One interface reports 1,200 conversions. The CRM counts only 830. Finance recognises 690 net sales. An experiment estimates that 240 would probably not have occurred without advertising. Which figure is correct? Perhaps all four, because they answer different questions.
That is the crux of the issue.
1. Definition: tracking, attribution and incrementality do not measure the same reality
Advertising tracking collects events associated with delivery: impression, click, visit, consent, add to basket, lead, appointment, purchase or repeat purchase. It depends on identifiers, deduplication rules, time windows and the ability to connect systems.
Attribution assigns credit to one or more observed touchpoints. Last click, first touch, a data-driven model or a multi-touch rule distributes that credit differently, but no attribution model can reconstruct the world without the campaign on its own.
Incrementality seeks this causal difference: how many additional actions did advertising generate compared with a comparable situation without exposure or spend? It relies on control groups, geo-tests, platform experiments or other causal methods.
Financial reconciliation finally translates outcomes into recognised revenue, contribution, cash flow and observed customer value. An order recorded and then refunded is not equivalent to a net purchase. A lead without sales capacity is not equivalent to a customer.
Four truths. Four uses.
2. Key figures: abundant data, but still little confidence
- IAB surveyed more than 400 senior decision-makers from brands and agencies for its State of Data 2026 report. According to the report, 60% to 75% of users consider advanced measurement inadequate on at least one criterion among rigour, speed, confidence and effectiveness.
- In the same study, no respondent believes that all paid channels are properly represented in marketing mix models. IAB estimates that better AI-assisted practices could unlock $26.3 billion in media investment and $6.2 billion in productivity over one to two years. These are self-reported estimates, not a guarantee of individual returns.
- In 2024, 95% of advertising and data decision-makers surveyed by IAB expected signal loss or regulatory pressure to continue; 55% expected greater difficulty tracking conversions, attributing performance or measuring ROI.
- To achieve the best model quality with data-driven attribution, Google historically recommended at least 200 conversions and 2,000 interactions over thirty days for the relevant actions. The model can operate below these levels, but a small sample reduces expected stability.
- In 2026, Google states that some incrementality tests, previously associated with budgets above $100,000, may be accessible from $5,000; the company also reports up to 50% more conclusive results. These performance figures come from the platform’s internal data and must therefore be reproduced in the advertiser’s context.
The market does not lack metrics. It lacks explicit links between the question, method and decision.
3. Register 1 — Prove that delivery and collection work
Before assessing effectiveness, verify that the campaign was delivered as planned and that events describe the activity correctly. This layer may appear technical. Everything else depends on it.
3.1. Build an event dictionary
Every event needs a name, business definition, trigger, source, owner, required parameters and quality rule. “Lead” is not enough: is it a submitted form, a validated enquiry, a contactable person or an opportunity accepted by sales?
The dictionary also separates diagnostic events from objectives sent to bidding. Viewing a pricing page can help explain a journey without deserving the same value as a purchase. If the algorithm optimises both actions indiscriminately, it will rationally find the easier one.
Define a deduplication key. When a purchase is reported by the browser, server and then CRM, a shared transaction identifier prevents three outcomes being counted. The rules must cover retries, split payments, cancellations and returns.
3.2. Test the journey end to end
A test plan covers real situations: devices, browsers, consent choices, logged-in or anonymous states, payment methods, forms, calls, delayed sales and offline imports. Verify triggering, parameters, order, currency, timestamps and receipt in every destination.
Do not test only the successful path. A declined payment must not create a purchase, a double-click must not duplicate the transaction, and a consent change must propagate according to the intended configuration. Negative cases often reveal errors that artificially inflate reports.
Continuous monitoring complements testing. Zero volume, a 300% jump, an inconsistent purchase-to-payment ratio, a longer import delay or a newly missing value triggers an alert to a named owner. Without an owner, an anomaly becomes a footnote.
3.3. Respect consent and data minimisation
CNIL’s consolidated recommendation on cookies and other trackers notes, in particular, that consent must be freely given, specific, informed and unambiguous when no exemption applies, and that refusing must remain as easy as accepting. The measurement plan translates these requirements into verifiable technical states.
Collecting more is not a sound default strategy. Limit parameters to defined purposes, document retention periods, control access and avoid sending sensitive or unexpected data to advertising services. Hashing personal data does not automatically make it anonymous.
A sound outcome is explainable.
4. Register 2 — Understand what attribution distributes
Attribution answers a descriptive question: among observable, eligible interactions, which receive credit for an outcome under a given rule? This wording is deliberately cautious. It avoids asking attribution for causal proof that it cannot provide.
4.1. Document the identity of every figure
A rate or total is interpretable only alongside its click window, view-through window, time zone, conversion or interaction date, model, device scope, consent logic and treatment of refunds. Record these elements in the reporting.
