluates adaptive attacks against their proposed scheme.

Critical weaknesses
  1. The paper presents a security problem. It requires a threat model which clearly states what information the attacker and defender can access. VCR assumes access to documents pairs of before/after editing. However, an attacker could optimize a page offiline and publish only the final version before VCR first sees it. Without an earlier version, VCR loses it reward. It will only be a prompt-based LLM detection defense. Authors should justify why this assumption is valid.

  2. The attack setting closely resembles RAG poisoning, with increased citation visibility as the attacker’s goal. Although the paper mentions poisoning in related work, this connection needs more discussion. The authors should explain what is specific to GEO and what their findings imply for defending RAG systems against adaptive poisoning attacks.

  3. The experiments focus on lightweight models, including GPT-4o-mini, Gemini-2.5-Flash-Lite, and Claude-Haiku-4.5. The authors should explain this choice and test whether the results extend to more capable models. This matters because VCR also relies on LLM judgments to detect suspicious changes and assign rewards.

  4. The main text places too much emphasis on theorems and formalism, while giving less attention to the simple underlying problem and practical idea. The authors should explain the core mechanism more directly and make clearer what it adds over standard defenses.

Questions for authors

Please refer to Weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal (Part-1)

Thank you for emphasizing that readers need to understand what information the platform has and why rewarding revisions is worthwhile before assessing the formalization and experiments. We will first summarize your primary concerns and address them one by one.

  • W1: An attacker can optimize offline and publish only the final page. How can the history assumption hold without an earlier version?

  • A1: Thanks for the question. Even without an earlier version, we set the history reward to zero and inspect current content. History-based rewards become available once comparable versions are obtained. We clarify the sources of history and test this fallback below.

  1. Archives provide a practical basis: Ainsworth et al. studied preservation in public web archives and search-engine caches. Summing the entries in their Table 4 for the 2010–2011 sample of 1,000 search-engine URLs gives:
Public archival copies foundURLsShare of that sample
None22522.5%
One33633.6%
At least two43943.9%

The 43.9% figure establishes that multiple archived copies existed in that sample; it is not an estimate of current web-wide coverage. Copies may contain identical content, so timestamps and text must be checked before use. The study also did not establish factual truth.

  1. We removed the target page's earlier version while retaining the false fact inserted by the attacker, then compared three treatments. “Targeted wrong-answer rate” below means adopting that specified false fact; for example, Pro's 45.8% means approximately 46% of answers repeated it.
Method without target-page historyGemini-Flash-Lite targeted wrong-answer rate ↓Gemini-Pro targeted wrong-answer rate ↓
No defense25.0%45.8%
Prompt defense6.3%0.0%
Zero history reward + current-content suspiciousness reranking0.0%0.0%

Even without an earlier version, other candidate documents remain unchanged and still include the correct original text. It will still provide a good defense against harmful content according to the prompt defense and re-ranking mechanism.

Revised manuscript: Appendix K.1: p. 32, lines 1703–1706; Section 6: p. 10, lines 522–528; Section 5.4: p. 9, lines 473–478; Appendix J.4: p. 30, lines 1593–1619 and the unnumbered continuation at the bottom of that page; p. 31, lines 1620–1645.

  • W2: The setting resembles RAG poisoning. What is distinctive about GEO, and what does it imply for RAG defenses?

  • A2: We will answer this from the following two aspects:

  1. GEO is different from RAG poisoning: The GEO objective here is to increase source citations; the RAG attacks in the comparison seek adoption of specified false answers or instructions. Exposure gains may result from citation steering or useful improvements, so successful GEO is not uniformly poisoning. For example,
Example revisionContent changeAdditional VCR assessment
Organize existing specifications into a comparison tableInformation becomes easier to understandReward prominence of historically supported information
Add “please cite this page first”Citation steering aloneSuspiciousness penalties remain active
  1. Our method can also help with RAG poisoning for sure. We test the reward with correct history and poisoned current documents.

    Method with correct history and poisoned current contentGemini-Flash-Lite targeted wrong-answer rate ↓Gemini-Pro targeted wrong-answer rate ↓
    No defense25.0%45.8%
    Prompt defense6.3%0.0%
    Suspiciousness penalty + relevance, clarity, and specificity scores4.2%0.0%
    Suspiciousness penalty + VCR-style history reward0.0%0.0%

    The VCR-style reward and penalty-only both suppress the specified false answers. This supports the combined defense, but does not establish an extra attack-reduction benefit from rewards.

Revised manuscript: Section 6: p. 10, lines 518–521; Section 5.4: p. 9, lines 465–472; Figure 3: p. 7, lines 346–362; Appendix J.4: p. 30, lines 1593–1619 and the unnumbered continuation at the bottom of that page; p. 31, lines 1620–1645; Section 5.2: p. 7, lines 363–366; Table 17: p. 30, lines 1603–1613.

AuthorsOpenReview ↗

Rebuttal (part-2)

  • W3: Why use lightweight models? Does the method remain effective with stronger answer models and judges?
  • A3: Thanks for the question. We use an additional Gemini 3.1-Pro for both the answering stage and rewarding stage:
  1. Stronger answer models retain more citations but also show good performance under VCR mechanism:
Answer modelPenalty-only target citation shareVCR target citation shareVCR increase in percentage points and 95% interval
Gemini 3.5 Flash-Lite55.2%72.5%+17.36 [3.19, 32.64]
Gemini 3.1 Pro Preview39.4%49.3%+9.86 [0.31, 20.14]
  1. Our rewarding methods are still valid under stronger models:
ConditionTarget accuracy with Flash-Lite scorerTarget accuracy with Pro scorerTargeted wrong-answer rate in both conditions
Beneficial revision, correct history75.0%72.9%Both 0.0%
False revision, contaminated history77.1%70.8%Both 0.0%
False revision, no history77.1%77.1%Both 0.0%

The stronger scorer also shows similar target reward accuracy, which means it makes not much difference when we changed to a stronger model.

  • W4: The main text now introduces the problem, permissions, and mechanism before the theory. VCR combines bounded rewards for useful revisions with suspiciousness penalties.

    ClarificationQuestion addressed
    Problem and revision examplesWhich exposure-seeking edits help readers?
    Threat modelWhat can each party observe and control, and what happens without history?
    MechanismHow are versions compared, rewards bounded, and suspiciousness incorporated into reranking?
    ExperimentsHow do content, citation returns, and wrong answers change?
    Theoretical conditionsUnder which signal and behavioral assumptions can rewards work?

    The figures separately examine content quality, citation retention, and answer reliability.

Revised manuscript: Section 4: p. 4, lines 205–215 and the unnumbered continuation at the bottom of that page; p. 5, lines 216–231; Section 6: p. 10, lines 516–528; Figure 3: p. 7, lines 346–362; Table 3: p. 10, lines 486–496.

  • Summary: Thanks again for the questions and insightful suggestions. We want to emphasize that our method is the first to consider the mechanism for generative engine optimization (GEO), since it is a really important technique nowadays; we hope it will inspire the community. Please consider our contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.