Skip to content

Welcome to the data floor of the Rafaverse

The Data World Store.

Mind the pipes. Read the board.

Analytics is not a philosophy club — it exists to move a commercial number. Nothing goes on the shelf unless it ladders to the objective and we can change it by Friday. And everything below the floorboards is there so the numbers above them mean something.

The door chimes. The shopkeeper looks up.

Aisle 01 / The Counter

Start at the objective. Work back to what you control.

Stoicism without a target is just a mood. The shop’s method is a chain with four links: name the commercial objective — LTV, retention, conversion, one number that actually matters; decompose it into the levers you control this week; work those levers relentlessly; and read the scoreboard only to check the chain, never to steer by vibes.

The corollary cuts both ways: a scoreboard metric you cannot move is weather — and a perfectly controllable metric that does not ladder to the objective is a hobby.

02 — The shopkeeper points deeper into the shop…

Aisle 02 / The Workbench

The Two-Column Board.

Pick the objective first. Then sort what actually moves it from what merely gets reported.

Objective on the wall: grow customer LTV — everything below serves it

The board — sort the twelve

0 / 12 filed

Unsorted — tap a chip, then a column. Or drag it straight in.

03 — Past the counter, something bubbles…

Aisle 03 / The Testing Lab

Conversion is a verdict, not a lever.

CRO is applied epistemology wearing a marketing badge. You do not control conversion — you control the quality of the question you ask it. A proper test is a question asked so cleanly that reality has no room to be misread: one hypothesis, one change, a sample big enough to mean something, and a runtime you committed to before you saw a single number.

The scientific method survives contact with marketing exactly as long as nobody peeks. Peek at day three, call the winner, and you have not run an experiment — you have run a ritual. The variants did not converge; your patience did.

And every test in the lab exists to move the number on the wall. An elegant experiment that cannot ladder to the objective is theatre with a p-value — interesting, publishable, and commercially inert. The lab runs on a simple sorting question: if this wins, which lever moves, and which objective feels it.

Peeking is prayer dressed as analysis.

Guess before you look

What share of online shopping carts get abandoned?

50%

No peeking — the truth only prints after you commit

Principle — the tester’s dichotomy

You do not control which variant wins.

You control the hypothesis, the sample, the runtime, and whether you peek. Optimise those and the win rate optimises itself — slower than you want, faster than you deserve.

  1. 01Observe
  2. 02Hypothesise
  3. 03Predict
  4. 04Run
  5. 05Verdict
  6. 06Compound

The loop runs on discipline, not luck — step six feeds step one.

Do not peek at the beakers

04 — A quieter corner, where the jars listen…

Aisle 04 / The Discovery Corner

Personalisation is applied listening.

Discovery is the shop’s quietest department and its most valuable. Every interaction a customer has with you is an answer to a question you did not have to ask — what they opened, what they ignored, what they almost bought at 11pm. Most companies file those answers in a drawer labelled “engagement metrics” and never read them again.

The Rafaverse method: treat every signal as inventory. An event becomes a trait, a trait joins a segment, a segment meets a moment, the moment gets a test, and the test teaches the next one. That chain — observe, resolve, decide, personalise, learn — is how intelligence compounds while you sleep. Personalisation is not a campaign; it is a shop that rearranges its shelves for whoever walks in.

And it feeds the till directly: discovery fills the hypothesis queue, the lab turns hypotheses into verdicts, verdicts move the levers, levers move LTV. Learn on the go, or pay to relearn on the next campaign.

  1. 01Observe
  2. 02Resolve
  3. 03Decide
  4. 04Personalise
  5. 05Learn

Observe — Every click is a customer telling you something for free.

The loop never closes — step five restocks step one.

Guess before you look

How many shoppers say they’re more likely to buy when the experience is personalised?

50%

No peeking — the truth only prints after you commit

Principle — the listening dividend

Discovery converts attention you already paid for into intelligence you do not have to buy.

The cheapest data you will ever own is the signal customers already gave you. Resolve it, act on it, learn from it — before buying a single new tool.

