Attribution that admits its limits, dashboards that answer the question you actually asked, alerts that catch problems the same day, and a data pipeline that respects consent by design.
From sources to decisions, with a privacy boundary in the middle
Sources pass a consent gate before anything is stored. Inside the dashed privacy boundary, the warehouse holds pseudonymous records that attribution, forecasting and anomaly models read. Outputs are dashboards for people and alerts for problems.
Attribution
Several lenses, shown together, settled by experiment
A buyer reads a guide found through an AI assistant, returns via a search ad, gets a nurture email, and books a call from a direct visit. Which channel deserves the credit? Every attribution model gives a different answer, and all of them are rules, not truth.
We implement the models that matter for your decisions, present them side by side, and flag where they disagree by more than a chosen margin. For the disagreements with money attached, we run incrementality tests: hold out a region or a segment, and measure what happens without the channel. That is the only attribution that stands up in a budget meeting.
We also account for the dark channels: AI assistant citations with no click, and word of mouth. Our visibility monitor and a simple "how did you hear about us" field on the contact form both feed the model.
Model
Credits
Flatters
Use for
Last touch
The final interaction
Brand search, direct, retargeting
Short cycles, quick reads
First touch
The first interaction
Content, ads, referrals
Judging awareness channels
Position-based
First and last most, middle shared
Both ends
A reasonable default for B2B
Data-driven
By statistical contribution
Whatever the data says
High volume with clean tracking
Incrementality test
Measured causal effect
Nothing
Settling budget decisions
Dashboards and alerts
Fewer dashboards, each answering a real question
Question-led dashboards
"Which pages produce qualified leads?" "What did last month's content earn?" One dashboard per question, with the definition of every metric written next to it. Built on your warehouse, not a vendor's black box.
Ask in plain English
A language model turns a question into a query against a governed data model, shows the query and returns the result. Built with the same care as our LLM applications: tested, logged, limited to what it is allowed to read.
Anomaly alerts
Models learn the weekly and seasonal shape of each metric and alert on deviations: a tag that stopped firing, a form that broke on mobile, a campaign overspending. The same approach as fraud and anomaly detection, applied to marketing.
Privacy by design
Consent is checked before a tag fires. Server-side tagging limits what third parties receive. The warehouse stores pseudonymous identifiers with a retention period. No personal data is sent to a model API; the question layer works on aggregates and ids. The approach is documented in our AI data privacy practice, and it is why we can run analytics for clients in regulated sectors.
Process
Foundation, then models, then habits
Measurement plan
The questions leadership asks, the metrics that answer them, the events needed, the consent model. Week 1.
Tracking and warehouse
Server-side tagging, event schema, sources connected, pseudonymisation and retention set. Our data engineering team. Weeks 2–5.
Dashboards
The first three question-led dashboards and the plain-English layer. Week 6.
Models
Attribution lenses and anomaly alerts once a few months of clean data exist. Forecasts after a year.
Review cadence
A monthly review where numbers lead to decisions and the measurement plan is updated.
None of them is true; each is a lens. Last-touch flatters the channel that closes, first-touch flatters the one that introduces, position-based and data-driven models spread credit by rules or by statistics. We implement two or three side by side, show where they disagree, and use incrementality tests to settle the arguments that matter financially.
Can we just ask the dashboard a question in plain English?
Yes, when the data model underneath is sound. We build a natural-language layer over your warehouse that turns a question into a query, shows the query it ran, and returns the chart with the numbers. Transparency matters: a chart with no visible query is a guess with a nice font.
What is an anomaly alert and why do we need one?
A model learns the normal pattern for a metric, including weekly and seasonal cycles, and alerts when a value falls outside it. It catches a broken tracking tag, a form that stopped submitting, a campaign that is overspending or a sudden traffic drop, usually within hours rather than at the monthly review.
How do you handle privacy and consent?
Consent is enforced before any tag fires; server-side tagging limits what third parties receive; personal data is pseudonymised in the warehouse with set retention; and no personal data is sent to a language model API. We document the lawful basis for each data use. See AI data privacy for the full approach.
How long does an analytics project take?
A tracking and warehouse foundation with the first dashboards takes four to six weeks. Attribution and anomaly models follow once there are a few months of clean data. We prefer a solid foundation to a quick dashboard on broken tracking.
Ready to make AI work in production?
Tell us about your systems, your data and your constraints. We reply within one business day with a straight answer and a suggested next step.