Attribution and real lift: did advertising bring new customers?

An advertising report can claim a hundred sales even when some buyers would have purchased without the advertisement. Someone already knew the business, returned to its website and encountered an ad shortly before ordering. Attribution assigns credit to a contact. An incrementality study asks how many additional customers a specific advertising change produced compared with what would otherwise have happened.

This guide addresses advertisers evaluating activity in the Russian market. That geographic scope remains unchanged in translation; the illustrative budgets use independent USD assumptions, not converted Russian prices. It does not explain how to launch an overseas advertising account in Russia. Primary sources were checked on 22 October 2026. The workflow, thresholds and both fictional scenarios are editorial teaching materials, not customer case studies or platform requirements.

Separate three different measurement jobs

Contractual accounting establishes whether a partner delivered the agreed outcome. Journey analysis describes contacts preceding a purchase. Causal measurement estimates what changed because of an intervention. A small estimated lift does not retrospectively invalidate a partner's properly accepted conversion. Pay according to the applicable agreement and negotiate any revised conditions for future activity.

Metrica documentation records the retirement of several older attribution models from 20 May 2026. Preserve the actual model name and export date; an old screenshot is insufficient.

Perfectly reconciled attribution reports still do not provide a control group. Comparing people who clicked with people who did not usually compares different levels of purchase intent. A rise after launch might reflect seasonality, product availability or a discount. Start with the precise intervention and a credible comparison. If the comparison cannot be established, describe an association rather than a proven advertising effect.

For a suitable partner channel, our AffiliateCore platform is an option for preparing your own offer. Register as an AffiliateCore advertiser and prepare the product description, landing page and outcome conditions. Product, country and traffic eligibility require separate checks. Audience assignment and lift analysis in this guide belong in your research process; we are not presenting them as built-in AffiliateCore features.

Write the study brief before spending

Use one testable sentence: “Adding the agreed advertising contact increases the proportion of new customers with a retained paid purchase.” Define a new customer as someone without an earlier completed purchase in the business records available to you. If records cover only one year, state that limitation. Several enquiries from the same person must not create several new customers in the primary outcome.

Record the assignment unit, intervention, primary metric, recruitment period, purchase maturation period, cancellation cutoff, budget cap and smallest commercially useful effect. Name the decision owner. Choose one primary outcome. Enquiries, complaints, cancellations and service workload help interpret it, but should not replace an unsuccessful primary metric after the results become visible.

You need owners for creative materials, assignment, sales records and finance. The analyst needs appropriately restricted access to outcomes in both groups, including purchases through direct visits or another channel. Individual assignment is feasible only where the business can genuinely control delivery and exclusion. Do not assume any advertising account offers a suitable holdout or that uploading a customer list is automatically permissible.

Ask an analyst to estimate sample size using the ordinary purchase rate, variability and minimum useful change. There is no universal one-week experiment. Rare purchases require more observations; long sales cycles require more time. When the available sample is inadequate, the project can test operational readiness, but it cannot promise a decisive estimate of commercial lift.

Prepare three practical documents

Intervention card. Include material version, audience, promise, destination, dates, exclusions and owner. A fictional service advertisement could say: “Scheduled home equipment maintenance. Review the work included and available dates; confirm the price before ordering.” The landing page must preserve those conditions. Avoid false urgency, hidden charges and unsupported savings claims. Keep an exact copy of the material used during the study.

Observation register. Create one row per assignment unit with an internal pseudonym, group, inclusion date, prior-customer flag, first purchase, cancellation, final status and contribution margin. Contact details remain in the protected customer system, not public worksheets or link parameters. A missing record is not automatically a non-purchase. First establish whether outcome capture works equally well in both arms.

Decision card. Record planned and actual sample sizes, initially accepted purchases, cancellations, retained results, full costs, estimated effect, uncertainty, protocol deviations and next action. Save the original plan before launch. Separately record changes to prices, stock, landing pages, sales handling and other campaigns so that the team can identify changes that compromise the comparison.

