NBER Working Paper 35953, "Competition in Generative AI: Evidence from Consumer Demand" (Egan, Russel, Shi, Yao; October 2026), assesses competition in the generative AI market using credit and debit card transaction data: 174 million consumers and 114 billion transactions. By April 2026, approximately 2.9% of US consumers pay for at least one major AI service, almost all on basic plans under $30 per month. Annual customer retention among major providers is 20–45%, and spikes in switching between services coincide with new model releases. The authors note high price sensitivity and weak preferences for specific providers: without a significant difference in model quality, growing substitutability will limit companies' pricing power, and surplus will shift to users.

image

What happened

A research team consisting of Egan, Russel, Shi, and Yao published NBER Working Paper 35953, "Competition in Generative AI: Evidence from Consumer Demand," in October 2026. The study is based not on surveys, but on credit and debit card transaction data: the sample covers 174 million consumers and 114 billion transactions. It shows that as of April 2026, approximately 2.9% of US consumers are subscribers to at least one major generative AI service, and almost all of them choose basic tariffs costing up to $30 per month. Annual customer retention among major providers is 20–45%, and switching spikes are synchronized with new model releases: consumer choice is driven by the release cycle, not the brand. The authors then built a structural supply and demand model that identifies high price sensitivity and weak preferences for specific services. The model's conclusion is that if the difference in model quality is imperceptible to the mass user, the growing substitutability of services will limit providers' pricing power.

Context

The generative AI subscription market—services like ChatGPT and Gemini—has previously been evaluated mainly through corporate disclosures and surveys, making the scale of the study's empirical base unusual: panel card data of this volume is rarely available to researchers. Annual retention of 20–45% is equivalent to 55–80% of subscribers not renewing their subscription within a year, which stands out sharply against the metrics of long-term loyalty typical for mature subscription businesses. According to the data, the mass paying user does not see the differences in model quality that labs consider key competitive advantages: they vote with their wallets around releases, not around leaderboard shifts. The work also brings to the forefront a central industry question about the gap between benchmark metrics and perceived quality, which requires a methodology for evaluating models based on behavioral, not just test, criteria.

Why this matters for the industry

The paper provides the industry with quantitative benchmarks that were previously unavailable: the share of the paying audience, annual retention, and the actual price range. The practical implication is that financial models for AI subscription products should logically be recalculated for 55–80% annual churn, and one should not expect consumer plan prices to rise significantly above the $30 mark: pricing is sensitive, and the "same access, but better" differentiation format does not sell. Model releases serve as a trigger for attracting competitors' customers, so it is likely that release cycles will be strengthened as a tool for churn competition, along with pricing experiments: tiers, bundles, and promo-switching. The signal for engineering teams is to not build the critical path on a single provider, to introduce a provider-agnostic client abstraction with fallbacks, and to fix an eval set and a protocol for running it before each major model release. If quality parity is maintained, the consumer segment will follow a path of commoditization with thin margins and possible consolidation, while differentiation will shift from raw model quality to the ecosystem—agents, integrations, memory—and to the corporate and API channels; an increase in multi-model configurations in production and demand for model selection and routing tools is also expected.

Why this matters for users

The main takeaway for readers: an AI service subscription is not a "forever" solution, but an easily switchable product. Changing providers with every major release is a rational strategy that allows users to avoid overpaying, because mass user choice relies on service competition rather than brand loyalty. If model quality continues to equalize, economic surplus will shift to consumers, and competition will determine subscription prices. Pricing experiments—tiers, bundles, promo actions for switching—are likely, while providers will respond to churn not only with price but also with memory, integrations, and agentic capabilities, i.e., value around the model. Demand for routers and "smart" subscriptions that automatically track releases and switch configurations may increase. A useful habit to adopt now: plan for model changes without losing context and review your tariff at the moment of each notable release.

What is still unknown / limitations

The study's limitations are significant. Card data records the fact of payment but not the perception of quality or the intensity of service usage. The thesis of substitutability and weak preferences is derived from a structural model based on assumptions, so market commoditization is a model forecast, not an established observation. The paper has working paper status, and conclusions may be revised: a second layer of econometric studies is expected, which will either confirm the results on other samples or attack them in replication. The sample covers US consumers, and the data extends to April 2026, so transferring conclusions to other markets and later periods requires verification. There is no data on the corporate and API channels, as well as on latency, uptime, and inference tariffs in the work, and the conclusions do not directly apply to these segments.

Sources

Author

Look at AI, editorial team