Points, tiers and member prices work on people because people get attached and switching takes effort. An AI agent shopping for someone has no feelings to attach and almost no effort to save. So does a loyalty programme still hold a customer when an agent does the buying? We read the research published so far. The honest answer is that nobody has tested it.

What the research does say

Agents make comparing stores cheaper. In The Coasean Singularity? (NBER, November 2025), Shahidi, Rusak, Manning, Fradkin and Horton argue that agents create efficiency gains “from lower search, communication, and contracting costs”. They also flag new frictions, “congestion and price obfuscation”, and they say the net welfare effect “remains an empirical question”. Microsoft researchers make a similar case in The Agentic Economy: the big economic change is less friction between consumer agents and business agents. Habit-based loyalty works because switching costs effort, so cheaper comparison is the threat.

The agent’s preferences are not stable. In What Is Your AI Agent Buying?, Allouah, Besbes, Figueroa, Kanoria and Kumar audit agent purchase decisions. Agents often pile demand onto a few “modal” products, and “model updates can drastically reshuffle market shares”. Agents also showed strong position bias, and how much price, ratings and reviews mattered “vary sharply across models”. If a model update can reshuffle market shares, then for the machine buyer, loyalty sits with the model provider, not with your brand.

Speed can beat quality. In Microsoft’s simulated two-sided market, Magentic Marketplace, every model tested showed “severe first-proposal bias, creating 10-30x advantages for response speed over quality”. That favours the store that answers first. It says nothing about the store the customer used last time.

The closest thing to a theory, and what it is not

The only paper we found that models agent loyalty directly is DVM-HALL (Madugula, de la Rosa and Shankar, July 2026). It predicts brand choice from human emotional equity, the agent’s “machine-experience utility”, calibrated trust, delegated authority and verifiable execution. It is a framework. It ends with a “proposed” three-stage validation plan. It does not report an experiment where an agent responds, or fails to respond, to points or tiers.

That is the gap: no study we found puts a rewards programme in front of a buying agent and measures what it does.

What a merchant can do today without guessing

A loyalty programme can already be published as data that machines read. schema.org has a MemberProgram type with MemberProgramTier, linked to your organisation by hasMemberProgram. An Offer, or its price, shipping or return policy, can carry validForMemberTier for member-only terms. We checked these terms against the vocabulary in the schema.org repository on November 17. An agent cannot apply a member price it cannot read. Publishing it costs little, whatever the answer turns out to be.

How we will measure it

Starting today, every AgentReady scan records whether a store declares a loyalty programme as structured data. The record never changes the store’s score. We link it to the agent-attention series we already keep per store and per week (method in readiness and agent attention, measured). The thresholds were set before any data came in: at least 30 stores that declare a programme and 30 that do not, observed for four weeks, compared within the same score band. If a store with a declared programme gets 1.5 times the agent requests of a comparable store without one, we will report a signal. If fewer than 30 stores declare one by March 1, 2027, we will report that instead. Until then we make no claim either way.

FAQ

Do AI shopping agents care about loyalty points?

No published study we could find has tested it. Research shows agents react to price, ratings, reviews, position and response speed, and that these reactions vary by model. Points and tiers have not been tested on a buying agent.

Should I stop investing in my loyalty programme?

Nothing in the research supports that. Your human customers still respond to it. What changes is how it reaches an agent: as readable data, not as a feeling. That is why publishing it in schema.org form is the low-risk step.

Which schema.org types describe a loyalty programme?

MemberProgram and MemberProgramTier, linked from your Organization by hasMemberProgram. Member-only prices or perks hang off an Offer (or its price specification, shipping details or return policy) through validForMemberTier.

Sources

AgentReady audits whether an AI agent can find, understand and buy on your store. Scan yours to see what an agent sees today.