ad platformsanalyticsCRMsourcesthe modelthe forecast, gradedtoday90% bandwhat happenedlast two weeks7142128
last six weekstoday90% bandHold spendholdPush paid +30%+30%Cut paid −25%−25%7142128
Marketing analytics · Kitchener-Waterloo and Toronto
I help B2B marketing teams make better decisions, and understand what worked, what didn't and why.
I build a data foundation you can trust: one that makes sense and is consistent across sources. On top of it, a 28-day forecast with a range, graded at days 7, 14, 21 and 28, and assistants that answer from your own data. For B2B companies that have outgrown their current foundation, need to grow it, or need better answers from it.
history 90% band options what happenedillustrative data
Hold spendPush paid +30%Cut paid −25%
Three sources, one model, a forecast with a range, and reality graded at four checkpoints. One miss, on purpose.
The problem
You're asked which channels earn their budget and what to spend next quarter. The numbers you have can't carry that answer.
Not because the dashboard is wrong. Because it was built to count, not to answer. So the strategy gets written on gut feel, and nobody can say afterwards whether it worked or what it returned.
Three tools, three answers
Ad platforms, GA4 and the CRM each count a conversion their own way. The first hour of every review goes to arguing about whose number is right. That is a missing foundation, not a reporting bug. I make the data make sense, so nobody has to wonder whether a number is accurate.
One line on a slide
Forecasts arrive as a single number with no range. When reality drifts, there is no way to tell a bad call from bad luck, and nothing to learn for the next call.
No record of the decision
Budget moves are made in a meeting and forgotten by the next one. Without a record of what was expected, return on spend is whatever the last dashboard said.
I build a decision system that grades the calls you made during the quarter: what worked, what didn't and why, where to improve, and what to do more of.
What I do
Three layers. Each one depends on the one below it.
Most firms do the data plumbing or the modeling. I do both, on B2B data: pipeline, CRM, long sales cycles. Start at whichever layer your team is stuck on.
ad platformsanalyticsCRMone warehouse, agreed definitionsmetric layer underneathforecast with a rangehistory, as columnsagentschemaqueryforecastnever the raw tables directly
ad platformsanalyticsCRMone warehouse, agreed definitions
forecast with a rangemetric layer underneath
agentschemaqueryforecastnever the raw tables directly
01
Foundations
A foundation your team can trust, built with data-science discipline. I move ad platform, analytics and CRM data into one warehouse, define each metric once with the people who use it, and build the reporting on top. Tracking and tagging get fixed on the way in. The point is not a prettier dashboard. It is that every forecast, test and assistant built afterwards rests on numbers nobody has to argue about.
This won't work if the CRM data is a mess, so that gets fixed first. If the foundation is already good enough, I'll say so and start at layer two.
For
Teams whose channel numbers disagree across tools, or who are about to hire an analyst and want the ground ready.
You get
A warehouse with marketing and pipeline data joined, a documented metric layer, and dashboards built on it.
Length
Six to ten weeks.
Example
A pipeline-by-channel report where marketing, sales and finance read the same number.
Stack
BigQuery, dbt, Looker and LookML, GA4, reverse ETL.
Know which channels earn their budget before you move it. On a trusted foundation I can do three things. A 28-day forecast of your main number with a range and the odds of hitting the target under each option. Attribution across channels where it is honest, which for long B2B cycles means pipeline-based models, not last click. Incrementality tests, geo or holdout, designed before the spend so lift can be separated from noise afterwards.
Marketing mix modeling only when spend and data volume justify it. For most companies under 500 people they don't yet, and I'll tell you that instead of selling it.
For
Teams that own a pipeline or revenue number and want the odds before they spend.
You get
A forecast with scenarios and a target probability, a channel-level read on where the money is working, and the test design for the next big decision.
Length
Four to eight weeks, then ongoing if you want it.
Example
A geo lift test for a regional campaign, with a power check before launch that says whether the test can detect the effect at all.
Assistants that answer from your warehouse, not from the internet. Agents and conversational analytics that run on the metric layer from layer one. Ask what is on pace this month and get the number, the forecast and the working. Each answer comes through a small set of tools I write, so it can be checked. Not a chatbot bolted onto a dashboard, and not workflow automation.
An assistant on an untrusted foundation produces confident wrong answers, faster. I won't build this layer until the first two are sound. Local models are available for teams that cannot send data to a third party.
For
Teams that want analytics on demand without a ticket queue.
You get
An assistant on your warehouse with a documented tool set, evaluation cases, and the option to run it on a local model inside your infrastructure.
Length
Six to twelve weeks.
Example
A Slack assistant that answers “are we on pace this month?” from the metric layer and shows its working.
Stack
Claude or local models through Ollama, Python, your warehouse.
It starts with a quick discovery call about your measurement, not a pitch. If the assessment fits, we scope it on that call.
Measurement Assessment
Three weeks · fixed scope
In three weeks I look at how your marketing, web and CRM data fit together, which of your current numbers can be trusted, and which decisions in the next two quarters deserve a forecast or a test.
What this won't do
It won't fix your data. It tells you exactly what is broken, what it costs you, and what fixing it takes.
It won't give you a forecast to bank on if the history is under a year old or the tracking changed halfway through. I'll forecast anyway and tell you how much to trust it.
It won't replace an analyst. It gives the one you have, or the one you hire, a plan.
A map of your data sources, the gaps, and which numbers can be trusted today.
The decisions worth measuring properly in the next two quarters, with the method for each.
A written plan for the foundation work, sized and sequenced, that you can run with me or without me.
A first 28-day forecast of your main number, with a range, so the next planning meeting starts from odds.
Case study · TrueLift
I built the tool that grades my own forecasts.
TrueLift takes a traffic or pipeline export, forecasts the next 28 days with a range, lays out the options and records the call. When reality arrives, the forecast is graded at days 7, 14, 21 and 28, misses included.
It is the evidence that I can build the whole ladder, from the data layer to the decision. It is not the product I sell; the engagement is.
I am Martin Guzman. I started in digital marketing in 2014 and learned statistics, Python, SQL, BigQuery and causal inference because the questions I was asked needed them. I have consulted in the USA, Canada, Mexico and Finland.
Today I work in marketing analytics for B2B software and take a small number of engagements a year, so the person you talk to is the person who does the work. I am based in Kitchener-Waterloo and work with teams in the Toronto corridor and remotely.