Two platforms can claim the same sale because each observed an interaction within its window. Their sum must not become a company total. Use platform data to operate the channel, then a CRM or transactional reference to count unique outcomes.
Timing matters. A B2B campaign can generate a lead today that closes in four months, while a weekly dashboard judges the cohort after seven days. It is better to track cohorts by entry date and maturity than continually rewrite the past without explanation.
4.2. Choose the model according to its use
Last click is easy to reproduce and useful for some diagnostics, but generally undervalues earlier touchpoints. First touch illuminates discovery while ignoring later assistance. A data-driven model uses available patterns, but remains dependent on observed data and the conversion definition.
Do not look for a universally “true” model. Choose a stable rule for day-to-day management, run a sensitivity analysis using other windows or models, then treat the differences as a measure of uncertainty rather than a competition to elect the best dashboard.
Changing the model creates a break in the series. It should be dated, documented and, if possible, recalculated over an overlapping period. Otherwise, an apparent increase may reflect the new credit distribution rather than commercial progress.
4.3. Connect the platform with the CRM
The minimum path associates a click or campaign identifier with the session, retains the source under the consent rules, sends the information to the CRM and then reports eligible sales stages back. A mapping table protects identifiers, controls duplicates and logs rejections.
For a lead, send back genuinely discriminating statuses: qualified, opportunity, won, lost, reason and value. For a purchase, add net amount, category, new-customer status, discount and return where the use case justifies it. Bidding then learns to seek an outcome closer to the real economics.
Beware of a closed loop. If quality is defined by a score influenced by the media source and then fed back into the same system as truth, the algorithm can reinforce an initial bias. Definitions and controls must remain independent of the performance they evaluate.
5. Register 3 — Estimate what advertising actually added
An attributed sale may be incremental, brought forward, displaced from another channel or simply claimed because the buyer would have purchased anyway. An experiment attempts to isolate these alternatives.
5.1. Use a randomised trial when available
Randomly divide an eligible population between treatment and control, then compare outcome rates. If 10,000 eligible people produce 600 purchases in the treatment group versus 500 in an equally sized holdout group, absolute lift is 100 purchases and relative lift is 20%, subject to significance and protocol integrity.
Randomisation does not remove the need for operational control. Avoid contamination between groups, check the initial balance, define the event before reading results, retain the duration required by the sales cycle and do not stop the trial as soon as a favourable result appears.
5.2. Use geo-testing for aggregate investment
Comparable areas are assigned an increase, decrease or absence of investment, while a model estimates what would have happened without intervention. The method suits channels that are difficult to link individually, but suffers when there are too few areas, campaigns spill over or a local promotion affects only one group.
iROAS is calculated as incremental value divided by incremental cost. An iROAS of 2 means the experiment estimates two euros of additional revenue for one additional euro. To assess profitability, contribution, operating costs and statistical uncertainty must still be applied.
5.3. Quasi-experiments and marketing mix modelling
When randomisation is impossible, time series, instrumental variables, difference-in-differences or synthetic controls can help, provided their assumptions are stated and tested. A sophisticated method does not magically correct for a simultaneous price change or forgotten seasonality.
MMM relates aggregate changes in sales to investment, prices, promotions, distribution, seasonality, the economy and other factors. It can inform cross-channel allocation over a long period, but is not designed to choose an advert tomorrow morning. Update frequency, granularity and uncertainty intervals determine its real use.
Do not set the methods against one another. An experiment calibrates causality within a scope, MMM generalises and allocates, attribution accelerates diagnosis, and commercial data grounds the outcome. Differences become useful when each tool retains its role.
6. Register 4 — Reconcile measurement with finance
The platform optimises the value it receives. The business funds costs and earns a contribution. Between the two lie VAT, discounts, returns, fraud, commissions, cost of goods, service, payment delays and sometimes several years of customer relationship.
6.1. Start with a shared cascade
Build the following cascade: platform event, deduplicated outcome, validated order, net revenue, contribution before media, contribution after media, cash flow and observed customer value. Every transition has a rule, frequency and owner.
For lead generation, replace the order with sales stages and calculate expected value from segmented historical rates. A qualified enquiry with a 30% probability of closing and €4,000 expected contribution is theoretically worth €1,200 before other costs. This estimate must be revised using cohorts that have actually closed.
Revenue ROAS can conceal deterioration. If a campaign generates €100,000 in revenue from €20,000 in media, its ROAS is 5. With a 25% contribution margin before advertising, only €5,000 remains before creative, technology and team costs — very different economics from an activity at 70%.
Cash creates another constraint. A campaign that is profitable over twelve months can suffocate a business that pays for media immediately and collects revenue slowly. Cash payback time should therefore accompany acquisition cost and lifetime value.
Clear. But demanding.
7. Minimum architecture for a measurement plan
A usable plan fits on one map before it becomes infrastructure. It connects six components.
