Syntha Lab RU

Measuring marketing: path attribution, advertising carry-over, the budget optimum, the price test, NPS and the referral reward

The document explains how to check that marketing money works: how attribution differs from causality, how to take the effect of advertising on the following weeks into account, how to allocate a budget between channels by marginal return, how to decide a price test by profit per visitor and how to tell a significant change in NPS from noise. The decision it supports: where to move the budget and which numbers to trust.

Commerce and buying Brand and retail Verified: 2026-11-17

Open interactively →

Why this document

Marketing is easy to confirm with any report: every channel has a model in which it is the best. The document sets a discipline: first see how much the assessment depends on the model, then check with an experiment, and allocate money by marginal return and not by the average. It is read by the marketer, the analyst, the finance director and the head of customer experience. There are no common norms of return, all parameters are taken from your data and updated every quarter.

At a glance

What is this document and why is it needed?
The document explains how to check that marketing money works: how attribution differs from causality, how to take the effect of advertising on the following weeks into account, how to allocate a budget between channels by marginal return, how to decide a price test by profit per visitor and how to tell a significant change in NPS from noise. The decision it supports: where to move the budget and which numbers to trust.
Who owns the document and how often is it rebuilt?
Brand and retail: the analyst gathers data and calculations, the marketer proposes the budget, the finance director approves; recalculate parameters every quarter.
What mistakes are the most common?
Taking attribution for causality → a channel may not have influenced the order → check with a test against a control. Comparing channels over different observation windows → assessments are not comparable → one window. Taking a short series for carry-over → λ is random → a series of at least 26 weeks cleaned of seasonality and price.

For whom

Brand
A brand sees more channels and a longer effect: the carry-over of advertising and the budget optimum matter more than for a shop.
Retail
Retail is closer to the till: a price test and attribution on its own site give a quick answer, and look at the influence of advertising on shops through a control.

What it consists of

Attribution divides credit but does not prove causality

  • A table of paths gives the sequence of channels before the order and the order value. Models divide the value by different rules: last click gives all to the last touch, first click to the first, linear equally, position-based 40/20/40 gives 40% to the first and last and splits 20% among the middle ones, with decay the weight of a touch equals 0.5 to the power “age ÷ half-life”.
  • The spread of models for a channel shows how much the assessment depends on the rule. If a channel wins in all models the assessment is reliable across rules; if it jumps, a decision without an experiment is unreliable. But even a consistent attribution does not prove that the channel produced the order.
  • After restrictions on ad systems the data is incomplete: some touches are outside the digital trace. Compare channels over the same observation windows and do not cut the “first touch” budget by one model.

Advertising carry-over and half-life

  • The effect of a week = the week’s spend + carry-over share × the effect of the previous week. Half-life = ln 0.5 ÷ ln λ, the full effect of a rouble = 1 ÷ (1 − λ). For TV, outdoor advertising and bloggers the carry-over is noticeable, for search advertising almost zero.
  • Fitting the carry-over share needs a sales series of at least 26 weeks cleaned of seasonality, price and promotions; the calculation searches values from 0 to 0.9 and takes the one at which the regression of sales on the accumulated effect explains most. A short or uncleaned series gives a random value.

The budget optimum by marginal return

  • The response of a channel is described by the curve a × (1 − e^(−spend ÷ b)): a is the revenue limit, b the spend scale. Marginal return = margin × (a ÷ b) × e^(−spend ÷ b) shows how much margin the next rouble gives. The budget goes in steps to the channel with the highest marginal return while it is above one.
  • Money whose marginal return is below one is not spent: it stays in the company. The parameters a and b are a hypothesis from fitting past data; update them every quarter and do not extend the optimum to spend outside the observed range. The calculation does not see the links between channels (brand and organic).

The price test by profit per visitor

  • Profit per visitor = conversion × (price ÷ (1 + VAT) − cost). The expensive variant almost always gets fewer orders but may earn more, so the decision is made by profit and not by the number of orders. Significance is counted separately for conversion and for profit: z for profit = difference ÷ √(SE_A² + SE_B²).
  • A test shorter than a week does not average out weekday fluctuations, with fewer than a hundred orders in a variant the estimates are shaky. Test price on your own site: platforms restrict A/B tests of price.

NPS with an interval and the reward for a friend

  • NPS = promoters − detractors; spread σ² = (P + D − (P − D)²) ÷ n, interval NPS ± z × σ. A change is significant if |z| = |NPS₂ − NPS₁| ÷ √(σ₁² + σ₂²) is not below the quantile of the confidence level. At small samples the interval is wide, and a rise of a few points is within the noise.
  • The referral reward is paid to all legitimate referrals while the benefit comes from incremental ones: break-even reward = incremental × lifetime margin ÷ legitimate. Pay after the first paid purchase; the tax treatment of payouts to individuals is checked by a specialist.

