Paid social attribution methods compared: what to ask an agency about each

Credit assigned across recorded interactions

Rule-based attribution allocates credit to eligible interactions under a named model and lookback window. It describes observed paths for a defined population, not the outcome that would have occurred without advertising. Changing the rule can shift reported value while the underlying customer events remain unchanged.

Record the conversion action, creditable channels, identity rule, model, window, version and date of change. Models may include last-click, first-click, linear, position-based, time-decay or data-driven approaches. Examples include Meta Ads Manager's 7-day click and 1-day view default, and a click window of up to 90 days in Google Ads.

Google's attribution page describes the setting as determining how credit is allocated to ad interactions and affecting conversion reporting. Treat attributed value as credit rather than incremental revenue: credit-based ROAS is not iROAS, and neither measure is MER. Keep the declared model and window beside key metrics.

Compare paths only under the same declared rule; do not present one model as a universal account of value. Review methodology can connect conclusions to supporting evidence while keeping them distinct from an impact estimate. This makes each comparison specific to its model and window, rather than making unlike totals appear interchangeable.

What an advertising platform records

Platform observation reports what an advertising product records under a particular account configuration. Its count or rate can inform campaign operations or a data-quality investigation, but the output alone does not establish that the campaign caused the recorded business outcome.

Before comparing extracts, name the event, eligible population, numerator, denominator or unit, account, period, timezone, export, query version, inclusions and exclusions. Meta Ads Manager, Google Ads, TikTok Ads Manager, LinkedIn Campaign Manager, Snapchat Ads Manager and Pinterest Ads Manager each provide attributed-conversion columns. Similar labels do not guarantee equivalent measures.

Google's conversion-data guidance says its Conversions column is shaped by primary actions and can include modelled events. Supplier definitions, missingness, modelling and reporting delay therefore affect interpretation. A platform analyst may use the output for controlled campaign management, but its account and time basis need to accompany the readout.

The quality checklist links supplier totals to operational checks rather than treating them as independently verified impact. A platform count says what its configured product recorded; the outcome itself does not show what would have happened without the campaign during the stated period.

Change measured against an alternative

Counterfactual impact design estimates change against a credible alternative condition, rather than assigning credit among recorded interactions. It can support a bounded impact claim when the comparison is appropriate, assumptions are stated and uncertainty is reported. A method's label alone does not make its result causal.

Name the design before launch: a randomised controlled trial, geo holdout or geo experiment, difference-in-differences, synthetic control, matched control or switchback test, or a platform lift product such as Meta Conversion Lift or Google Conversion Lift. The qualified evaluator owns the work and reports an appropriately bounded estimate.

The evaluator should identify the population, outcome and analysis period in advance, then report the limits of the estimate. Evaluation standards for government provide a public reference point; quality assurance of analysis is also relevant when checking how analysis is presented. Neither validates a commercial design or supplies an England benchmark.

Use strategy framework when the decision needs impact evidence and a feasible causal design exists. Preserve the other evidence layers when they answer useful operational questions, but do not combine their outputs into a causal claim. Agree the evidence layer and test design with the client before launch.

Matching agency and client records

Internal operational linkage identifies which platform and client records satisfy declared identity and timing rules. Show matches and unmatched categories in both directions because the denominators differ. This is a reconciliation result, not a causal estimate, and the matched set should not automatically represent every record.

Freeze both extracts and define the accepted business event, match fields, conflict rules, deduplication, period and timezone before joining them. Keep cancellations, refunds, missing identifiers and unresolved records visible. Eligible records matter because linkage loss and selection shape the population represented by the matched set.

Privacy and PECR checks sit alongside the attribution model, not inside it. A successful join does not itself supply permission to collect or use data, and matching logic does not replace a separate privacy review. Keep the record set labelled as the population actually assessed, including its unmatched categories.

Use measurement and reporting to set out match rules, source extracts, unmatched categories and outcome checks for a client. Keep the result labelled as operational reconciliation, not campaign lift. Any later impact claim needs a separate comparison design and declared assumptions.

What to ask an agency about its attribution method

The four methods above are what an agency can offer you. Asking which one it uses tells you how far its claims can go, and the answers are easy to compare across agencies.

Be wary of a single headline ROAS with no declared model. Put each answer in the same form for every agency you compare, and read it next to the agency's reference-call answers and the review gates.