Marketing Analytics and Attribution AI
Working out which marketing actually caused revenue, using experiments where they are possible and modelling where they are not — and being clear that no attribution model recovers ground truth.
Every platform reports that it drove the conversion. Add the platform numbers together and they exceed the total orders, which everyone notices and nobody acts on. The uncomfortable truth is that last-click attribution measures who was standing nearest when the customer decided, and signal loss has made even that unreliable.
Marketing attribution estimates which activities caused which outcomes. Modern practice combines controlled experiments that measure incremental effect directly, marketing mix modelling that estimates contribution from aggregate data, and platform and journey data used as directional evidence rather than as truth.
The three methods, and what each is good for
| Method | What it gives you | Limitation |
|---|---|---|
| Incrementality experiments | Causal evidence for one channel | Costly, slow, one question at a time |
| Marketing mix modelling | Contribution across all spend, privacy-safe | Correlational; needs history and spend variation |
| Multi-touch attribution | Journey detail at user level | Degraded by signal loss; not causal |
| Platform-reported results | Fast, granular, always available | Self-reported and systematically overstated |
The workable answer is a triangulation: experiments to anchor the truth on your biggest channels, mix modelling to allocate across everything, platform data for day-to-day operation with a known correction applied. Any single method presented as the answer is overclaiming.
How we build it
Start with experiments on the largest spend
Geo holdouts, matched-market tests or scheduled pauses give causal evidence, and the finding is regularly uncomfortable — brand search and retargeting in particular often prove far less incremental than their reported performance suggests. Experiments on the top two or three channels anchor everything else.
Build mix modelling on honest inputs
Spend, exposure, seasonality, price, promotion, competitor activity and external factors. Models fitted to spend alone attribute everything to marketing, including the effect of a price cut or a competitor's outage. Calibrating the model against experiment results is what makes it credible rather than merely fitted.
Treat platform numbers as a source with a known bias
Not discarded — they are the only granular daily signal — but corrected using the experimental findings, so the operating dashboard reflects something closer to reality than the sum of self-reported credit.
Model the lag properly
Marketing effects decay over days or months depending on the channel and the purchase cycle. Attribution windows chosen for convenience systematically under-credit upper-funnel activity, which is how brand budgets get cut on the strength of a seven-day window.
Reconcile to finance
Modelled contribution must sum to actual revenue. This sounds obvious and is frequently skipped, and it is the check that stops an analytics function reporting numbers the finance team does not recognise.
Report uncertainty, not point estimates
A channel contribution of 'somewhere between eight and eighteen per cent' is honest and useful. A single figure to one decimal place is neither, and it invites decisions the evidence cannot support.
Some things cannot be attributed, and saying so is the professional answer
Word of mouth, an AI assistant recommending you in an answer nobody clicked, a conversation at a conference, a competitor's failure. These influence revenue and appear in no model. The honest position is a stated unattributed share rather than distributing it across the channels that happen to be measurable.
What changes once it exists
- Budget moves on evidence. Usually away from channels whose reported performance was largely capture of demand that already existed.
- Upper-funnel activity gets a fair hearing, because lag is modelled rather than truncated by an attribution window.
- Platform optimisation improves, since better conversion signal fed back is what automated bidding actually runs on. See media buying.
- Marketing and finance stop disagreeing, because the model reconciles to actual revenue.
- Testing becomes routine rather than a one-off project, with a calendar of experiments that keeps the model calibrated.
How the engagement runs
Experiments anchor the model, because a mix model calibrated against nothing is a well-fitted opinion.
Data assembly
Spend, exposure, conversions, price, promotion and external factors assembled and reconciled to finance.
Experiment design
Incrementality tests designed for the largest channels; power calculated so results will conclude.
Experiments and modelling
Tests run; mix model built and calibrated against the experimental results.
Reconciliation and reporting
Contribution reconciled to revenue; uncertainty and unattributed share reported explicitly.
Handover
Refresh process, experiment calendar and interpretation guidance handed to your team.
What you receive
Contribution estimates anchored by experiments, reconciled to finance, with uncertainty stated.
Incrementality test results
Causal evidence for your largest channels, with the uncomfortable findings included.
Marketing mix model
Contribution across all spend, calibrated against experiments and reconciled to revenue.
Corrected platform reporting
Daily operating numbers with the known overstatement adjusted.
Lag and decay modelling
Effects modelled over their real horizon rather than truncated by a convenient window.
Uncertainty reporting
Ranges rather than point estimates, with the unattributed share stated.
Experiment calendar
An ongoing testing programme that keeps the model calibrated after handover.
Is this the right engagement?
Worth being direct. Marketing Analytics and Attribution AI is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- Platform-reported conversions exceed actual orders.
- Budget allocation is argued rather than evidenced.
- Signal loss has degraded your existing attribution.
- Upper-funnel spend cannot be justified with current measurement.
- Marketing and finance report different numbers.
Choose something else if
- Spend is too small or too concentrated for experiments or modelling to conclude.
- Insufficient history exists for mix modelling and no appetite for experiments.
- You want a single attribution number with no uncertainty attached.
- The findings will not be allowed to change budget allocation.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Is last-click attribution still usable?
As an operational signal, with known bias. As a basis for budget decisions, no — it credits whoever was nearest at the moment of purchase and systematically under-credits everything that created the demand. Signal loss has degraded it further, so it should be one input among several rather than the answer.
What is incrementality testing?
Deliberately withholding or varying marketing for a group or a region and measuring the difference in outcome. It is the only method that produces causal evidence, and the findings are frequently uncomfortable — brand search and retargeting in particular often prove far less incremental than their reported performance implies.
Is marketing mix modelling worth it for a mid-size business?
It depends on spend, history and whether spend has varied enough for a model to learn from. Without genuine variation there is nothing to fit. Where the data is thin, a programme of incrementality experiments on the largest channels delivers more decision value than a poorly identified model.
Why do platform numbers overstate?
Because each platform uses its own attribution and counts conversions it claims credit for, with no visibility of the others. Summing them double-counts, often substantially. They remain the only granular daily signal, so we correct them using experimental findings rather than discarding them.
What about things we cannot measure at all?
They are reported as an unattributed share rather than distributed across the channels that happen to be measurable. Word of mouth, an AI assistant recommending you, a conference conversation, a competitor's failure — all influence revenue and appear in no model. Pretending otherwise is how attribution loses the confidence of the finance team.
Often paired with this
Most clients combine two or three engagements from the AI for Marketing and Growth pillar. These are the ones that most often run immediately before or after.
AI Ad Creative and Media Buying
Variant production with rights and disclosure handled, and control of what automated bidding leaves you.
Read more →AI Lead Scoring and Enrichment
Scoring trained on real outcomes, enrichment coverage measured by segment, and a score sales will act on.
Read more →Conversion Rate Optimisation with AI
AI to find what to test and read behaviour at scale; proper experiment discipline to decide what worked.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.