05.19 Omnichannel E-Commerce Multi-Touch Attribution and BI Analytics Methodology
Governance metadata | Owner: Sales Lead | Last reviewed: 2026-10-22 | Next review: 2027-01-31
Tool positioning: Unifies the multi-touch attribution (Attribution) logic and the business intelligence (BI) dashboard system for omnichannel smart-hardware marketing, solving the classic predicament of "I know half of my ad spend is wasted, but I do not know which half." By introducing a Shapley marginal-contribution model based on cooperative game theory, it breaks the attribution bias of single last-click reporting and supports sound decisions on supply-demand coordination, traffic budget allocation, and single-product unit economics.
1. Cross-Evaluation of Mainstream Multi-Touch Attribution Models
Smart hardware carries a high average order value (typically USD 100 to 1,000+), and users usually travel a long decision path from first interest to final paid order — "YouTube review seeding ➔ Facebook feed re-engagement ➔ Google brand-term search ➔ Amazon price comparison ➔ DTC-site abandoned-cart recovery EDM ➔ final order" (typically 7 to 28 days).
| Attribution Model | Core Algorithm Logic | Typical Use Case | Core Defects and Misleading Risks |
|---|---|---|---|
| Last Click | 100% of the conversion credit goes to the last touchpoint before payment | Simple broad-blast media, search bidding | Severely over-credits brand terms and retargeting ads, starving upstream YouTube/KOL seeding budget |
| First Click | 100% of the conversion credit goes to the first touchpoint that introduced the visitor | Brand new-product cold start, pure awareness/buzz evaluation | Severely ignores mid- and lower-funnel order-driving conversion and abandons the key value of cart-recovery email |
| Linear | Splits the conversion credit equally by touch count across all touchpoints in the path | Rough sizing in the early stage of touchpoint coordination | Cannot distinguish the critical value of the "closing touch" versus the "ice-breaking seeding" touch |
| Time Decay | The weight assigned to a touchpoint increases exponentially the closer it sits to the conversion timestamp | Promo sprint period (Black Friday / mega-promo) | Still systematically undervalues the awareness-building role of high-value in-depth review videos from two weeks earlier |
| Positional (U-shaped / W-shaped) | The first and last touchpoints take 40% each, with the remaining 20% split evenly across the middle transition touchpoints | Reasonably balanced multi-channel media teams | Weight allocation rests on experiential human assumptions, with no rigorous statistical or game-theoretic proof |
| Shapley Marginal Contribution (Data-Driven) | Based on cooperative game theory, computes each channel's marginal increment across all possible coalitions | Omnichannel multi-channel blended media, budget reallocation | Depends on sufficient conversion-path sample data (typically ≥ 500 monthly orders) |
2. Hands-On: Game-Theoretic Shapley Marginal-Contribution Multi-Touch Attribution
1. Mathematical Logic and Marginal-Contribution Measurement
The Shapley Value originates in Lloyd Shapley's cooperative game theory (Nobel Memorial Prize in Economic Sciences). Treat every channel that participates in a conversion in omnichannel marketing as a "player (Players)" in the game, and the generation of the conversion as the "coalition value (Coalition Value)".
The Shapley attribution weight φ_i of a channel i equals the weighted average of the marginal conversion increment (Marginal Contribution) that this channel brings across every possible subset combination of channels:
Core Shapley attribution weight formula:
φ_i = Σ [|S|! × (|N| - |S| - 1)! ÷ |N|!] × [v(S ∪ {i}) - v(S)]
Where:
Nis the set of all touchpoints participating in the omnichannel journey (for example:N = {YouTube seeding, Meta paid media, Google search});Sis any subset of touchpoints that does not include channeli;v(S)is the historically observed conversion rate or order volume of touchpoint combinationSwhen it exists on its own;v(S ∪ {i}) - v(S)is the pure marginal gain brought by introducing channeli.
2. Projection Example: 3-Touchpoint Sample for the AstroNext USD 399 Smart-Hardware Product
⚠️ Projection notice: this section is a teaching projection of the Shapley model. The conversion rates and the distribution of 1,000 users are self-consistent assumptions and do not represent measured data from any real brand or platform. The model itself can be applied directly; replace the figures with your own backend data.