Where the data comes from

Data sourceFields
Analytics and CRM systemscustomer paths with channels and the order value, spend by channel and week, sales by week
Response curve fittingchannel parameters a and b, margin on revenue, minimum spends
The site and surveysvisitors and orders by price variant, shares of promoters and detractors, number of responses

How to build it step by step

  1. Step 1: Gather customer paths and count the value by channel in five models.
  2. Step 2: Find channels whose assessment jumps between models and set an experiment for them.
  3. Step 3: For channels with a long effect count the accumulated effect and the half-life.
  4. Step 4: Fit the response curves with the accumulated effect and get the parameters a and b.
  5. Step 5: Allocate the budget in steps by marginal return and find how much does not pay back.
  6. Step 6: Check the price with a test of at least a week and decide by profit per visitor.
  7. Step 7: Report NPS with the interval and the number of responses, announce only a significant change.
  8. Step 8: Update the parameters every quarter and compare plan with actual.

Formulas

MeasureFormulasNote
Marginal return of a channelmargin × (a ÷ b) × e^(−spend ÷ b)Stop when below 1.
Accumulated advertising effectA_t = x_t + λ × A_(t−1); half-life = ln 0.5 ÷ ln λThe full effect of a rouble = 1 ÷ (1 − λ).
Profit per visitorconversion × (price ÷ (1 + VAT) − cost)Decide by profit and its z.

Formulas in notation

Accumulated effect
A_t = x_t + \lambda A_{t-1}
x: the week’s spend, λ: the carry-over share.

Worked example

Illustrative data: a RUB 20 million budget, three channels; a price test of RUB 3,990 against RUB 4,490

MetricValue
Spent at the optimum, m RUB6.00
Profit at the optimum, m RUB1.06
Profit at an equal split, m RUB−2.11
Profit per visitor A, RUB47.06
Profit per visitor B, RUB52.65
z for profit1.96

The undistributed RUB 14 million stay in the company: a rouble in them does not pay back. z for profit is at the significance bound 1.96. All numbers are illustrative.

Figures

Budget by channel: the optimum and an equal split
Bar chart: Search 3.25, Social 1.75, Marketplace 102468SearchSearchOptimum: 3.253.25Equal split: 6.676.67SocialSocialOptimum: 1.751.75Equal split: 6.676.67MarketplaceMarketplaceOptimum: 11Equal split: 6.676.67m RUB
  • Optimum
  • Equal split
Data as a table
MeasureOptimumEqual split
Search3.256.67
Social1.756.67
Marketplace16.67

Budget by channel: the optimum and an equal split. The optimum spends less and only where a rouble pays back.

Spend and accumulated effect by week
Line chart: Spend of the week: 100 → 0; Accumulated effect: 100 → 15.50204060801001234567Spend of the week, 1: 100Spend of the week, 2: 0Spend of the week, 3: 0Spend of the week, 4: 50Spend of the week, 5: 0Spend of the week, 6: 0Spend of the week, 7: 0Accumulated effect, 1: 100Accumulated effect, 2: 60Accumulated effect, 3: 36Accumulated effect, 4: 71.6Accumulated effect, 5: 43Accumulated effect, 6: 25.8Accumulated effect, 7: 15.5conventional units
  • Spend of the week
  • Accumulated effect
Data as a table
MeasureSpend of the weekAccumulated effect
1100100
2060
3036
45071.6
5043
6025.8
7015.5

Spend and accumulated effect by week. The effect fades gradually and overlaps with the next burst of advertising.

Specifics by category

Apparel
Apparel: many sizes and seasonality, so count redemption, purchase gaps and day indices by season.
Footwear
Footwear: redemption is lower because of fitting, and reverse logistics noticeably changes the cost of a redeemed unit.
Accessories
Accessories: the check is small, pairs with apparel give a lift; the delivery threshold and bundles work stronger.
Kids
Kids: frequent repurchases by height, gaps between purchases are shorter, the lapsed threshold is lower.
Men
Men: a steady core range, less frequent purchases, a longer lapsed threshold.
Women
Women: more novelties and seasonal peaks; count day indices and pair lift separately by season.

Common mistakes

  • Taking attribution for causality → a channel may not have influenced the order → check with a test against a control.
  • Comparing channels over different observation windows → assessments are not comparable → one window.
  • Taking a short series for carry-over → λ is random → a series of at least 26 weeks cleaned of seasonality and price.
  • Using the parameters of one period → the optimum goes stale → update every quarter.
  • Deciding a price test by the number of orders → the expensive variant is underrated → decide by profit per visitor.
  • Comparing NPS without an interval → noise is taken for growth → count z and the interval.

Russian specifics

  • Advertising is subject to marking, customer data is personal: use aggregates and consent. Payouts to individuals in a referral program are checked by a tax specialist.

Open interactively →