Marketing Attribution Modelling for Multi-Market APAC Brands

Quick Answer: Multi-market APAC brands should assign attribution models by market tier rather than mandating one globally: algorithmic multi-touch attribution where volume and path data support it, rules-based models in small or data-constrained markets, marketing mix modelling for marketplace and dark channels, and geo experiments as the referee.

A regional marketing director at a beauty group I worked with had a problem that will sound familiar. Her dashboard said paid search delivered 62% of revenue across seven markets. Her Taiwan country manager said LINE was where every sale actually started. Her Japan team spent 40% of budget on retail media that never appeared in the global report at all. And her CFO wanted one number: what did the last quarter's incremental marketing spend actually return?

All three of them were right, and the dashboard was wrong. That gap is the core problem in marketing attribution modelling for multi-market APAC brands: the measurement infrastructure most global companies import from headquarters was designed for a two-channel world (Google, Meta) in a single legal jurisdiction. Asia-Pacific is neither. You are measuring across LINE, WeChat, Shopee, Lazada, Rakuten, Naver, Zalo, TikTok Shop, Instagram, YouTube, plus a live-commerce layer and a physical retail network — under eight different privacy regimes, in five currencies, with buying cycles that differ by 3x between Sydney and Jakarta.

Related reading: Ecommerce Platform Comparison APAC 2026: 15 Platforms Ranked

Related reading: B2B Ecommerce Platform Migration 2026: An APAC Buyer's Guide

Related reading: How EU Companies Build Engineering Squads in Singapore: A Step-by-Step Setup

Related reading: Vercel Security Incident: Impact on APAC Teams and What to Audit

This guide is the implementation sequence I'd use. It assumes you have a real budget under pressure, not a research project. It's opinionated about where multi-touch attribution earns its keep and where it's a very expensive way to make last-click look sophisticated.

Prerequisites: What You Need Before You Model Anything

Attribution is downstream of data engineering. Teams that skip this section spend six months arguing about model weights on data that was never trustworthy in the first place. Treat these as gates.

A single conversion definition, agreed across markets

If Hong Kong counts a marketplace order at checkout and Japan counts it after the 8-day cancellation window, your models are comparing different things. You need one written definition of the conversion event, the currency conversion rule (fixed monthly rate vs. spot), the returns treatment, and whether marketplace GMV or net revenue is the target variable. Get this signed off by finance before touching a modelling tool.

Server-side tagging, not just client-side pixels

Apple's App Tracking Transparency has permanently reduced device-level signal. AppsFlyer and Adjust have both reported ATT opt-in rates clustering in the 20–30% range across most verticals since 2021, with meaningful variation by market and app category. Client-side pixels alone will undercount, and they undercount unevenly across your markets — which is worse than undercounting consistently. You need a server-side collection layer (server-side Google Tag Manager, a customer data platform, or your own event API) and server-to-server conversion feeds into the ad platforms.

A consent and privacy map per market

Singapore's PDPA, Hong Kong's PDPO, Australia's Privacy Act (mid-reform), Japan's APPI, South Korea's PIPA and China's PIPL are not interchangeable. PIPL in particular constrains cross-border transfer of personal data out of mainland China, which affects whether WeChat-side identifiers can join your global warehouse at all. Consult the Singapore PDPC and Hong Kong PCPD guidance directly rather than relying on a US vendor's compliance summary.

At least 200–300 conversions per market per month

This is the practical floor for any data-driven or algorithmic attribution model. Google's own documentation notes that data-driven attribution in Google Ads and GA4 requires minimum conversion and path volumes before the model is available. Below that, algorithmic MTA is curve-fitting noise. Small markets should stay on a rules-based model and get measured with experiments instead.

An identity spine, however imperfect

You need one table that resolves hashed email, phone number, loyalty ID, marketplace order ID and app install ID to a single person ID. In APAC, phone number is often the stronger key than email — mobile-first markets, and messaging-app commerce, mean the phone is the login. The GSMA's Mobile Economy Asia Pacific reporting has consistently shown the region as the world's largest mobile subscriber base, and that shapes which identifier your spine should be built on.