Assume we select the conversion-path data of 1,000 final purchasers, covering three core channels: channel A (YouTube tech reviewers), channel B (Meta social ads), and channel C (Google brand search):
| Touchpoint Combination (Coalition S) | Users Covered by the Combination | Standalone Conversion Rate v(S) | Path and Business Context |
|---|---|---|---|
| ∅ (no prior touchpoints) | 0 | 0.0% | Natural un-intervened baseline |
| {A} (YouTube only) | 3,000 | 1.2% | Pure video seeding, no active search behaviour |
| {B} (Meta ads only) | 5,000 | 0.8% | Passive social feed exposure, no deep mindshare formed |
| {C} (Google brand-term search only) | 1,000 | 2.5% | Native brand-term search (existing customers or word of mouth) |
| {A, B} (YouTube + Meta) | 1,500 | 2.8% | After watching a review video, re-targeted by a Meta feed ad |
| {A, C} (YouTube + Google) | 800 | 4.6% | Actively searches the brand term on Google after an in-depth review video |
| {B, C} (Meta + Google) | 1,200 | 3.5% | Searches the brand term after an ad sparked interest |
| {A, B, C} (full 3-touchpoint path) | 600 | 6.0% | Completed the full seeding, re-engagement, and search-closing loop |
After the Shapley permutation-and-combination calculation, the projected weights of the three channels in driving a single conversion are (this dataset is a sandbox projection, not a real measured report; basis: 00.04 Source Ledger item #5):
- Shapley weight of YouTube seeding (channel A): 31.9% (under the traditional last-click model its attribution weight is only around 10%, a severe undervaluation);
- Shapley weight of the Meta feed (channel B): 19.4% (it plays the key role of cross-platform awareness re-engagement, but carries the lowest marginal contribution in this set of assumptions);
- Shapley weight of Google brand search (channel C): 48.6% (highest under these assumptions — brand search converts 2.5% on its own, far above the other single touchpoints, and stays strongest once prior seeding is stripped out).
💡 Operating conclusion: under these assumptions, in-depth reviews by top YouTube tech influencers carry 31.9% of the Shapley conversion momentum, but weight alone understates their value — removing A drops total conversion from 6.0% to 3.5% (−41.7%, leaving {B,C}); by contrast, removing C drops it to 2.8% (−53.3%, leaving {A,B}), so C has the strongest closing effect while A's value shows up mainly through its synergy with C ({A,C} reaches 4.6%).
The right move is therefore not "cut YouTube based on last-click", but: look at how much total conversion falls when a channel is removed (counterfactual test), then set budget. Deciding on last-click alone (in this example C is always the final touch, so A scores 0) systematically undervalues every upstream seeding channel.
3. Unified Omnichannel E-Commerce Metric Definitions and Calculations
To eliminate data disputes across departments entirely, omnichannel metrics strictly follow the standardized definitions in 00.01 Unified Terminology Quick Reference and 00.02 Core Metric Definition Dictionary:
1. Advertising and Traffic Efficiency Metrics
| Metric | Standard Calculation Formula | Industry Health Benchmark / Target | Operating Warning and Circuit-Breaker Red Line |
|---|---|---|---|
| ACOS | Ad spend ÷ direct revenue generated by ads | Mature stage < 18.0% (new-product ramp stage < 30.0%) | > 35.0% rigidly pause loss-making keywords |
| ROAS | Total sales generated by ads ÷ total ad spend | Omnichannel blended target ≥ 3.0 • Amazon on-platform: 3.0~4.0 • TikTok social paid media: 2.0~3.0 • DTC standalone-site cold start: 1.8~2.5 | Blended omnichannel < 2.0 immediately triggers creative and targeting review |
| CPC | Total ad spend ÷ clicks (Clicks) | Benchmark hardware < USD 1.50 (varies by category and region) | Core keywords breaching USD 3.00 for 3 consecutive days require a bid cut to defend position |