05 — A trapdoor behind the counter. Mind your step…

The junction, mid-repair — two identifiers becoming one customer

Aisle 05 / The Cellar — below the floorboards

Your CDP is a promise, not a product.

Every number on the shop floor is only as honest as the pipe feeding it. Identity resolution is a governance habit — the software just keeps score.

6 min below the floorboards · Segment · Redshift · Braze

Nobody buys a CDP. They buy a promise — a single view of the customer, real-time everything, every downstream tool finally speaking the same language. Then the contract is signed, the connectors light up green, and eighteen months later someone in a QBR asks why the email platform thinks a customer is two different people. I have sat on both sides of that meeting.

CDP

A beautiful machine, a familiar problem.

Here is the uncomfortable part: the software was never the problem. Segment does what it says. Redshift holds what you give it. Braze sends to whoever you tell it to. The gap is always in the promises the organisation quietly declined to keep — the naming convention nobody enforced, the mobile team that ships events without telling anyone, the "temporary" user ID format from a 2021 migration that is still generating profiles today.

The 'Temporary' Fix

[Temp_ID]2021

The Unconnected Pipe

web_events mobile_events

The Forgotten Treaty

Schema v1

Identity resolution, specifically, is not a feature you configure. It is a governance habit you either practise or you do not. Every identifier that enters your stack is a small treaty between teams: this is what it means, this is who owns it, this is when it may change. The CDP just keeps score of how well you honour those treaties. When the score is bad, buying a different scorekeeper does not help.

The plumbing lessons from running Segment, Redshift and Braze as one system, in no particular order. Write your tracking plan before the event exists, and make the plan the contract — an event that is not in the plan does not ship. Treat anonymous-to-known identity stitching as a product with an owner, not a checkbox in a connector. Keep one table in Redshift that is the arbiter of who a customer is, and make everything downstream — including your warehouse-sync into Braze — read from it rather than deriving its own opinion.

01

The Blueprint

Tracking Plan

02

The Master Valve

Identity Stitching

03

The Central Reservoir

Source of Truth

The part that has changed recently: agents raised the stakes. When a human builds a segment, bad identity data produces a slightly wrong audience and a shrug. When an agent queries your warehouse and acts on the answer in real time — which is exactly what our MCP-based analytics setup does — bad identity data produces confident, automated, wrong actions at machine speed. Clean data used to be a quality goal. It is now a safety requirement.

Before: Human Analyst

“Hmm, this segment looks a bit off.”

After: AI Agent

CONFIDENT. AUTOMATED. WRONG.

So my advice, unfashionably, is to spend less time evaluating CDP vendors and more time writing down what your identifiers mean. The schema audit nobody wants to run will do more for your personalisation programme than any migration. A CDP is a promise. The product is the discipline of keeping it — and the discipline, inconveniently, is not available on a usage-based pricing tier.

The right tool

Discipline

06 — Back up the ladder. Under the jar shelves, a drawer of blueprints…

Aisle 06 / The Blueprint Drawer

The best customer data template is a contract.

Every shop keeps one drawer it reaches for more than all the others. This one holds the best customer data template I know — the tracking plan this site runs on, generalised. It is not a list of events to copy; it is the order you decide things in: six questions the business will act on, then the naming contract, then identity, and only then a single event table. Journey tracking becomes actionable data at exactly the moment every row can name who acts on the answer.

Inside: Object + Action names that cannot drift; a property dictionary where one name means one thing; identity rules that admit there is exactly one honest identify moment; a context layer that stamps who, where, and how far in onto every event; and — the unusual part — a defect ledger, because a plan that admits its faults is the only kind whose clean rows mean anything.

And it leaves the shop as a skill. Drop the file into your agent’s skills folder and “draft our tracking plan” stops producing a list of button clicks: the agent interviews for the six questions first, refuses events that answer none of them, and writes the defect ledger before anyone is allowed to be proud of the plan.