Step 1: Map the Real Channel Graph, Market by Market

Build a channel inventory that includes the non-Western platforms

Start with a spreadsheet, one tab per market, four columns: channel, spend, whether it emits click-level data, and whether it emits impression-level data. You will immediately find channels that carry real budget and emit nothing usable — LINE Official Account broadcasts, WeChat Moments, KOL seeding, offline retail counters, Shopee in-platform ads where the reporting stops at the marketplace boundary.

This matters because Asia's commerce is platform-native. Google, Temasek and Bain's e-Conomy SEA research has tracked Southeast Asia's digital economy growing to well over US$200 billion in GMV, with e-commerce as the largest component and marketplaces dominating that channel. If Shopee and Lazada are where the transaction happens, your attribution model's blind spot isn't a rounding error — it's the majority of the funnel.

Classify each channel as measurable, partially measurable, or dark

Three buckets:

• Measurable: click IDs land on your own domain (Google, Meta, TikTok traffic to your DTC site). MTA works here.
• Partially measurable: platform reports conversions inside its own walls, no user-level export (marketplace ads, retail media, Rakuten, Naver shopping).
• Dark: no per-user signal at all (LINE broadcast, OOH in Hong Kong MTR, KOL gifting, TV in Australia).

The split determines your method mix. A brand where 70% of spend sits in "measurable" can lead with MTA. A brand where 60% sits in "partially measurable" and "dark" — common for Japan, Korea and Greater China portfolios — must lead with marketing mix modelling and geo experiments, using MTA only for in-market optimisation of the measurable slice.

Measure path length before you choose a model

Pull the conversion path report from GA4 or your warehouse and get the median number of touchpoints per market. If Australia shows a median path of 5.2 touchpoints and Vietnam shows 1.4, multi-touch attribution will change decisions in Australia and change almost nothing in Vietnam. That single diagnostic saves more money than any vendor demo.

Step 2: Fix the Data Foundation

Stand up server-side collection

Route events through your own endpoint first, then fan out to platforms. A minimal server-side GTM container plus Conversions API feeds gets you the majority of the recovery. Meta's Conversions API documentation and Google's Enhanced Conversions guidance both describe hashed-identifier matching; the practical benefit is that your first-party CRM data becomes the join key rather than a third-party cookie.

Related reading: How to Audit a Failing CRM Implementation: A 5-Step Framework

A basic server-side event payload should carry every ID you own:

1{
2  "event_name": "purchase",
3  "event_time": 1731052800,
4  "event_id": "hk-order-884213",
5  "market": "HK",
6  "user_data": {
7    "em": "<sha256_lowercased_email>",
8    "ph": "<sha256_e164_phone>",
9    "external_id": "<loyalty_id_hashed>"
10  },
11  "custom_data": {
12    "currency": "HKD",
13    "value": 2480.00,
14    "order_channel": "dtc_web",
15    "gclid": "...",
16    "fbclid": "...",
17    "ttclid": "..."
18  }
19}

Note event_id — deduplication between browser and server events is where most implementations quietly double-count. Test it before you trust any model built on top.

Build the identity spine in the warehouse

Your warehouse — BigQuery, Snowflake or Databricks — is where attribution actually gets computed. The spine is one table mapping every identifier to a person. Phone-first, because APAC.

1-- person_identity_map: one row per (identifier_type, identifier_value)
2with orders as (
3  select order_id, market, order_ts, net_revenue_usd,
4         sha256(lower(trim(email)))       as email_hash,
5         sha256(e164_normalise(phone))    as phone_hash,
6         loyalty_id, gclid, fbclid, ttclid
7  from raw.orders_all_markets
8  where order_status = 'confirmed'
9)
10select
11  coalesce(loyalty_id, phone_hash, email_hash) as person_id,
12  market, order_id, order_ts, net_revenue_usd,
13  gclid, fbclid, ttclid
14from orders;

Crude, deliberately. A resolution graph with survivorship rules comes later. What you need in week one is a person_id that lets you count a customer once instead of three times.

