María S. López Oller

Each brand collected web data its own way. I created the Google Tag Manager governance framework for 60 brands^M1 and more than 100 websites, across every region including China.

I led the migration of the consent management platform, including the Consent Mode validation under GDPR and CCPA, and built the business case for Google Analytics 360 by measuring what poor data quality was costing. I also ran BigQuery analysis for the brands' marketing teams and put in place what they asked for.

• Stack
• Google Tag Manager
• Consent Mode v2
• Chrome DevTools
• Google Tag Assistant
• GA4
• BigQuery

Digital identity is harder to measure than most government services: the UK has no national ID card, so a service has to verify who someone is without one. I worked in several Agile, multidisciplinary teams at once, and created and validated new metrics and KPIs whose data quality I owned.

For my team's analytics work, I wrote the specifications developers built from: user stories and UAT acceptance criteria. With the Head of Privacy, I implemented GDPR across the unit, covering consent and sensitive data handling. Our team automated the recurring reporting with Python and ETL pipelines and cut reporting effort by three quarters^M2, with weekly reporting to 30+ stakeholders^M3, 100% on time.

I also trained as a GDS service assessor and sat on Service Standard assessment panels as the performance analyst.

ResourcesBlog post · writing analytics specs developers can build

• Stack
• User stories
• UAT criteria
• KPI definitions
• GDPR
• Python
• ETL
• Jira

Marketing and ecommerce didn't trust each other's figures. With the director, I started from a measurement plan and agreed objectives with each area before defining any KPI. I ordered the backlog by dependency and by how many teams each piece unblocked, so no one built a metric that hadn't been agreed. Each delivery came with a working dashboard and training.

New analytics requests grew 70%^M4 year on year, with 90%^M5 stakeholder satisfaction.

ResourcesPresentation available on request

• Stack
• Measurement plan
• KPI definitions
• Looker
• Looker Studio
• Power BI
• Jira

The business wanted the checkout redesigned. Before anyone committed engineering time, I compared the funnel step by step with the retail benchmark, and found a page carrying 40%^M6 of checkout traffic that had never been tracked. I fixed that first.

Three teams wanted three different fixes. Marketing wanted Customer Care on video calls, Customer Care wanted chat on product pages, and the evidence pointed at the checkout. I showed each team, step by step, what data existed at each stage and what each option would tell us, and we started with the checkout.

Then we went after the reasons, with a Hotjar survey categorised through the OpenAI API and live chat reasons logged by Customer Care in Salesforce. Payment failures came out as the main cause. After the changes shipped, checkout conversion rose 4%^M7 against the pre-launch baseline, and Customer Care kept using the weekly report in its own meetings.

ResourcesBlog post · the full story

• Stack
• GA4
• Google Tag Manager
• BigQuery
• SQL
• Python
• Hotjar
• Microsoft Clarity
• OpenAI API
• Salesforce

Nutrition professionals had no AI tool built around their client work. I took the product from concept to beta. I coordinate the team and decide what to build next from interviews with more than ten of them.

We chose to give the AI to nutritionists only, so a professional reviews each recommendation before it reaches a client. The project reached the final phase of the INCOOVA incubator, one of 12^M8 ventures.

ResourcesBlog post · why we put nutritionists in the loop

• Stack
• LLM APIs
• Vertex AI
• Google Cloud
• BigQuery

The target was double the previous year's leads while the budget was being cut. I led a team of four, three marketers and a designer, and ran the marketing relationship with white-label partners. I agreed what counted as a lead, read the results channel by channel and moved the money to what was paying back.

Leads rose 112%^M9 in six months with a marketing budget 48%^M10 lower, and cost per lead went from €36.20 to €9.38^M11.

ResourcesPresentation · PDF, 10 slides

• Stack
• Google Analytics
• Google Ads
• Facebook Ads
• Google Display Network
• Google Trends
• CRM

Goldcar Club, the car rental loyalty programme, needed a way to tell which members were worth more to the business. With the Chief Digital Officer, I defined customer lifetime value^M12 for its members.

I built it in Python and SQL, starting from RFM segmentation: how recently, how often and how much each member had spent. Marketing then used it.

In the same role I led the measurement plan across 15 countries.

• Stack
• Python
• SQL
• RFM
• Customer lifetime value

The order of a category page decides which products customers see first. The performance team asked to test new ranking algorithms in Sprinter's internal merchandising tool: one that reordered the grid from what the visitor had been browsing, and one that pushed up products related to those they had looked at.

It was a true A/B test: visitors were split at random into groups that saw each ordering at the same time, so the comparison wasn't skewed by season or campaigns, unlike a before-and-after. Success was revenue per visitor, with clicks on the grid and add-to-cart as early signals.

The hard part was the unit of the split. Splitting by user is the right choice here: the personalised ordering learns from what a visitor browses, so a split by session can show the same person both versions on different visits and blur the difference. But recognising a returning visitor needs a cookie, and on Sprinter that cookie depended on consent, so the technical team couldn't split by user and the test ran by session, against a control group with the current ordering.

How I'd check a test like this today: three arms (the current ordering as control and the two algorithms) need a multivariate test with a correction for multiple comparisons. Clicks and add-to-cart are proportions, so a two-proportion test fits. Revenue per visitor is skewed by a few big baskets, so it needs a test that copes with that, such as Welch's t-test on a large sample or a bootstrap. And with a split by session, the session has to be the unit of analysis too. My free calculator covers the planning, the multivariate set-up, several metrics at once and the risk of picking the wrong winner.

ResourcesA/B test calculator · built by me

Experiments were a team effort, around five a quarter^M13. I worked with the technical team to validate each test before it went live, and I evaluated the results.

A practice project on how I'd structure analytics a team can trust. Fourteen business questions about a fictional clinic, from 30-day readmissions to medication adherence, each answered by its own dbt model on synthetic data. No real patients are involved.

The data comes with defects planted on purpose: duplicated intake forms, a physician with no department, stays discharged before they were admitted and impossible lab values. Bad rows are flagged, not hidden, and each model decides what to exclude. 50 tests^M14 and a data quality scorecard catch the defects: known ones warn without blocking the build, real breaks fail it.

It also works as a data catalogue. Every model has an owner, a data classification and its personal data columns tagged, and the lineage graph runs from the raw data to the reports that use it. A check in the pipeline fails the build if any of 34 models and seeds^M16 is left undocumented.

How it was built, with AI agents, is in Agentic AI in my own work.

ResourcesGitHub · code, tests and resultsCatalogue and lineage · dbt docsUse case and slides available on request

• Stack
• dbt
• DuckDB
• SQL
• Python
• GitHub Actions
• dbt docs