Readiness means that another analyst can trace an assigned unit through to a final commercial outcome. Rehearse duplicate submissions, cancellation and status correction using clearly marked noncommercial records excluded from results. Do not manufacture paid conversions or genuine purchases merely to make tracking look successful. If the rehearsal reveals missing outcomes, repair measurement before testing marketing.

Scenario one: an additional contact with prospective customers

A fictional maintenance business has 20,000 eligible prospects without a prior purchase. Assume the contact is permitted and exclusion is technically reliable; these are scenario prerequisites, not claims about any particular platform. Before delivery, random assignment places 10,000 people in treatment and 10,000 in control. Normal service and other conditions remain the same in both groups.

Prepare the intervention card above, one landing page, a handling script and the observation register. The script says: establish the need, confirm the work and appointment, and do not promise unavailable services. Treatment receives the additional contact; control does not. Analyse people in their original assignment regardless of whether they click or actually see the material. This estimates the effect of assignment, rather than selecting only enthusiastic responders.

Recruitment and delivery run for four weeks. Freeze outcomes after another thirty days of follow-up for the last included record. These illustrative periods must be adapted to a real sales cycle. Count every eligible new purchase in each arm, including purchases attributed elsewhere. Use only a permitted, verified identity link across devices; disclose unresolved identity limitations instead of silently assuming complete matching.

Treatment initially records 330 accepted purchases and later 30 cancellations, leaving 300 retained new customers. Control records 220 accepted purchases and 20 cancellations, leaving 200. The retained purchase rates are 300 / 10,000 = 3% and 200 / 10,000 = 2%. The difference is one percentage point. Estimated additional customers in the treatment population are 10,000 × (3% − 2%) = 100; relative lift is 100 / 200 = 50%.

Assumed additional costs are USD 600 for delivery, USD 100 for creative production, USD 100 for setup and analysis, and USD 100 for extra enquiry handling: USD 900 altogether. These are independent teaching assumptions, not supplier prices or exchange-rate equivalents. There is no separate affiliate commission in this scenario. If your arrangement includes one, add it explicitly rather than hiding it outside the budget.

Full cost per additional customer is USD 900 / 100 = USD 9. Assume USD 12 contribution per retained purchase after variable fulfilment, payment and accounted return costs, but before the listed experiment costs. Additional contribution is 100 × USD 12 = USD 1,200. Subtracting USD 900 leaves USD 300. This is not company net profit: common fixed costs have not been allocated.

Suppose the attribution report credited the campaign with 240 retained purchases. Using the same full cost gives USD 900 / 240 = USD 3.75 per credited purchase. That is a different denominator and a different answer. The difference does not identify 140 specific people who were “unnecessary”; the study estimates a population effect, not each person's unobservable alternative behaviour.

Now account for uncertainty. Under independent binary outcomes, the approximate standard error of the rate difference is √(0.03 × 0.97 / 10,000 + 0.02 × 0.98 / 10,000) ≈ 0.002207. A simple approximate 95% interval uses the difference plus or minus 1.96 standard errors: about 0.567–1.433 percentage points, corresponding to 57–143 additional customers. This is an illustrative frequentist calculation, not a vendor report or a probability that the campaign is profitable.

Break-even requires USD 900 / USD 12 = 75 additional customers. The lower interval endpoint is below 75. A positive estimated effect therefore does not establish sufficiently secure unit economics. Under a predeclared rule requiring the lower endpoint to exceed break-even, this version is not ready to scale. Plan a separate adequately powered confirmation study instead of extending the experiment until a favourable result appears.

Stop delivery immediately if control exclusion fails, the promise is wrong or outcome collection breaks. Make the commercial decision at the planned time after cancellations mature. If a subsequent independent study meets the economic and service-quality criteria, expand gradually while maintaining a comparison group. A broader audience or larger spend changes the conditions; the previous lift is not a promise of constant returns.