Keep China data resident and modelled separately

Under PIPL, moving personal data out of mainland China triggers transfer requirements that most brands are not resourced to satisfy for marketing analytics. The pragmatic architecture: aggregate inside China, export only aggregated, non-personal spend-and-conversion series, and model China with MMM rather than user-level MTA. This is a constraint, not a workaround — and it's one reason a single global MTA model rarely survives contact with an APAC portfolio.

Step 3: Choose the Attribution Model Per Market, Not Per Company

Here's where most global rollouts break. Headquarters picks one model and mandates it across 12 markets with wildly different path lengths, channel mixes and data quality. That's like making the whole squad run the same training programme regardless of position.

The model options, honestly assessed

Last-click (last non-direct touch): 100% of credit to the final touchpoint. Structurally biased toward branded search and retargeting, structurally punishes upper-funnel. Also: cheap, stable, reproducible, and understood by every country manager and every agency in the region. It remains the most widely used approach in practice, and Adobe, Salesforce and Matomo all note last-click's continued dominance despite its known flaws.

First-click: useful only as a companion diagnostic. Run it alongside last-click; the delta between the two tells you how much upper-funnel work your channels are doing.

Linear: equal credit across all touches. Defensible, transparent, wrong in a predictable direction — it over-credits high-frequency low-intent channels like retargeting display.

Time-decay: exponential weighting toward recent touches. A reasonable default for fast-cycle markets (fashion, beauty, F&B in Southeast Asia) where the median path closes in under 72 hours.

Position-based (U-shaped, 40/20/40): credits discovery and conversion, splits the middle. The best rules-based compromise for considered purchases — jewellery, appliances, travel, B2B — where a Hong Kong or Singapore buyer researches for weeks.

Data-driven / algorithmic (Shapley, Markov): assigns credit from observed path data rather than a rule. Genuinely better where you have volume and clean paths. Available off the shelf in Google Ads and GA4 data-driven attribution; buildable in Python against your warehouse.

A market-tier framework you can actually defend

Tier the markets by data quality and volume, then assign methods:

• Tier 1 — high volume, DTC-heavy, long paths (Australia, Japan DTC, Singapore, Hong Kong DTC): algorithmic MTA for in-channel optimisation, MMM quarterly, geo holdout tests twice a year.
• Tier 2 — high volume, marketplace-heavy (Indonesia, Philippines, Vietnam, Thailand, Malaysia): position-based or time-decay MTA on the DTC slice, MMM as the primary budget allocator, platform-native incrementality tests inside Shopee/Lazada/TikTok Shop.
• Tier 3 — low volume or data-constrained (New Zealand, smaller SEA markets, mainland China): rules-based last-click for reporting continuity, plus experiments. Do not build an algorithmic model here.

One group rule that keeps everyone honest: every market reports last-click and its assigned model side by side. The delta becomes the conversation, not the number.

Step 4: Build the Multi-Touch Attribution Model

Assemble the path table

Every MTA model starts from the same shape: one row per touchpoint, ordered within a person and a lookback window. Choose your lookback window per market from the path-length diagnostic in Step 1 — 30 days is a habit, not a finding. For considered categories in North Asia, 60–90 days is often closer to reality.

1select
2  p.person_id,
3  p.market,
4  t.touch_ts,
5  t.channel,
6  t.campaign,
7  row_number() over (partition by p.person_id, c.conversion_id
8                     order by t.touch_ts) as touch_seq,
9  c.conversion_id,
10  c.net_revenue_usd
11from marts.touchpoints t
12join marts.person_identity_map p using (person_id)
13left join marts.conversions c
14  on c.person_id = p.person_id
15 and t.touch_ts between timestamp_sub(c.conversion_ts, interval 60 day)
16                    and c.conversion_ts
17where t.market = p.market;

Critically, include non-converting paths. Markov and Shapley models need the failures to estimate the value of a channel; if you only feed the model converting journeys, you have built an expensive description of your winners.

