The agent filters; the consumer chooses.
FromRanking→ToUser-controlled fit
In a marketplace, whoever controls the ranking controls the market. An agent that simply picks “the best provider” would rebuild that gatekeeper behind a friendlier interface. The alternative is an agent that filters, explains itself faithfully, and hands the choice back to the consumer.
The ranker is the gatekeeper
Marketplaces rank by some mix of popularity, engagement, sponsored placement and price, and the ranking is both their power and their revenue. If an AI agent simply announces that it has selected the best provider, it has recreated that gatekeeping behind an AI interface. The original concept put the principle plainly: the agent filters the open supplier universe but does not make the final choice for the user[CN].
Narrow before showing, and admit that narrowing is ranking
The consumer should see neither 10,000 local businesses nor 50 algorithmically ranked products. They should see a small, relevant set of three to five providers, with who they are, what they specialise in, evidence of their work, credentials, approximate price, availability and service area.
Cutting a universe down to five is itself a ranking decision, and it is exactly where paid placement would creep in. So the shortlist rule should be published and stable, and placement in it should never be for sale. Ties between equally qualified providers should rotate, which also gives new providers a fair chance (Shift 6).
Explanations must be faithful, and the consumer sets the weights
Each provider on the shortlist should come with a reason: shown because they specialise in balcony installations, serve your area, have completed [n] similar projects and can meet your deadline. The evidence review names this kind of decision transparency as a design requirement, along with separating verified claims from unverified ones. An explanation is only honest if it reflects the real selection logic. The practical rule: explanations are generated from the same filter trace that produced the shortlist, never written afterwards as a plausible story, and the trace is logged so any shortlist can be audited.
The usual order of popularity, engagement, sponsored placement and price gives way to intent fit, capability, trust, availability, user preferences and, only then, narrative. The consumer controls the weights, with filters such as verified providers only, craftsmanship over price, available this weekend, within 10 km, established businesses or include independent specialists. When a chosen filter leaves too few providers, the agent says so and asks whether to relax it[CN]. The agent becomes an interface for the consumer’s preferences rather than an invisible commercial gatekeeper.
Go deeper
Fit before price, as a hypothesis
Once every merchant is reduced to price, rating and delivery time, the network becomes another price-comparison engine. The agent can instead treat price as one visible constraint among fit, capability, trust, availability, quality, location and service level, and let the consumer decide the weighting. That changes the merchant’s pitch from why I’m ₹50 cheaper to why I’m particularly good at this.
The nearest verifiable precedent shows the risk. Closebuy, listed on Google Play, lets a user send a text, voice or photo request. Nearby shops respond with offers, and the user compares price, distance, delivery time and ratings. Its listing showed 50+ downloads when we checked, which shows the product exists, not that it has traction[1]. The review warns that if the product collapses into nearby shops bidding on price, it becomes a hyperlocal marketplace with a conversational front end. A negotiation agent may have a place after the consumer has chosen a provider, but not as the discovery mechanism[CN].
Keeping big players from buying the top of the list
Large brands and well-funded providers will want visibility, so the safeguards need to be mechanisms, not principles. Brand authorisation should work as a filter the user can choose, never as a ranking boost. Shortlist placement should not be for sale. Ties should rotate so that incumbency doesn’t harden into a moat.
What could go wrong
- Price compression.Comparison may train users to solicit the cheapest quote and squeeze merchant margins.
- Incumbency lock-in.A shortlist that rewards history can shut out new providers. Shift 6 covers the cold-start problem.
- Explanation drift.Explanations written to persuade rather than generated from the filter trace would quietly make the agent opaque again.
- Paid placement creep.The moment the business needs revenue, placement is the easiest thing to sell. Shift 9 covers who pays.
How we’d test it
Choices collapse to the lowest price, or any systematic mismatch between an explanation and its filter trace.
| The test | What we would measure | What would end the idea |
|---|---|---|
| Fit versus price (Experiment 4) | What drives selection, by filter; realised margin; share of choices made on non-price attributes; merchant retention | Choices collapse to the lowest price |
| Explanation fidelity | Sample shortlists and compare each explanation with its logged filter trace | Any systematic mismatch between explanation and trace |
How strong is the evidence?
- Established
- Request-based local agents already let nearby shops compete on price (Closebuy; existence, not traction).
- Evidence-supported
- Decision transparency is a design requirement for agentic discovery.
- Hypothesis
- Fit-first shortlists hold up against price comparison, and faithful explanations are practical.
- Unknown
- How consumers weigh fit against price in real local missions.
Sources
- [1]Closebuy: AI Local Shopping, Google PlayChecked; the listing showed 50+ downloads
- [CN]Voice-session working notes (internal)Provided
Spotted a source we missed or a rule we misread? Tell us what we got wrong.
A shortlist only works if its entries can be believed. What should a provider have to show?