Scenario two: regional placements for an office service

A fictional office-maintenance provider cannot reliably exclude advertising at individual level. It chooses two comparable sets of regions for a limited additional placement. This is an observational comparison of changes, not a randomised causal demonstration. Inspect several weeks of trends, customer mix, seasonality and delivery capacity beforehand. Similar totals in one previous month do not establish comparable underlying trends.

Prepare a regional information card: “Scheduled office maintenance: included work and quotation process.” Use the same enquiry brief and qualification script across regions. Collect the premises type, required work, region, desired timing and preferred response method. Staff apply identical acceptance criteria and do not prioritise records based on treatment status. Verify contractual conditions and permission for the actual placement before spending.

Use eight weeks of history, four weeks of placement and another forty-five days for the last deals to mature. For simplicity, the figures below are normalised to equal observation durations; retain the original weekly regional series for analysis. Before intervention, treatment regions have 200 retained new customers and control regions have 160. These are completed outcomes, not unqualified enquiries.

After intervention, treatment records 270 accepted purchases minus 20 cancellations, leaving 250. Control records 190 minus 14, leaving 176. Control growth is 176 / 160 − 1 = 10%. If treatment would have followed the same proportional background growth, its counterfactual is 200 × 1.10 = 220. Estimated additional customers are 250 − 220 = 30, rather than the entire observed increase of 50.

Additional costs are USD 360 for placement, USD 80 for material adaptation, USD 80 for analysis and USD 80 for extra handling: USD 600 in total. Assume USD 15 contribution per retained customer before those costs. Cost per additional customer is USD 20; incremental contribution is USD 450; the result after expenditure is minus USD 150. Break-even requires 40 additional customers.

Test the background assumption without calling the exercise a confidence interval. At 5% background growth, the expected total is 210, the estimated increment is 40 and the economic result is zero. At 15%, the expectation is 230, the increment is 20 and the result is minus USD 300. Four aggregate numbers cannot honestly supply the sampling uncertainty of a regional experiment.

Do not scale this version. Finish handling existing enquiries, investigate the offer, material costs and regional differences, and prepare a fresh protocol after corrections. A stronger future design may use several non-overlapping matched regions with random assignment within pairs, where technically feasible. Statistical analysis then respects regions as the assignment units instead of pretending that every purchase is an independent randomised observation.

If a partner channel fits, sign in to AffiliateCore as an advertiser and use the prepared description and landing page for your offer draft. Maintain the external study register separately. A directly operated partner programme is another option where you can manage agreements and accounting. Choose according to control and cost, without assuming either route automatically produces incremental sales.

Common errors that change the answer

Do not remove cancellations from treatment while leaving them in control. Do not compare mature purchases in one group with fresh leads in the other. Do not redefine a new customer after seeing results. The budgets here contain additional costs relative to normal operation, so identical baseline work is not counted twice. Any claimed savings in other channels need their own evidence before being deducted.

Travel across regions, shared recommendations and overlapping campaigns can contaminate the comparison. Document these problems rather than silently deleting inconvenient records. Assess whether the estimate remains usable and redesign if contamination is material. When the estimated increment is zero or negative, do not advertise a seductive negative acquisition cost: state that positive additional customers have not been established.

Your assignment and source notes

Create one study brief, one material card and an empty outcome register. Confirm that both groups can be measured using the same rules. Ask the analyst to calculate a realistic sample requirement and finance to approve contribution margin and the full budget. The output should be a documented decision to start, repair readiness or acknowledge inadequate data, rather than a request to spend more and hope measurement follows.

Primary methodological references: Google: incremental versus attributed conversions; Google: Conversion Lift availability varies by account; Vaver and Koehler, 2011: randomised regional experiments. These references do not establish campaign availability in Russia.

Related work: audit the path from advertisement to CRM and review mature cohorts before scaling. The specific output here is an estimate of additional customers with an explicit comparison, complete costs and honest limits.