acceptodds
Under review as a conference paper at ICLR 2027

Learning Dynamics of Strategic Publishers in Generative AI Ecosystems

Abstract

Generative AI (GenAI) search systems are transforming how users access information. Given a user's question, GenAI systems generate an answer, often accompanied by attributed sources (e.g., in the form of citations). Content creators (publishers) seeking to increase exposure might behave strategically and compete with each other for users' attention. In these generative systems, the incentives take on a novel form: publishers may gain exposure through generated responses and/or attributions to these responses. We introduce a novel game-theoretic model of the emerging GenAI ecosystem in which publishers compete for attribution-based exposure. We study the learning dynamics of strategic content creators under better-response dynamics. We associate the convergence of learning dynamics to equilibrium with ecosystem stability. Using the concept of potential games, we study the stability of GenAI ecosystems under several known content selection mechanisms. We demonstrate the instability of mechanisms representing real-world modern systems and characterize a mechanism that induces a stable ecosystem. We conduct extensive simulations to analyze the stability and welfare of GenAI ecosystems under various mechanisms. The simulations support our theoretical findings and reveal an interplay among stability, publisher welfare, and user welfare. We then introduce a study illustrating how platforms can select GenAI mechanisms to achieve desired welfare trade-offs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
generative AI ecosystems, game theory, learning dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_E3cxOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how LLM-based search platforms shape competition among content publishers. It develops a game-theoretic model of publishers adapting content to gain citation exposure, analyzes the resulting dynamics under different attribution mechanisms, and uses simulations to examine stability and welfare.

Critical strengths

1 The model captures an interesting feedback loop: publishers’ content changes influence the generated response against which their exposure is evaluated. This offers a useful conceptual extension of competition in traditional search. 2 The paper examines both stability and welfare, showing that convergence need not imply better outcomes. This distinction provides useful insight into the trade-offs involved in attribution mechanism design.

Critical weaknesses

1 The practical relevance of the model is insufficiently established. The centroid-based generation function, attribution-based utility, and feedback needed for better-response learning lack adequate justification. Since the simulations retain these assumptions, they do not establish whether the findings extend beyond the proposed abstraction. 2 The interpretation and scope of the results require clarification. Convergence is a central objective, yet its practical value remains unclear when nonconvergent mechanisms can yield higher welfare. Publisher welfare also reduces to minimizing content drift, limiting its economic interpretation. Additionally, Theorem 2 needs clearer quantification: the case n=2,α=1,k=nn=2,\alpha=1,k=n admits an exact potential function. These issues weaken the resulting mechanism-design conclusions.

Questions for authors

1 Model validation: What evidence supports the centroid-based generation function as an approximation of LLM-based search? Which main conclusions are expected to hold beyond this particular function? 2 Behavioral assumptions: Which publishers plausibly maximize attribution scores, and what feedback enables them to identify utility-improving content modifications? 3 Theorem clarification: Does Theorem 2 assert that every induced game is non-potential, or that softmax does not guarantee a potential game? How does the statement accommodate n=2,α=1,k=nn=2,\alpha=1,k=n, where attribution scores are constant and an exact potential exists? 4 Practical significance: What costs of persistent content modification are omitted from the welfare measures, and why should convergence matter for mechanism selection when nonconvergent mechanisms can achieve higher welfare? 5 Welfare interpretation: Since total publisher welfare equals one minus aggregate content-drift costs, which findings provide insight beyond reduced content drift, and how should they inform claims about publisher welfare?

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

The paper studies how publishers change their content to get cited in generative search. It compares how different attribution rules affect stability and the welfare of users and publishers.

Critical strengths
  • The paper tackles an important question about publishers’ incentives.
  • The theory provides useful results, especially the convergence guarantee for linear attribution when all documents are used.
Critical weaknesses
  1. I am concerned that what is specific to GenAI. The main change is that publishers compete around a reference point shaped by their documents, rather than a fixed query. But the same model could describe traditional search with relevance feedback, where retrieved documents update the query without generating an answer.

Which findings actually depend on answer generation, rather than this general feedback mechanism?

