Commerce Analytics

Data foundations, reporting, dashboards, internal tools

The dashboard
is the easy part.

Commerce Analytics builds the layer underneath it: pipelines out of the source systems, one set of agreed definitions, and reporting that says what moved and why. Almost every argument about a number happens before anyone opens a chart, and that is where the work is.

  • Selling Partner API Seller Central and Vendor Central: orders, settlement, catalogue, inventory, fees
  • Amazon Ads Sponsored Products, Sponsored Brands, Sponsored Display, and the demand-side platform (DSP)
  • Amazon Marketing Cloud Clean room analysis: reach, frequency, overlap, path to purchase
  • Retailer portals Grocery and pharmacy retail media exports, scan and loyalty-matched sales
  • Finance and cost systems Invoiced revenue, landed cost, freight, promotional funding

Why it is hard

Four honest answers, and none of them is wrong

Ask how much a brand sold last week and there are several defensible answers, because the surfaces are measuring different events. Put the raw exports side by side and the totals will not agree, and they are not supposed to. Choosing which one the business runs on, writing that down, and holding the line is the job.

  1. 01 Sales and Traffic Ordered product sales, gross of promotions
  2. 02 Settlement Net of promotions, fees, and refunds, on a settlement cycle
  3. 03 Shipped cost of goods sold Amazon's sell-in to itself, not sell-out to a shopper
  4. 04 Finance system Recognised on invoice, in the reporting currency
Illustrative. The axis carries no scale and the bars carry no values, on purpose: the point is the spread between the surfaces, not the level of any one of them. Every real version of this chart belongs to a client, so there is not one on this page.

  • Amazon restates history

    Settlement lands on a cycle, Brand Analytics backfills, and vendor reporting revises. A figure pulled on Monday and the same figure pulled on Friday can both be correct. If the pipeline does not keep the date of the pull alongside the date of the sale, nobody can tell a restatement from a mistake, and the meeting turns into an argument about the data instead of the business.

  • Timezone and currency are not formatting

    Advertising change history is returned in Coordinated Universal Time (UTC), so a single local-day query run from Australia quietly misses part of the day the team actually worked. Marketplaces are separate accounts on separate calendars in separate currencies. Two markets are comparable only after somebody decides how they are made comparable, and records the decision.

  • Margin needs data the platform does not hold

    Landed cost, freight, and promotional funding live in the finance system, not in Seller Central or Vendor Central. Net pure product margin in Vendor Central is the retailer's margin, not the brand's. Profitability reporting is a modelling exercise before it is a reporting exercise, and treating it as a chart problem is how brands end up optimising toward the wrong number.

What we build

Four things, in the order they have to happen

The sequence matters more than the list. Reporting built before the definitions are agreed has to be rebuilt, and a dashboard built on a fresh export is a second version of the truth on day one.

01

Data foundations

Pipelines, modelling, definitions

Automated pulls from the source systems into one modelled layer, with the source systems left canonical for their own objects.

The unglamorous half is the definitions: which source wins for each attribute, what counts as a sale, how refunds and restatements are handled, and what happens when a surface changes its own history. We write those down where the business can read them, because an undocumented definition becomes a folk belief inside a quarter.

  • Incremental extraction
  • Modelled layer
  • Canonical source per attribute
  • Restatement tracking
  • Data quality tests
02

Reporting

Weekly, monthly, quarterly

The recurring pack, built from the modelled layer so a figure in the weekly matches the same figure in the quarterly.

Commentary is written to the finding, not the metric. What moved, why it moved, and what changes because of it. A report that lists percentage changes without naming a cause or a next step is a status update, and a status update does not need a data team to produce it.

  • Business review packs
  • Variance commentary
  • Cause and action, not movement alone
  • One number, one definition
03

Dashboards

Self-serve views

Views for the people who own the number, built on the modelled layer and never on a fresh export.

Two dashboards built on two exports will disagree within a fortnight, and the disagreement will surface in front of a client. Each view is scoped to a decision rather than a subject, with the question it exists to answer written on the face of it. A dashboard nobody opens is a maintenance cost, so we go looking for those and retire them.

  • Decision-scoped views
  • Shared definitions
  • Access and row-level scoping
  • Usage reviewed, not assumed
04

Internal tools

Applications, alerting, connectors

The small software that takes manual work out of the week and puts data where the work already happens.

Account audits, scorecards, and alerting on the conditions that actually cost money: a lost Featured Offer, a suppressed listing, a campaign out of budget by mid-morning. Alerts are outcome-led rather than activity-led, because activity monitoring gets gamed and burns the goodwill you need for the alerts that matter. Tooling is judged on hours returned and errors prevented, not on how much of it exists.

  • Audit and scorecard systems
  • Exception alerting
  • Workflow connectors
  • Retired when superseded

How it runs

We measure ourselves on how much work leaves

A data team that does everything personally becomes the bottleneck it was hired to remove. The counter to that is a rule about the third time, and we apply it to ourselves rather than describing it in a deck.

First time

We do it

Someone needs an answer. We get it, fast, without a process conversation attached.

Second time

We templatise while doing it

Same shape of request, so the answer gets built as something repeatable rather than as a one-off.

Third time

You get the template

One link, with the instructions inside it, that a non-technical person can finish without coming back to us.

  • Definitions before charts. The definition of the metric is agreed and written down before anything is built on top of it. Retrofitting a definition means rebuilding everything downstream of it.
  • Every recurring build has an owner outside this team. If the only person who can maintain a report sits here, the report is a dependency rather than an asset.
  • Source systems stay canonical. We do not rebuild the tools you already run on. The warehouse is authoritative only for the state it derives itself.
  • We retire as willingly as we ship. Every quarter something gets turned off. Reporting estates grow by default and nobody is ever thanked for the pruning, so it has to be scheduled.

Who this suits

Useful to a specific kind of team, and openly wrong for others

Worth establishing before anyone books time. The second column costs us work and is the more useful of the two.

Worth a conversation

  • Somebody owns the number. A data team with no counterpart on the commercial side builds views that nobody acts on.
  • The question is a decision. Pricing, range, media split, inventory. Something is waiting on the answer.
  • Your existing reports disagree and nobody can currently explain why. That is a solvable problem and a common one.
  • Willing to agree one definition and live with it, including when it makes a number you have already reported look worse.

Probably not us

  • Hoping a dashboard will settle an argument that is actually about strategy or ownership. It will not, and it will get blamed.
  • Wanting a copy of a report you already have, rebuilt in a different tool with a different logo on it.
  • Needing a live view of a source that reports daily. Most of these surfaces settle over days and revise afterwards. Refresh rate cannot outrun the source.
  • Looking for a licence rather than a team. There is software that does the generic version of this well and costs less.

Get in touch

Send the question and the systems it lives in

Tell us what decision is waiting on the answer, what you are already pulling and from where, and the place your reporting currently disagrees with itself. We will come back on whether this is a data problem and what it would take, including when the honest answer is that it is not one.

Enquiries

hello@commerceanalytics.app

Analytics, reporting, and internal tooling for ecommerce and retail media teams. Australian market.