Seller analytics

​

Overview

When you ask Murph for seller analytics, Murph can attach a sellerAnalytics payload to the chat response for visual rendering. The payload summarizes your recent intelligence runs — discovery-run mix, buyer asks, top surfaced products, and the commercial outcomes attributed to those runs — and turns them into directional negotiation guidance. This page is the canonical reference for that payload. Each section below is one top-level field of sellerAnalytics.

​

historicalPerformance

historicalPerformance summarizes the analytics window into the negotiation signals the Merchandising Agent can use.

​

strategySignals[]

strategySignals[] turns historicalPerformance into deterministic strategy guidance.

​

outcomes

outcomes summarizes attributed outcomes across the window. Each item in runs[] carries an outcome object for that run: latestType is one of media_buy_submitted_for_approval, media_buy_forwarded, media_buy_forward_failed, media_buy_rejected, or delivery_reported.

​

buyers[]

buyers[] rolls the same window up by buyer identity.

​

seasonality[]

seasonality[] buckets the same window by UTC month.

​

postureConversion

postureConversion is the learning readout: which negotiation posture your Merchandising Agent actually used is booking deals, and whether following the history-derived recommendation booked better than diverging from it. It is the proof behind the recommendedPosture that appears throughout this payload. It distinguishes two postures per run: the selected posture (what the agent used this time) and the recommended posture (the history-derived suggestion — the same one documented under historicalPerformance.recommendedPosture).

​

postureConversion.byPosture[]

Conversion rolled up by the posture the agent actually used (the selected posture). Runs where no composition ran — and so have no selected posture — are excluded from every denominator.

​

postureConversion.adherence

How often the agent followed its own history-derived recommendation, and whether following it converted better than diverging from it. How to read it: when followedWinRate is higher than divergedWinRate, following the learned recommendation is converting better — the signal the learning loop is working. The reverse is a prompt to revisit your operating instructions. Both rates stay null until there are enough comparable runs to compare.

​

learnedDefault

learnedDefault is the negotiation posture your Merchandising Agent has saved as its own default — the first rule the agent writes for itself. Where postureConversion is the live readout of what is converting, learnedDefault is the persisted rule distilled from it: a durable, visible, revertible default that survives across sessions. learnedDefault is null until a default has been saved. Two conditions must both hold before the agent saves one:

1. A sustained, evidence-gated signal. A single posture must be both converting above the floor and backed by enough recent runs — a one-off win never becomes a saved default.
2. The learning loop is switched on for your storefront. Saving a default is gated and off by default; until it is enabled, the agent reads its posture live on every request and saves nothing.

How it is used: the agent applies learnedDefault only as a fallback — when it has no fresher, buyer-specific read for the request in front of it. It is never a forcing override. Your operating instructions and any buyer-specific rules always take precedence, so you stay in control of how your storefront sells. Clearing or pinning it: you stay in control of the saved default. Ask your agent to pin a specific posture as the default yourself — the saved reason then reads "Pinned by you" and the agent will not relearn over it, even while it keeps learning, until you change it. Or ask it to clear the default (revert to agent-managed): the saved default is removed and the agent may learn a new one from outcomes again. Either way the learned default remains a fallback the agent uses only when it has no fresher read; pinning or clearing never changes which postures the agent considers, only which one it falls back to.

​

sellerRecommendations[]

sellerRecommendations[] turns those directional signals into seller actions.

​

Learned auto-approve suggestions

When the media buys from one buyer that came to you for review have all cleared and forwarded with no rejections, the agent surfaces a learned_auto_approve recommendation flagging that buyer as consistently low-friction and pointing you at the per-buyer auto-approve control — so you can opt that buyer in with one step if you trust them to keep clearing. This looks only at media-buy approvals — not creative review (a separate gate) and not delivery performance. It is suggest-only: the agent recommends, but you author the carve-out — it never changes an approval gate on its own. It learns the pattern from your own decisions, and it’s evidence-gated — a short or mixed approval history yields no suggestion.

​

Related