← Paper page Mechanism Design for AI Overviews: Creator Incentives and Long-Term Profit Ask this paper OpenReview ↗

Mechanism Design for AI Overviews:
Creator Incentives and Long-Term Profit

AI summaries keep users on the search page. If the websites they are built from lose clicks and stop trying, does the search engine still win in the long run, and how can it fix the incentives?

Yihang Wu, Jiajun Tang, Jinfei Liu, Haifeng Xu, Fan Yao · NeurIPS 2026

Press → or click to step through

Who does what on a search page

creators choose how much effort (effort is costly) a search query page at rank 1page at rank 2page at rank 3 page at rank 4page at rank 5 search engine ranks pages by creators’ effort users look at top spots far more often and click good pages “position bias” AI Overview written from the pages below its quality depends on creators’ average effort clicks are creators’ income

The search engine’s profit grows with how well users are served, by the pages and by the AI Overview, minus anything it pays creators.

A click experiment: the AI Overview takes the attention

chance that a user looks at each spot on the page (estimated “position bias”) 100%50%0 AI Overview pages it cites web results, by rank

50 participants, 24 queries, 1,200 searches on a mock engine built from real Google results. Estimates from the paper’s Tables 1–2. Dashed outlines: no AI Overview.

The long-run worry: a vicious cycle

does it pay off? the AI Overview answers right on the page fewer clicks reach the creators creators put in less effort pages, and the overview built from them, get worse users are served worse; profit falls

The paper models creators competing for ranks, finds where their competition settles, and compares long-run profit with and without the AI Overview.

With identical creators the competition has a unique outcome, in which each creator mixes her effort levels, trying harder on some queries than on others (Theorem 4.1). The paper also characterizes it for a mix of experts and non-experts (Theorem 5.2).

Result 1: on its own, an AI Overview lowers long-run profit

long-run profit with an AI Overview, as a share of profit without one no AI Overview 100%75%50% more capable overview less capable overview content costly to create content cheap to create biggest loss here

Below 100% in every tested setting; worst where content is costly to create.

Right after launch, before creators adjust, the drop is smaller; in some settings profit even rises slightly. The long-run loss comes from creators’ reaction.

Traced from the paper’s Figure 1 (approximate); band = range over the AI Overview quality settings tested. Short-run comparison: Appendix I.5.

Two ways to give creators their incentive back

Citation AI Overview [1][2][3][4] pagepagepagepage cited pages win back attention; no money paid
Compensation search engine creator a bonus per unit of effort, at a rate set by rank costs money, but adds incentive directly

In the model both do the same job: they raise what a creator earns for effort at each rank. The design question is how much to give each rank.

In the click experiment, citing four pages raised total attention to web pages from 3.10 to 3.82 looks per search (Section 6.1).

Result 2: a simple shape is enough

extra reward for effort at each rank (schematic, 10 creators) rank on the results page the same boost for every rank… …but none for the last higher boost (experts’ rate) lower boost none for the last

In numerical tests, simple citation rules of this shape reach at least 60%, often over 80%, of the improvement the best citation design could bring (a conservative estimate; Appendix I.3).

Result 3: with these mechanisms, long-run profit can recover

long-run profit, as a share of profit without an AI Overview no AI Overview = 100% ≈ 57% ≈ 80% ≈ 210% content costly to create (e.g., specialized topics) ≈ 91% ≈ 99% ≈ 100% content cheap to create (e.g., lifestyle, entertainment) AI Overview alone + mechanisms, low profitability + mechanisms, high profitability

Gains are largest for costly-to-create topics where each satisfied user earns the search engine a lot. Where earnings are low, show AI Overviews there with care.

For cheap-to-create topics, payments add little: improve the overview and its citations instead.

Mechanisms beat the AI Overview alone in every tested setting, and the no-overview level in most. Bars read approximately from the paper’s Figures 1–2 (least capable overview setting); profitability = revenue per unit of user satisfaction, lowest vs highest value tested.

What this means for AI-enhanced search

  1. 1AI Overviews have a hidden long-run costThey take attention from the pages they are built from. Once creators adjust, profit falls in every setting tested.
  2. 2Citations and payments can restore the incentiveSimple rules (the same boost for every rank but the last, set per creator type) provably beat citing nothing and come close to the best design in tests.
  3. 3Tailor the policy to the queryStrong incentives pay off most for costly, specialized topics with high earnings; for everyday topics, improve the overview itself.

Paper page · OpenReview · arXiv · Ask this paper