  1. The assumptions are not checked against real systems. The experiments are all simulations. They do not test whether real systems generate answers or choose citations in the way the model assumes.

  2. The reported convergence is not fully verified. Figure 1 treats early stopping as reaching an ε\varepsilon-PNE. But the algorithm only checks a finite set of moves. Other moves could still improve a publisher’s payoff by more than ε\varepsilon.

At the final states, how do the authors rule out untested moves that could improve a publisher’s payoff by more than ε\varepsilon?

  1. Attribution outside the generation context needs more explanation. The model allows documents that were not used to generate an answer to receive attribution. It is unclear what real system setup this is meant to represent.

Why can documents outside the generation context receive attribution? Would the main results still hold if attribution were limited to documents used in generation?

Questions for authors

Please refer to the questions in the Weaknesses

Reviewer_nLU4OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a game theoretic model that describes dynamics between publishers of content and answer aggregators (eg. an AI-generated answer akin to Google’s AI overview). In the model, content publishers gain utility when the generated answer cites their content, and incur costs when they deviate from their original content. In particular, a main feature is that the publishers’ documents influence the generated answer (as opposed to being measured against a static query/optimum), and thus the better-response dynamics evolve over time, and change depending on what attribution rules the generated answer uses. The authors study three different attribution rules and characterize the stability of their respective dynamics, proving convergence for linear attribution (when all publishers contribute to generation) and constructing examples of non-convergence (eg. cycling behavior) in other settings. They also empirically validate on simulations and analyze welfare tradeoffs between publisher welfare, response relevance, and attribution relevance.

Critical strengths

In my view, the primary strength of this paper is its modeling of the generated answer as being a function of the publishers’ content. I think this is a novel contribution that captures an important property of AI-generated answers/content aggregators, and creates an interesting framework with which to study the stability of these publisher-generator/aggregator dynamics.

Another strength of the paper is its distinguishing between all publishers and a smaller retrieved subset of publishers. I felt that the exposition about the differences between these two settings was clear and well-motivated, and the theoretical results on the difference between these two settings (namely that approximate equilibria exist in some cases and not others) was also compelling.

Critical weaknesses

The model substantially builds on Madmom et. al (2025), which already introduced a framework including content embeddings, deviation costs, the softmax and linear attribution rule families, potential functions as an analysis tool, etc. It would have been helpful to have a more explicit comparison to clarify which results of the current paper introduce qualitatively new insights (as opposed to extending earlier findings).

To my understanding, the theory proves distinct characterizatioons of the regime of including all publishers and the regime of including only a selected subset of publishers. It would have been interesting to have an intermediate result (say, something to the effect of “retrieving a sufficiently large fraction x of publishers guarantees an approximate equilibrium with error y, and we characterize how y shrinks as x approaches 1”). I consider this a small weakness.

Also to my understanding, the authors did not provide a baseline comparison in their empirical simulations measuring answer relevance/welfare allowing publisher edits to a baseline of keeping publishers’ original content fixed. I felt that this would help disentangle effects - for instance, if increasing α\alpha reduced relevance by 0.03 with the fixed content baseline but 0.06 with allowing publisher edits. I also consider this a small weakness.

Questions for authors

I believe it would potentially improve readability if the figures presented in the paper had more labeling (eg. in Figure 1, it is not immediately clear what the different colored lines represent or what the y-axes represent).

Reviewer_TKQyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies a stylized model of publishers in a genAI ecosystem. Publishers both wish to be promoted and publish content near their signal. They study a variety of settings including k = n (where there are only a few choices to choose from and it can show them all). They show that if a certain selection is linear, this induces a potential game when k=n. This means that “better response” dynamics will converge. If n > k though, they show cases where the better response dynamics may not converge.

Critical strengths

The paper is fairly well written. Moreover they explore the model showing various results that are not altogether expected—like the result above.