- Decisions: increase, reduce, maintain, move or learn.
- Questions: delivery, attributed effectiveness, causal effect and profitability.
- Metrics: one primary metric and guardrails for each decision.
- Data: platform, analytics, CRM, transactions, costs and context.
- Methods: reporting, cohorts, attribution, experiment or MMM.
- Governance: definition, access, quality, schedule and owner.
This sequence prevents technology being purchased before the decision it should improve is understood. It also reveals gaps: an advertiser may have excellent tagging but no contribution data, or an impressive incrementality study that cannot be refreshed.
8. A twelve-week implementation programme
8.1. Weeks 1 to 3 — Align the language
Inventory conversions, reports and decisions. Choose an outcome reference, define events, map flows and document windows. The deliverable is a dictionary accepted by marketing, data, sales and finance.
8.2. Weeks 4 to 6 — Repair and monitor
Test journeys, deduplicate, connect the CRM, configure consent controls and create quality alerts. An incident log retains dates, impacts and corrections so that breaks in the series remain visible.
8.3. Weeks 7 to 9 — Reconcile the economics
Reconcile conversions, net sales and contribution by cohort. Estimate delays, qualify missing data and build a marginal view. Gaps are not hidden; they become issues to resolve.
8.4. Weeks 10 to 12 — Launch a causal test
Select a decision important enough to justify a trial, verify power with a specialist, register the protocol before reading the results and plan the action associated with every possible outcome. A test without a planned decision is merely an expensive curiosity.
9. Logiks recommendations: nine controls that prevent false precision
- Never add together conversions reported by multiple platforms.
- Separate diagnostic events from the objective used for bidding.
- Display windows, the model and the reference date in the report.
- Reconcile mature cohorts rather than judging only the current week.
- Test rejection, failure, duplicates and refunds, not only the happy path.
- Maintain a financial reference independent of advertising interfaces.
- Pre-register an experiment’s hypothesis, population, duration and metric.
- Report the uncertainty interval alongside the central estimate.
- Assign every anomaly an owner and deadline.
This framework is a Logiks recommendation. It must be adapted with legal, data and finance teams to the organisation’s context. It is neither a universal interpretation of regulation nor a statistical formula that remains valid without testing its assumptions.
10. FAQ
10.1. Does server-side tracking solve signal loss?
It improves control, quality and sometimes transport resilience, but removes neither consent requirements, platform restrictions nor causal limitations. A poorly governed server-side architecture can also amplify excessive collection.
10.2. Which attribution model should you choose?
Choose the model that serves the use case, remains documented and has enough data. A stable model is often sufficient for operations; for allocation, complement it with cohorts, experiments and economics.
10.3. Why do the CRM and Google Ads report different totals?
Windows, time zones, click or conversion dates, consent, duplicates, import rejections, cancellations and scopes differ. Create a reconciliation table instead of forcing an artificial equality.
10.4. How long does it take to measure incrementality?
It depends on the conversion rate, expected effect, sales cycle, traffic and protocol. A power analysis should precede launch; some experiments take several weeks and may still remain inconclusive.
10.5. Is a view-through conversion useless?
No. It describes an observable sequence within a window, but its causality is more uncertain. Analyse its sensitivity and use an experiment when its weight strongly influences the budget.
10.6. Can performance be managed without third-party cookies?
Yes, by combining consented first-party data, aggregate measurement, CRM, experiments, MMM and platform data. This provides less exhaustive individual tracking, but potentially better decisions.
11. Conclusion
A good measurement system does not manufacture a single figure to close the discussion. It organises different forms of evidence, states their limitations and connects each one to a decision.
Begin by making delivery and events reliable. Then describe the credit assigned. Measure the causal difference when the value at stake justifies it. Finally, reconcile everything with contribution and cash flow.
Useful precision is honest.
12. Main sources
- IAB, State of Data 2026: The AI-Powered Measurement Transformation, 2 February 2026: https://www.iab.com/insights/2026-state-of-data-report/
- IAB, announcement Project Eidos, sample and main findings from the 2026 report: https://www.iab.com/news/iab-announces-project-eidos/
- IAB, State of Data 2024, signal loss and measurement challenges: https://www.iab.com/news/iab-state-of-data-report-2024/
- Google Ads Help, About data-driven attribution: https://support.google.com/google-ads/answer/6394265
- Google Ads Help, About attribution models: https://support.google.com/google-ads/answer/6259715
- Google Ads Help, Conversion Lift based on geography and definition of iROAS: https://support.google.com/google-ads/answer/14102986
- Google Ads Help, 2026 incrementality update: https://support.google.com/google-ads/answer/16719772
- CNIL, Cookie and other tracker recommendation — consolidated version, January 2026: https://www.cnil.fr/sites/default/files/2026-01/recommandation_cookies_consolidee.pdf