Bidding toward nothing
A primary conversion action imported eighteen months earlier from an analytics event the site no longer sent. Smart Bidding had been optimising toward a conversion that never fired, at full spend, for a year and a half.
Marko Cvijić / Analytics and attribution
Analytics and attribution. The paint.
Verified tracking, honest attribution, and reporting that maps to the P&L rather than to a dashboard. This is the layer I am known for holding, and the one every other number on this site depends on - plus the AI-native operations I build on top of it, never underneath it.
The problem
Every decision above this layer inherits its errors.
This is the layer nobody wants to buy and everything else stands on. Budget allocation, channel strategy, bidding, hiring, board reporting - all of it is downstream of measurement. If the measurement is wrong, the decisions are wrong with complete confidence, and the reporting will look healthy the entire time.
The failures are mundane and almost universal. Last-click attribution crediting the channel that closed the deal and defunding the channels that created it. A primary conversion pointing at an event the site stopped sending. Share of voice reported as one equation when it needs five, depending on what you are actually asking. Analytics and the back office producing revenue figures that differ by fifteen per cent, with a standing agreement not to mention it. A dashboard with forty metrics on it, none of which is the number the business is run on.
I hold the paint here: the measurement, the attribution and the governance under everything else. Nobody applauds it, and every number anyone quotes depends on it.
What the work covers
Verified data, then honest attribution.
01
Tracking integrity
Verified, not assumed
Full measurement audit
Every event, every conversion, every property and every integration: what fires it, when it was built, whether it still works, whether it double-counts, and whether anyone still uses what it produces.
Reconciliation
Analytics against the ad platforms against the back office. Where they disagree, by how much, and which one is right. Establishing the size of the gap is often more valuable than closing it.
Server-side and first-party
Server-side tagging, first-party data collection and enhanced conversions where the client-side path has degraded past usefulness. Built to survive the next browser change rather than the last one.
Consent and compliance
Consent mode implemented so it satisfies the regulation without silently discarding half the data - a failure mode that is common, expensive and nearly invisible.
Data layer and governance
A documented data layer, a naming convention, an owner, and a change process. Measurement decays because nobody owns it; this is the fix.
Quarterly re-verification
Sites change and events break. Every primary conversion gets re-checked on a schedule, because the systems consuming that signal do not stop when it fails.
02
Attribution
Which euro did what
Multi-touch attribution
A model that reflects how your customers actually buy, with the assumptions written down and the limitations stated. Every model is wrong in a specific way; you should know which way yours is wrong.
Incrementality
Separating demand you generated from demand you harvested. Brand search, retargeting and shopping on your own product names are where paid most often takes credit for revenue that was already coming.
Offline and long-cycle
Deals that close on a phone call, in a showroom or nine months later, connected back to the origin. Without this, every long-cycle channel looks unprofitable and gets cut.
Margin-aware reporting
Return calculated on contribution rather than revenue. Channels change rank - sometimes reverse entirely - once cost of goods, returns and fulfilment enter the equation.
Share of voice, properly
Five equations, not one, depending on whether you are asking about visibility, click share, category share, paid coverage or answer-engine presence. Collapsing them into a single figure is how a market gets misread.
Reporting to the P&L
One report, in the language of the person who owns the number, with a small number of lines that map to how the business is run. Diagnostics live underneath it, not on it.
03
AI-native operations
Built on top, never underneath
Semi-autonomous systems
AI-powered marketing systems that scale: monitoring, classification, briefing, drafting and reporting handled by machines, with a human owner on every output that touches a customer or a budget.
Where AI does not belong
Automation applied to an unverified data layer produces wrong decisions faster and more consistently than a person would. Measurement gets fixed first. This ordering is not negotiable.
Answer engine measurement
What AI traffic is actually worth once you look at volumes instead of headlines. It is real, it converts differently, and it is currently a fraction of what the discourse implies - which is exactly why it should be measured rather than guessed.
Operating capability
Every business will run more AI than it can staff. The constraint is not the models, it is having someone who can tell whether the output is right - so I build the systems and train the people who own them.
From real audits
What this actually turns up.
A match rate that was fine
Forty-six per cent between ads and analytics, escalated as a crisis. For a long-cycle B2B business that is a normal figure. The genuine problem was three tables away, and a month nearly went into the wrong one.
Budget eaten by crawl
Faceted navigation generating tens of thousands of near-duplicate URLs inside a large catalog, absorbing crawl budget that never reached the pages that sell. Visible only in the logs, invisible in every report.
Operating rules
The rules that make the number defensible.
01
One number, defined once, written down. If two teams disagree about what a conversion is, nothing else discussed in that meeting means anything.
02
Verify before reporting. No performance commentary on data that has not been checked against a second source.
03
State the model's limitations alongside its output. An attribution model presented without its assumptions is a claim, not a measurement.
04
Under a ninety per cent match rate triggers an attribution audit - adjusted for long-cycle B2B, where the realistic floor is much lower.
05
If the honest answer is that we cannot know, that is the answer. Manufacturing a confident figure to fill a slide is worse than an admitted gap.
06
Fix measurement before automating anything on top of it. Otherwise you have built a machine for being wrong at scale.
Fit
Who this is actually for.
A good fit
Your analytics and your back office do not agree and nobody has resolved it. You are allocating significant budget on last-click data. A board or an investor is asking questions the reporting cannot answer. You are about to automate marketing operations and want the foundation checked first. Or you have several markets and properties with no consistent definition across them.
A poor fit
You want a prettier dashboard. You want the numbers presented more favourably. You need a tool implemented to spec and nothing questioned - a good analytics engineer is cheaper and faster for that. Or the organisation is not prepared to act on a finding that contradicts the current strategy, in which case the audit is an expense with no return.