A Markov removal-effect model in Python

The most interpretable algorithmic approach is the Markov chain removal effect: build a transition matrix across channel states, compute baseline conversion probability, then remove each channel and measure how much the probability drops. That drop, normalised, is the channel's credit.

1import pandas as pd, numpy as np
2from collections import defaultdict
3
4# paths: list of (channel_sequence, converted_bool)
5def transition_matrix(paths):
6    counts = defaultdict(lambda: defaultdict(int))
7    for seq, conv in paths:
8        states = ["start"] + list(seq) + ["conv" if conv else "null"]
9        for a, b in zip(states[:-1], states[1:]):
10            counts[a][b] += 1
11    states = sorted({s for a in counts for s in [a, *counts[a]]})
12    M = pd.DataFrame(0.0, index=states, columns=states)
13    for a in counts:
14        total = sum(counts[a].values())
15        for b, c in counts[a].items():
16            M.loc[a, b] = c / total
17    for absorb in ("conv", "null"):
18        M.loc[absorb, :] = 0.0; M.loc[absorb, absorb] = 1.0
19    return M
20
21def p_convert(M, iters=200):
22    v = pd.Series(0.0, index=M.index); v["start"] = 1.0
23    for _ in range(iters):
24        v = v @ M
25    return v["conv"]
26
27def removal_effects(paths, channels):
28    base = p_convert(transition_matrix(paths))
29    effects = {}
30    for ch in channels:
31        pruned = [([c for c in seq if c != ch], conv) for seq, conv in paths]
32        effects[ch] = 1 - (p_convert(transition_matrix(pruned)) / base)
33    tot = sum(effects.values())
34    return {k: v / tot for k, v in effects.items()}   # credit shares

Run it per market, not pooled. Pooling Australia and Indonesia produces a transition matrix that describes neither. And run it per month — channel dynamics in APAC shift fast around 6.6, 9.9, 11.11 and 12.12 mega-sales, and a model trained on November behaviour will misprice February.

Sanity-check credit shares against spend and against last-click

Before anyone reallocates budget, produce three columns per channel: spend share, last-click revenue share, model revenue share. Where the model and last-click agree, you have a finding you can act on immediately. Where they diverge by more than about 30%, you have a hypothesis that needs an experiment — not a reallocation.

Step 5: Pair MTA With MMM and Incrementality Testing

What's the difference between MTA and MMM?

Multi-touch attribution works bottom-up from user-level path data and answers tactical questions: which keyword, creative, audience or campaign to shift budget between, this week. Marketing mix modelling works top-down from aggregated time-series data — spend, price, promotion, seasonality, competitor activity, weather, macro — and answers strategic questions: how much to put into TV, OOH, retail media and dark channels next quarter.

For multi-market APAC brands the choice isn't either/or, because MTA structurally cannot see the channels that carry a large share of your regional budget. MMM covers the dark and partially measurable buckets from Step 1. Google's open-source Meridian and Meta's Robyn have both made Bayesian MMM accessible without a six-figure vendor contract, which changes the build-versus-buy calculus considerably for regional teams.

Incrementality tests are the referee

MTA and MMM are both models. Experiments are measurement. Geo holdouts — suppress a channel in matched cities or states, compare against control — are the arbiter when the two models disagree. Australia and Japan support clean geo splits; Southeast Asia is harder because of geographic concentration, so use platform-native conversion lift tools and audience holdouts instead.

Run at least two structured tests per year in each Tier 1 market. Nielsen's ongoing ROI research has repeatedly found large gaps between marketers' confidence in their measurement and their ability to validate it — experiments are how you close that gap. Feed the results back as priors into MMM and as calibration on MTA outputs.