Critical weaknesses

The biggest weakness is that the model is quite complex, and, in part, because of this, it is not clear how the results translate back to the problem purported to be studied: how strategic producers behave in genAI contexts. I think the paper provides rather limited insight that translates over with any confidence. For example, it may be no equilibrium exists if the genAI chooses the closest match. Okay, do we really believe that Alice and Bob are going to continually update their websites/papers/etc to battle each other for first place as the model predicts? Maybe, but probably not because they did not model the cost to change strategies. The agents might not play any sort of dynamics. And even if this does happen, it is already predicted by previous literature where we have race conditions. And this is one of the more robust results in the paper. I do not really buy that agents play better update dynamics in a publishing setting. Is there any evidence for this. Seems like a strange assumption. But without this, many of the results fail to hold.

For example, Theorem 3 shows that a certain class are potential games. But this only predicts that iterative updating will converge if players, in fact, iteratively update (uncertain) and have utilities that are well modeled by those in the paper (nearly certainly false). Otherwise this result does not hold and there is little we can take from it.

Theorem 4 shows that iterative dynamics will likely not converge for certain settings, but, again, this is only illuminating if agents are running iterative dynamics and not, for example, playing equilibrium.

I think that the first paragraph of related work is telling: there are many citations, but the document does not share the results of any of them. This might be because the results in this field (of Game Theory in Search and Recommendation) are that easy or interesting to start. It is likely a series of papers leading to limited insight. (To be fair the next paragraph does pull a few insights out of the papers).

Questions for authors

Do you think there are practical lessons about strategic agents in GenAI that can be gleaned from this paper? Or is this step 1, and you think some more iterations of models are required before this comes?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Rebuttal and Discussion

Dear reviewers and area chair,

We thank the reviewers for their feedback. As the rebuttal period draws to a close, we summarize the main concerns raised by the reviewers and our responses.

Contribution and relation to prior work. We present a novel model for the emerging GenAI-based search ecosystem in which publishers compete for attribution-based exposure. Unlike prior models, ours includes a novel generation function which combines retrieved documents with the platform’s own knowledge (e.g., knowledge encoded in an LLM’s weights). Publishers’ actions (modifications) can therefore change the response against which their attribution score is determined. This feedback distinguishes our setting from fixed-query models of traditional search (e.g., Madmon et al., 2025). By studying the interplay between generation and attribution, we show that stability guarantees from these models need not carry over to GenAI-based search ecosystems. Under our model, we study the impact of content selection mechanisms on the ecosystem’s stability. Moreover, we analyze the effect of GenAI mechanisms on the users’ and publishers’ welfare and reveal new Phenomena.

Modeling assumptions. In GenAI-based search, platforms often ground their answers with attributed sources (e.g., in the form of citations). These attributions allow for a new visibility source for content creators. Thus, following a wide range of prior work on generative engine optimization (GEO; Aggarwal et al., 2024), we assume that the exposure of publishers is through attributions.

Additionally, we claim that the better-response behavioral model is, in fact, a relatively weak behavioral assumption on how publishers adapt over time. To see why, notice that, compared to, e.g., no-regret or best-response adaptations, better-response captures a very wide range of possible adaptation techniques: for the given utility model of our work, we only require that the adaptation introduces some myopic improvement (among many possible ones in the typical scenario)! We do not assume optimal response as in best-response or long-term minimization of cumulative regret. This behavioral model supports many different possible trajectories, under all of which stability is guaranteed. In particular, studying this general framework allows us to abstract away the fact that interactions are made in complex modalities (typically natural language) and study the interactions in the embedding space directly.

Simulation robustness. We conduct code-based empirical simulations for content optimization competitions, as suggested by prior work (Yao et al., 2023; Madmon et al., 2025). To check the validity of our empirical simulations, we sampled 100 additional directions at each final state. We found improvements above the tolerance in only 0.004% of the cases, each of which stopped after one additional move. The welfare measures were statistically indistinguishable from those at the original final states.

If the paper is accepted, we will highlight these points in the camera-ready version. Thank you for considering these clarifications.

References:

Madmon et al. The search for stability: Learning dynamics of strategic publishers with initial documents. Journal of Artificial Intelligence Research, 2025.

Aggarwal et al. Geo: Generative engine optimization. KDD, 2024.

Yao et al. Rethinking incentives in recommender systems: Are monotone rewards always beneficial?. NeurIPS, 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.