Blueprint — customer-data-template

  1. 01Questions before events — who acts on the answer
  2. 02The naming contract — Object + Action, Title Case
  3. 03Identity — every identifier is a treaty
  4. 04The event table — call sites, not intentions
  5. 05The property dictionary — one name, one meaning
  6. 06The context layer — to whom, where, how far in
  7. 07The defect ledger — append and strike, never delete
  8. 08Governance — the rituals that keep it true
Take the skill — SKILL.md

Free with any visit · no email asked · yours to keep

Principle — the door policy

Events earn shelf space by answering questions.

Name the question, name who acts on the answer, and only then name the event. Everything else is storage with ambitions.

07 — Through the door marked staff only, where the ledgers are kept…

Aisle 07 / The Back Office

Clean data is an agreement, not a cleanup.

The template has exactly one natural predator: other teams. Web ships Signed Up, mobile ships sign_up, the relaunch mints user_registered, and two roadmaps later a funnel is comparing three names for one behaviour — a dashboard lying without a single false number in it. Consistency is not a matter of style. It is the difference between data and noise wearing data’s clothes.

Which is why accuracy is not a cleaning job downstream; it is an agreement upstream. The back office runs four rituals, and none of them is software: the plan lives in the repo, so a new event is a pull request; a check runs in the build and fails it the moment code and plan disagree — this page’s own plan ships behind exactly that gate; every identifier has one named owner, a person rather than a team; and a rename is a migration with a date, never an edit.

The payoff is trust, and trust is brutally asymmetric: the first time a number is caught being wrong, every number after it is negotiable. Keeping every team aligned on what the words mean is the main challenge of producing real data — not because it is hard to understand, but because it is boring, weekly, and only works when it is enforced.

Guess before you look

What share of companies' data meets basic quality standards?

50%

No peeking — the truth only prints after you commit

Principle — the treaty test

A metric is only as real as the treaty behind it.

The cellar named the treaty — what an identifier means, who owns it, when it may change. This is the room where treaties are countersigned: honour them and accuracy stops being a project and becomes a habit; break them and no amount of downstream cleaning puts the meaning back.

08 — Out front again. New stock rattles in from the AI era…

Aisle 08 / New Arrivals

The instruments change. The discipline does not.

What the shop stocked lately — the analytics shelf after AI, restocked August 2026. Same sorting question for every shiny new instrument: which lever does it move, and which objective feels it.

New in

Talking is table stakes

Every instrument on the shelf now answers plain English. Nobody differentiates on chat anymore — the only question left is what grounds the answer.

New in

The truth shelf wins

Raw text-to-SQL agents score around 65% — one enterprise study watched a 91% benchmark hero collapse to 21% on real data. Put a governed semantic layer underneath and frontier models hit 98–100%. Buy the semantic layer, not the chatbot.

Restocked

Agents mind the store

Agentic analytics is a named Gartner category now. Amplitude’s agents already run a quarter of its platform queries. Dashboards are being demoted to the evidence an agent shows when you ask to see its working.

Restocked

One plug fits every socket

MCP became the connective standard — Power BI, Tableau, Looker, Qlik, ThoughtSpot, Metabase, Amplitude, Mixpanel all ship servers. If your agent cannot reach a tool, treat that as a red flag on the tool.

Staple

New shopkeepers on the street

Omni — built by Looker’s old guard on a governed-semantics thesis — reached a $1.5B valuation, profitable. Sigma raised at $3B chasing the same wave. Mode, a decade’s darling, was bought and quietly retired. Shelves get retired.

Staple

The discipline did not move

Eighteen months rearranged the entire shelf. The discipline never moved: govern your meanings, control your inputs, keep a human veto on anything that acts, and distrust any confident answer without a semantic spine.

Shelf stocked August 2026 · figures true at stocking — recheck before you lean on one

09 — Everything rings up at the till.

Aisle 09 / The Till

Where the measuring got real.

One last thing before you go. The whole time you were browsing, this shop was doing to you exactly what it teaches — watching which aisles held you, logging your guesses, counting your stamps. Every good shop itemises.

Itemised from your own walk through the store

Stoic Analytics — the shop’s writing deskAgentic analytics on the MCP stackSanta AI: stoic analytics in productionThe measured half, on paper

Keep wandering

Martech Garage →← Back to the villageTalk pipes with me