Attribution allocates recorded credit under a model. It does not prove that the credited interaction caused the sale.
Google Ads documentation recommends at least 200 conversions and 2,000 ad interactions within 30 days for data-driven attribution to identify patterns more precisely. Google also says the model still functions with less data. That threshold describes model input volume inside one platform. It does not convert observed paths into a controlled causal experiment.
The honest use of attribution begins by separating three questions. What did the system observe? How did the model assign credit? What would have happened without the advertising?
Only the third is a causal question.
The path is already filtered
An attribution report does not contain every influence on a buyer. It contains events the measurement system could connect under its identity, consent, lookback and channel rules.
It may miss exposure on another device, an untagged referral, offline conversation, organic research, a blocked identifier, activity outside the lookback window, or a purchase connected to a different account.
The recorded path is useful. It is not the whole path. Before interpreting channel credit, document which events are eligible, which channels can receive credit and which identities can be joined.
Every model makes a different claim
Last click gives all eligible credit to the last credited interaction. Data-driven attribution distributes credit using patterns in converting and non-converting paths. Neither model changes the underlying sale.
Google Analytics explains that its data-driven model estimates how adding each interaction changes the probability of a key event, using a counterfactual approach within available path data. It also states that conversions can be reattributed for up to seven days after the event.
That means a report can change after the business outcome occurred, because the allocation model received more information.
Use model comparison to understand sensitivity. If a channel appears essential under one model and negligible under another, the correct conclusion is not that one report has revealed truth. It is that allocation depends heavily on assumptions.
Attribution is a bookkeeping rule applied to observed paths. Incrementality asks whether the outcome changed.
Platform attribution has a boundary
A platform can model interactions it sees with more detail than an independent analytics tool. It also has an incentive to demonstrate value. This does not make platform reporting useless. It means the boundary must be named.
For every reported number, record the platform, attribution window, click and view treatment, conversion definition, time zone, currency, model, consent and identity conditions, and date extracted.
Do not place Google Ads, Meta and analytics totals in one table as if they were mutually exclusive sales. The same order can receive credit in several systems. Reconcile against the commercial source of truth, such as payment records or a qualified CRM outcome.
Allocation and proof serve different decisions
Attribution can help allocate budget inside a stable operating system. If a campaign repeatedly appears in valuable paths and the business needs a practical bidding signal, modelled credit may be sufficient.
It is not sufficient for every decision. Use stronger evidence when asking whether the campaign created additional sales, whether brand activity changed later demand, whether retargeting captured buyers who would have purchased anyway, or whether one region improved because of media rather than another change.
These questions require a comparison with a credible counterfactual.
Build experiments around decisions
A randomised holdout can compare eligible people exposed to treatment with a control group not exposed. Geo experiments can compare matched regions when user-level randomisation is unavailable or inappropriate.
Google published the open-source CausalImpact methodology for estimating the effect of an intervention using Bayesian structural time-series models. The method depends on a valid control series that was not affected by the intervention. A sophisticated model cannot repair a contaminated control.
Meta’s Conversions API documentation recommends lift studies for understanding the impact of Meta ads on in-store purchases. Eligibility, design and available measurement products change, so confirm current account access before designing the measurement plan.
The experiment should begin with the decision it will change. A lift estimate that arrives after the next six budgets are committed is research without an operating route.
Read uncertainty, not just the point estimate
An estimated lift of 8 percent is incomplete without uncertainty and test design.
Ask what the confidence or credible interval was, whether the test was powered for the effect the business cared about, whether groups were balanced before launch, whether another campaign spilled into control, whether pricing, stock or site performance changed, and whether the analysis was specified before results were seen.
A result compatible with both meaningful gain and no effect should not be presented as proof of gain.
Do not extend a test until it becomes significant without correcting the method. Repeated checking changes the false-positive risk. The same sampling discipline applies to creative testing, covered in building a reliable workflow.
Create an attribution receipt
Attach a compact receipt to every performance report.
Observed. Commercial outcomes, analytics events and platform events.
Allocated. Credit assigned under each named model and window.
Estimated. Modelled conversions or behaviour where direct observation is unavailable.
Tested. Incremental effect from a defined experiment, with uncertainty.
Unknown. Material influences the system cannot observe or isolate.
This stops one column labelled “conversions” from carrying five different meanings.
Do not optimise a proxy without a value map
Automated bidding learns from the conversion actions supplied to it. If a low-value form completion and a qualified sale are treated equally, the system is being asked to maximise the wrong unit.
Map the event, its business meaning, validation, value, delay, and cancellation or refund treatment. Import later commercial outcomes where the platform and privacy basis allow it, and keep the raw commercial record outside the advertising platform.
Attribution quality cannot exceed conversion-definition quality.
Write the decision before opening the report
Attribution becomes less misleading when the team states the decision in advance. “Should we move ten percent of prospecting spend from market A to market B?” can be examined. “Which channel caused growth?” invites a story after the numbers arrive.
For each review, record the action under consideration, the metric, the comparison period, known tracking changes and the evidence that would reverse the decision. Then show attributed conversions beside total orders, spend, consent rate and any experiment result. If the platform report rises while the business total does not, the discrepancy is the finding.
Keep that statement beside the chart, not in a methodology note nobody opens.
How we handle this
Scale uses attribution for allocation while labelling its model, window and observation boundary. Causal claims require a test design that can support them.
Origin helps identify brand effects that short lookback windows can miss, without inventing a direct-response number. Merch reconciles platform credit against actual orders, refunds, fees and fulfilled revenue.
Send us one report and the decision it is supposed to support. We will separate the observed, allocated, estimated and unknown before recommending another dashboard.
Sources
Google Ads, About attribution models.
Google Ads, data-driven attribution requirements and methodology.
Google Analytics, get started with attribution.
Google Research, Inferring causal impact using Bayesian structural time-series models.
Meta Business Help, About Conversions API.
Sources and product documentation checked 30 July 2026.