| CVR | Paid orders ÷ page visitors (Sessions) | Benchmark hardware product detail page (PDP) > 10.0% | < 5.0% check the unboxing video, review ratings, and above-the-fold selling points |
| CPA | Total ad spend ÷ ad-attributed orders | < USD 15.00 (for fast-moving, impulse-driven 100-yuan-class devices) | Breaching the per-unit marginal contribution means the spend is loss-making |
2. Customer Lifetime Value Metrics
| Metric | Standard Calculation Formula | Target and Control Benchmark | Mandatory Discipline and Definition Conventions |
|---|---|---|---|
| Customer Acquisition Cost (CAC) | Total actual marketing investment across all channels ÷ total newly won customers | < 14.0% MSRP (equivalent to < 15.2% of net revenue) | Mandatory denominator discipline: never mix up the MSRP and net-revenue denominators (align with 00.02 Core Metric Definition Dictionary) |
| Customer Lifetime Value (LTV) | Net profit on hardware shipments + accessory revenue after the 3-year warranty + subscription software revenue share | ≥ 3.0 × CAC | < 2.0 × CAC is judged an "acquisition loss trap" |
| Payback Period (Payback) | Customer acquisition cost CAC ÷ monthly average net gross profit contribution per customer | < 6 months (< 3 months recommended for fast-moving hardware) | > 9 months easily triggers a working-capital cutoff and sudden death |
3. Inventory and Supply-Demand Turnover Metrics
| Metric | Standard Calculation Formula | Target Baseline Band | Operating Actions and Risk-Control Red Lines |
|---|---|---|---|
| Days of Inventory (DOS) | Current sellable plus in-transit inventory across all channels ÷ average daily sales over the last 7 days | 45~75 days (30~45 days ocean in transit + 30 days warehouse stock) | < 20 days critical stockout and listing suspension risk; > 110 days dead-stock warning |
| Inventory Turnover (ITO) | Cumulative 12-month cost of goods sold (COGS) ÷ average inventory asset value | > 6.0 times/year | Squeeze working-capital occupancy efficiency and prevent write-down provisions |
| Sell-Through Rate (STR) | Actual shipped units in the period ÷ (opening inventory + units received in the period) | > 70.0% | If the target is missed within two weeks after a major promo, immediately launch B-Stock discounted clearance |
4. Deriving Break-Even ROAS (Any Category Can Compute Its Own)
The ROAS targets given above (omnichannel ≥ 3.0, Amazon 3.0~4.0, DTC 1.8~2.5) are reference values, not a pass line you can copy. The real pass line is set by your own gross-margin structure:
Break-even ROAS = 1 ÷ gross marginDerivation: let the average order value be P, gross margin be m, and ad spend be S. When the ad produces one order, gross profit is P×m and spend is S. Break-even means:
P × m = S ⟹ ROAS = P / S = 1 / mWhat it means directly:
| Gross margin | Break-even ROAS | Note |
|---|---|---|
| 60% | 1.67 | Every USD 1 of ad spend must return at least USD 1.67 in sales |
| 50% | 2.00 | |
| 40% | 2.50 | |
| 30% | 3.33 | |
| 25% | 4.00 | |
| 20% | 5.00 | Low-margin categories lose money easily on paid media |
Three usage rules:
- Pass line ≠ target line: break-even ROAS only means "no loss". The actual target must layer operating-profit requirements on top (e.g. for 10% net margin, target ROAS ≈ 1 ÷ (gross margin − net-margin target)).
- The gross-margin definition must be consistent: channel shipment gross margin vs retail gross margin yield completely different pass lines; use the definition in the 07.06 Single-SKU P&L.
- ROAS is not comparable across channels: Amazon on-platform, TikTok and DTC have different margin structures and therefore different pass lines; never add "Meta ROAS + Amazon ACOS" together as total efficiency (see Section 6).
Note: ROAS only looks at revenue ÷ spend and ignores returns and fulfillment cost. If return rates or last-mile costs are high, include refunds and fulfillment in the denominator, or switch to back-end ROAS (see 05.21 Paid Media Attribution in Practice).
