Speak the
language of profit.

The human gate. Every money-moving action waits here.

The queue of agent-proposed actions awaiting human approval before they touch a platform. Each entry logs the prompt, the simulation, the proposed payload, and the eventual approver. Reversible in one click.

Revenue minus the costs that scale with the order.

Revenue minus landed COGS, fees, shipping, returns and incremental CAC. The number your bank account already knows, the truth that the platform-reported numbers obscure.

The metered unit of agent work. Predictable, never per-seat.

Each deep investigation, scheduled report, or simulator scenario consumes a fixed number of credits. The model is fixed plan + credits so a 2-person founder team pays the same per investigation as a 200-person analytics team.

A candidate cause of a metric move, tested by the agent.

When an investigation decomposes a question, each candidate cause becomes a driver, labelled CONFIRMED, TESTING or REFUTED. Drivers must sum to the actual move on the books before the investigation is allowed to close.

How much of the attributed revenue is actually caused by the campaign.

The estimated share of attributed revenue that wouldn't have happened anyway via organic, email, or retargeting overlap. Measured via geo-lift, holdouts, or synthetic-control modelling. The single biggest gap between reported and true contribution.

A deep agentic run answering one question.

A self-contained agent run that decomposes a question into drivers, fans out across every connected source, reconciles to a bridge, and ends in a queued recommendation. Costs 10 credits and typically completes in 3–8 minutes.

The ROAS that survives the holdout.

Return on ad spend computed only against the incremental revenue, the orders that would not have happened without the campaign. Typically 30–60% lower than the platform-reported ROAS. The number that determines whether you should keep spending.

Your true unit cost, supplier + freight + duty + warehouse-in.

Cost of goods sold computed at landed cost, the supplier price plus inbound freight, duty, customs, and the warehouse-receiving fee. The number that turns gross margin into a real number.

The waterfall that explains a margin move dollar by dollar.

A contribution-by-contribution waterfall: revenue, COGS, fees, returns, CAC, mix, each driver stacked until the deltas reconcile to the actual margin move on the books. Used to validate every investigation result.

Drivers sum to the actual move, or the investigation re-runs.

The rule that keeps the agent honest. If the proposed drivers of a metric move don't add up to the move on the books, to the dollar, the investigation is rejected and re-runs. The math doesn't ship until it sums.

Platform-reported attribution vs the truth-after-holdout.

ROAS is what the ad platform reports, revenue divided by spend, against its own attribution model. iROAS is the same ratio but using only the incremental revenue. For most operators, iROAS is 40–60% of reported ROAS.

Counterfactual modelling before committing real money.

A what-if engine that models the outcome of a proposed action, pausing a campaign, reallocating budget, repricing a SKU, with side-effect and worst-case panels. Used as the input to every approval-queue decision.

Contribution profit, computed end-to-end and bank-reconciled.

margininfo's name for the contribution-margin P&L: revenue minus landed COGS, fees, shipping, returns and incremental CAC, computed 12 months back, per SKU and per channel, reconciled to your bank total.

The flagship question class the agent is built for.

Any natural-language metric question of the form 'why did X move?'. The diagnostics pillar of the platform. Decomposes into drivers, reconciles to a bridge, and ends in a ranked queue of fixes.