A triangulation cadence that fits real operating rhythms

• Weekly: MTA outputs for in-channel optimisation, by market.
• Monthly: last-click vs. assigned-model delta review with country managers.
• Quarterly: MMM refresh, regional budget allocation, cross-market reallocation.
• Half-yearly: experiment results, model recalibration, lookback-window review.

If your reporting cadence doesn't match a decision that actually gets made, kill the report.

Step 6: Select the Platform Stack

Warehouse-native beats black box for multi-market work

The pattern that holds up across regional portfolios: collection layer (server-side GTM, Segment, RudderStack or a self-hosted event API) → warehouse (BigQuery, Snowflake, Databricks) → transformation (dbt) → modelling (SQL rules-based models plus Python for Markov/Shapley plus Meridian or Robyn for MMM) → BI layer (Looker, Power BI, Metabase).

Why warehouse-native: you control the lookback windows per market, you can hold China data separately, you can join marketplace order exports that no SaaS attribution vendor natively ingests, and you own the logic when your agency of record changes. The trade-off is real — you need an analytics engineer, and time-to-first-insight is months, not weeks.

When a packaged tool is the right call

If you have fewer than four markets, mostly DTC, mostly Google and Meta, and no analytics engineering resource, a packaged platform gets you 80% of the value in 20% of the time. GA4's data-driven attribution is free and, for single-domain DTC businesses, materially better than last-click. Above that complexity level, packaged tools tend to fail on exactly the APAC-specific things you need: LINE, WeChat, Shopee, Lazada, Rakuten and Naver connectors, multi-currency handling, and per-market privacy segregation.

The marketing attribution software category is growing fast — Fortune Business Insights projects the global market well above US$5 billion by 2026 — which means a lot of vendors will tell you they handle APAC. Ask for two named reference customers running the exact channel set you run, in your markets. Most cannot produce them.

Staffing and vendor management

A workable regional model: one analytics engineer owning the pipeline and models (this role is the constraint, and the hardest to hire in Hong Kong and Singapore at market rates), one measurement lead owning definitions and stakeholder alignment, and market analysts who consume rather than build. Where the analytics engineering capacity doesn't exist locally, distributed hiring across Taiwan, Vietnam, Malaysia and the Philippines has become a practical answer — the talent pool for warehouse and dbt work in those markets is considerably deeper than the Hong Kong salary band suggests.

On the agency side, put one clause in every contract: platform data access sits with the brand, not the agency. Losing three years of path history in an agency transition has ended more attribution programmes than any technical failure.

Step 7: Operationalise the Model

Publish a measurement charter

One document, two pages, versioned: conversion definitions, lookback windows per market, assigned model per market, currency rules, what MMM governs versus what MTA governs, and who can change any of it. Without this, every quarterly business review reopens the methodology debate and no decision compounds.

Change the incentives, or nothing changes

If country managers are bonused on last-click ROAS, they will optimise to last-click ROAS regardless of what your model says. This is the single biggest reason attribution rollouts stall — it's a compensation problem wearing a data costume. Shift a portion of the target to blended metrics: contribution margin per market, new-customer acquisition cost, or MMM-derived incremental revenue.

Version and monitor the models

Treat models like code. Version the transformation logic in git, snapshot monthly credit shares, and alert when a channel's credit share moves more than a set threshold month-over-month. Most "the model is broken" escalations turn out to be a tracking regression — a pixel dropped in a site release, a UTM convention broken by a local agency, a marketplace export schema change.

Common Mistakes and How to Troubleshoot Them

Model output swings wildly month to month

Usually insufficient volume for the model class. Check conversions per market per month against the 200–300 floor. If you're below it, drop to a rules-based model and lengthen the aggregation period. Second most common cause: a mega-sale month (11.11, 12.12) training the model on behaviour that doesn't generalise. Exclude or flag promotional periods as separate regimes.

Total attributed revenue doesn't match finance

Almost always deduplication or currency. Check event_id dedup between browser and server events, check that marketplace orders aren't counted in both platform reports and your warehouse, and confirm everyone is using the same FX rule. Reconcile to finance monthly and publish the variance — a known 3% gap you can explain beats a mystery.

