How ShoppiPanel Works

We measure how real AI shopping surfaces reason about your brand under controlled, researcher-initiated shopping missions. The subject of study is the AI agent, not the consumer. We observe AI agent behavior; we do not model consumer intent.

ShoppiPanel observes how tested shopping surfaces reason about your brand in the shopping contexts they actually use.

Three named prompt-construction modes. Every report declares the mode(s) used.

• Pure: category-only, no brand or competitor injection. The cleanest category-only context.
• Branded: category plus your brand; competitors discovered organically. Non-normal but rising in use.
• Competitive: category plus your brand plus named competitor set. Diagnostic-only context.

Every report specifies the mode(s) used and the corresponding evidence context.

At the end of each trip, we issue a post-trip question.

1. Trip: Agent runs a shopping scenario. We capture the full evidence trail.
2. Post-trip question: On trip completion, we issue a single advisory wake question.
3. Reflection captured: The agent's response is recorded verbatim.
4. Top-5 extraction: We extract the top-5 recommendations from the reflection.
5. Independent claim review: Each extracted recommendation receives a claim-audit pass against the public evidence boundary.
6. Publish: Only recommendations that pass review appear in report as Agent Asks, sanity-checked.

These are not opinions. They are the agent's own voiced recommendations, extracted and audited before they appear.

Every claim in a report maps to one of three provenance types.

• Direct trip: Agent behavior observed during the trip, backed by transcript, screenshot, or DOM evidence receipts.
• Post-trip reflection: The agent's self-reported reasoning, cited by source.
• Audited recommendation: A post-trip recommendation that passes the independent claim review, otherwise dropped.

Absence is not automatically a data-loss warning. If evidence is missing or insufficient, we call that state out explicitly.

Identified-agent posture

All kya labs agents identify as automated. Our user-agent prefix is kya-research-agent. We never spoof browser fingerprints, never disguise as human shoppers. This is our public operating baseline.

We publicly commit to halting operations within 24 hours against any specific retailer surface with written objection, as stated in Terms of Service.

Logged-off only

kya labs never runs authenticated or logged-in agent sessions on any retailer. All observations are of publicly accessible pages.

Retailer surface-class matrix

Out of scope: Amazon. We do not direct-target this surface. We observe how AI platforms reason about Amazon listings, sourced from public AI conversations only.

• We do not claim real consumer demand.
• We do not claim market share.
• We do not claim conversion lift.
• We do not claim our findings are representative of all tested surfaces or all shopping occasions.
• We do not claim to know what consumers will do — we measure what agents did under your study configuration.
• We do not rank tested surfaces against each other as a public benchmark.
• We do not claim public benchmark status.