Morning Metrics

01

Eight tabs, eight versions of the truth

It is 7:40am. Coffee, and eight tabs.

Google Ads, Meta, Microsoft, GA4, Search Console, Merchant Centre, the email platform, the store’s own admin. Always in the same order, because I have been doing it long enough to have an order.

Every one of them is telling me a story, and every story is true. Add the platforms up and they have collectively sold more than the business actually took.

None of them are lying. They are each answering a slightly different question: a click they touched, a window they chose, a conversion they counted their own way. Alone, every number is defensible. Together they do not reconcile, so people stop trusting all of them and argue from instinct instead.

That was the moment. Not a gap in a market. A spreadsheet I rebuilt every morning to answer a question eight systems already had the data for.

02

Three things that bothered me every morning

None of it was hard. It was wrong in the same three ways every day, which is worse.

1. 01

   Nobody could agree on a number. Not out of malice, out of definitions. Revenue meant six different things depending on the tab, and nobody had written down which one we ran the business on.

2. 02

   Every question was a project. Which search terms burned money last month? meant an export, a pivot table and half a morning. So mostly, nobody asked.

3. 03

   The answer always arrived after the decision. Reporting was archaeology. By the time the picture assembled, the week it described was over.

03

Land everything, define it once, serve it everywhere

Morning Metrics is one idea, repeated.

Every source lands raw in a warehouse overnight. A transform step turns that history into a small set of daily summary tables: spend, revenue, orders, keywords, products, organic queries, email flows. The definitions live in code rather than in someone’s head, and those summaries are the only numbers anything is allowed to read.

Three things read them. The app, which is dashboards, campaigns, products, attribution, and a nightly rules engine that turns wasted spend and silent email flows into a dollar-valued to-do list. The brief that lands in your inbox at 7am. And an MCP server that lets a marketer connect Claude to the workspace and just ask.

That last one matters more than it sounds. Ask the chatbot what our MER was last week and you get the figure the dashboard shows, because there is exactly one place that figure lives. No more “the dashboard says X but the report says Y”.

04

Building it

I built it solo, with AI coding assistants doing the typing while I did the deciding. That is not a footnote. It is why a marketer shipped a data platform instead of a deck about one.

It is ELT rather than ETL, deliberately. Land the raw history first and transform later, so a definition I get wrong in month two can be recomputed in month three without re-asking eight APIs for data they no longer hold. The warehouse keeps everything, cheaply. The app database keeps only the summaries a page load can afford.

But the interesting problem was never the dashboard. It was making a dozen inconsistent feeds behave like one dataset.

1. 01

   Contracts, not conventions. Every record arrives in a versioned envelope, against an explicitly named dataset rather than whatever shape the API felt like returning that day. Money stays as strings and goes through decimal arithmetic, because floats quietly lose cents at scale. Ratios are recomputed from totals, never averaged. Averaging an average is the most common way a dashboard tells a confident lie.

2. 02

   Capabilities, so nothing guesses. Each connector declares what it can actually provide. Ask a source for keyword data it does not have and the request is rejected before the adapter runs, rather than returning an empty result that reads on screen as a zero. A wrong number is worse than a missing one.

3. 03

   Idempotency, so nights are repeatable. Loads stage into a temporary table, then transactionally replace only that source’s slice of that date window. Re-running a failed night is a no-op, not a duplicate. Every pipeline I had ever been handed as a marketer failed that test, usually on the day I needed it.

4. 04

   Isolation and secrets. Every query is filtered by workspace membership, so one brand’s data cannot surface in another’s. Refresh tokens are encrypted at rest with key versioning, and a source credential can never be used by a destination.

5. 05

   The AI layer, treated as untrusted. The MCP server runs on OAuth 2.1 with PKCE, a consent screen, revocable grants re-checked on every request, rate limits that fail closed, and an audit event per query. Ad copy from the platforms is handed to the model labelled as data, not instruction, because the one thing an attacker can write into your pipeline is ad text.

05

What is missing from this page

There is not a single result on this page, and you should notice that.

Morning Metrics exists to make numbers trustworthy, so it would be a strange place to put an untrustworthy one. I have no user count worth quoting, no measured hours saved, and no before-and-after on a decision that got made faster because the brief arrived at 7am. Each of those would take a minute to write and would not survive being asked about.

The version of this page I would rather be showing you is the one with a baseline on it. I never timed the eight-tab morning before I replaced it, so the saving is obvious to me and unprovable to you. Measuring the thing you are about to change is the cheapest work in any project and the easiest to skip, and I skipped it.

What is here instead is the architecture and the reasoning under it. If you want to test a claim on this page, test a design call: why ELT, why money is a string, why the MCP server treats ad copy as hostile.

06

What I took from it

I started this as a marketer with eight tabs and a spreadsheet. I finished it having designed a schema, written a connector layer, and made real calls about idempotency, tenancy, credential handling and prompt injection.

The most useful thing I learned was not technical. Most reporting problems are not reporting problems. They are definition problems wearing a dashboard.

The hard part was never plotting the number. It was getting everyone to agree what the number was, then putting that agreement somewhere only code can edit.