Direct traffic absorbs 40% of credit

Symptoms of app-to-web handoff loss, missing UTMs on messaging-app links (LINE and WhatsApp strip or wrap URLs), and dark social. Fix the tagging convention first: every LINE broadcast, WeChat article and KOL link gets a UTM. Then check whether "direct" is really "typed brand URL after seeing OOH" — in which case it belongs in MMM, not MTA.

Marketplace channels look unprofitable in MTA and profitable in MMM

Expected, and MMM is usually closer to the truth. MTA cannot see inside the marketplace walls, so it credits only the sliver of the journey that touched your own properties. Don't reallocate away from marketplace spend on MTA evidence alone. Run a platform-native lift test.

The model contradicts strong local intuition

Sometimes the model is right and the country manager is anchored. Sometimes the country manager knows about a channel your data pipeline never ingested. Before overruling either, ask one question: does the touchpoint table contain every channel the market actually spends on? In my experience, roughly half of these disputes resolve to a missing data source, not a modelling disagreement.

Nobody uses the output

The most common failure, and the least discussed. If the model doesn't change a decision someone makes on a specific day, it's overhead. Tie each output to a named decision, owner and cadence, or cut it.

Why Last-Click Still Dominates — and When That's Defensible

Be honest about this. Last-click is the standard across most APAC marketing teams for reasons that aren't only inertia: it's deterministic, agency-verifiable, unaffected by model drift, and every platform reports it natively. When a Hong Kong general manager challenges a number in a Monday meeting, last-click can be traced to a single click ID in about ninety seconds. An algorithmic model cannot.

Last-click is genuinely defensible when your median path is one or two touchpoints, when a market's volume can't support a data-driven model, or when the cost of the modelling programme exceeds the budget it would reallocate. A market spending US$30k a month does not need a Shapley model — it needs a holdout test and a clear-eyed look at contribution margin.

Last-click becomes actively expensive when you're funding upper-funnel work — brand video, KOL, retail media, OOH — that it structurally cannot credit. That's the moment MTA and MMM earn their cost, and it's the argument to take to your CFO: not "better attribution", but "we currently have no way to know whether 35% of regional spend does anything."

What to Do Monday Morning

Three things, in order. None of them require a vendor contract.

1. Run the path-length diagnostic. Pull median touchpoints per conversion, per market, for the last 90 days. This one table tells you which markets multi-touch attribution will actually change and which ones should stay rules-based. It takes an analyst a morning.

2. Audit your channel coverage against spend. List every channel carrying budget in every market, and mark which ones appear in your conversion data at all. The gap between total regional spend and spend-that-is-measurable is your real problem statement — and it's the slide that gets a measurement programme funded.

3. Book the reconciliation meeting. Marketing, finance and one country manager, one hour, one agenda item: agree the conversion definition, the currency rule and the lookback window. Every technical decision downstream depends on it, and it will not happen by itself.

Marketing attribution modelling for multi-market APAC brands isn't a tool purchase — it's an operating discipline that combines rules-based reporting for continuity, algorithmic MTA where data supports it, MMM for the channels MTA can't see, and experiments as the referee. Get the sequence right and you stop arguing about dashboards. Get it wrong and you'll buy a very sophisticated way to keep believing last-click.

If you're building this measurement stack across several APAC markets and need the analytics engineering and data-pipeline capacity to do it properly, talk to Branch8 about how we resource regional data and commerce teams across Hong Kong, Singapore, Taiwan and Southeast Asia.

Sources

• Google & Temasek & Bain — e-Conomy SEA
• GSMA — The Mobile Economy
• Google — Meridian open-source marketing mix modelling
• Google Analytics — GA4 developer documentation
• Meta — Conversions API documentation
• Apple — App Tracking Transparency framework
• Singapore PDPC — Personal Data Protection Act guidance
• Hong Kong PCPD — Personal Data (Privacy) Ordinance