4. Three-Tier BI Dashboard Architecture and Data-Flow Integration
┌─────────────────────────────────────────────────────────┐
│ Smart Hardware Omnichannel BI Data Integration Hub │
└─────────────────────────────────────────────────────────┘
│
┌───────────────────────────────────────┼───────────────────────────────────────┐
▼ ▼ ▼
[E-Commerce & Ad Platform APIs] [Supply Chain & WMS/ERP] [After-Sales & VOC Feedback]
• Amazon SP-API / Ads API • ECang / Lingxing / Tongtu ERP • Zendesk / Gorgias work orders
• Shopify GraphQL API • GoodCang / 4PX / FBA in-transit • Amazon 1~3 star review scraper
• Meta Marketing API / Google Ads • OEM production schedule (PSI) • DTC return/exchange RMA records
│ │ │
└───────────────────────────────────────┼───────────────────────────────────────┘
▼
[Data Warehouse Layer ETL / Snowflake / ClickHouse]
│
┌───────────────────────────────────────┼───────────────────────────────────────┐
▼ ▼ ▼
[L1: CEO / Board Strategy Dashboard] [L2: GTM Lead Battle Dashboard] [L3: Front-Line Execution Dashboard]
• Global net revenue and margin (EBIT) • Rolling 8-week PSI balance • Ad spend by keyword and hour + ACOS
• Cash cycle (CCC) and cash pool level • Omnichannel DOS + stockout alerts • Listing CVR and Buy Box share
• LTV:CAC by product line • Multi-touch attribution model • Week-1 return reason attribution1. Agile BI Tool Selection Recommendations
| Tool Option | Suitable Scale | Typical Advantages | Deployment Cost and Technical Barrier |
|---|---|---|---|
| Feishu Bitable + Automation | 10~30-person startup team | Zero-code agile build, notifications straight into Feishu groups, low cross-team collaboration cost | Free / low cost; query performance degrades beyond 50,000 rows |
| Metabase (open source, self-hosted) | 30~100-person growth-stage team | Deep and flexible SQL querying, automated report push, fully private data | Requires basic data-warehouse engineering capability and a lightweight cloud server |
| Tableau / Power BI | Mature overseas-expansion enterprise with 100+ people | Complex multi-dimensional report analysis, fine-grained permission control, group-level governance capability | Software licence annual cost is relatively high (USD 10~70/user/month) |
5. Daily Data-Analysis and Decision-Loop SOP
1. 15-Minute Morning Standup Health Check (Daily Loop)
- Step 1: Take the pulse: check yesterday's omnichannel sales, total ad spend, blended ACOS, and real-time inventory DOS;
- Step 2: Flag anomalies: has core-keyword CPC risen by > 20%? Has the DOS of any hero SKU fallen below 25 days? Is the return rate rising abnormally?
- Step 3: Set the actions: immediately coordinate with the media optimiser to cut low-efficiency ad budgets, and notify the warehouse team to investigate inbound-to-shelf delays.
2. Bi-Weekly Supply-Demand and Budget Post-Mortem Review (Bi-Weekly Loop)
- Re-tune attribution: run the Shapley multi-touch attribution algorithm to evaluate the cross-platform conversion contribution of YouTube reviews and TikTok short videos;
- Move budget: raise the budget of top-tier seeding channels that perform poorly on last click but high on marginal incremental contribution by 15%~20%, and cut long-tail low-efficiency search terms;
- Balance supply and demand: feed the latest frontline sales forecast into 03.05 Rolling 8-Week PSI Balance Sheet to size the reorder purchase quantities that follow.
6. Common Data Traps and Pitfall-Avoidance Guide
- Survivorship bias: never analyse only the conversion paths of buyers; you must sample-analyse high-value non-converting users who "added to cart and then abandoned checkout (Abandoned Cart)" to find the shipping-cost or payment-method gaps blocking payment;
- Confusing correlation with causation: you cannot draw the absurd conclusion that "buying accessories prevents returns" merely because "users who bought accessories have a lower return rate" — the common cause is that such a user is a high-loyalty geek fan;
- Algorithmic hallucination caused by insufficient sample size: for small and mid-sized standalone sites whose monthly converting order volume has not reached 300~500 orders, it is strictly forbidden to blindly apply complex algorithmic attribution models; use positional attribution (U-shaped) as a transition and focus on in-site funnel conversion.
Execution Checklist
[ ] Sample size checked: below 300~500 monthly orders, do not apply complex algorithmic models — use positional attribution as a transition
[ ] Non-converting sample included: high-value abandoned-cart users analysed, not only purchasers
[ ] Attribution basis unified: never add Meta ROAS and Amazon ACOS together as total efficiency
[ ] Unknown-attribution share checked: above 40%, "last-click ROAS" may not be used alone to cut brand or social
[ ] Projection separated from measurement: teaching examples are labelled "projection" and never cited as real results
[ ] Three-tier BI dashboard built: overview / channel / creative, with traceable data lineage
[ ] Decision loop closed: every attribution conclusion maps to an executable action and a review point