Last updated: August 2026 • By the SMK Vietnam team — a decade in SEO, content and AI-assisted marketing, working with Vietnamese brands and foreign companies entering Vietnam.
Every Vietnamese marketing team I talk to in 2026 has the same quiet panic: the numbers in Facebook Ads Manager, TikTok Ads, Google Ads and Shopee Seller Centre add up to roughly three times the revenue the finance team actually booked. Nobody is lying. The tracking is simply broken — and it is not coming back. This is the practitioner’s guide to what replaces it.
Quick answer: how do you measure marketing in Vietnam now that cookies are gone?
Stop trying to rebuild a perfect user-level path. In Vietnam in 2026, working attribution is a three-layer stack: (1) first-party data — server-side events, hashed phone numbers and Zalo/CRM IDs as the identity spine; (2) marketing mix modelling (MMM) — a statistical model on weekly spend and revenue that needs no cookies at all and finally values Zalo OA, KOC seeding, livestream and offline stores; (3) incrementality tests — geo holdouts and conversion lift studies that tell you what would have happened if you had not run the ad. AI does not replace these methods; it makes them affordable, by writing the modelling code, cleaning messy Vietnamese channel data, and translating the output into a decision a CMO can act on. A small Vietnamese team can now stand up a usable open-source MMM in about two weeks instead of the two quarters an agency used to quote.
ON THIS PAGE
Why attribution broke in Vietnam →
Layer 1: the first-party identity spine →
Layer 2: marketing mix modelling with AI →
Layer 3: incrementality testing →
Which method answers which question →
The 6-week rollout plan →
Five mistakes I keep seeing →
FAQ →
Why did attribution break in Vietnam specifically?
Third-party cookie loss and Apple’s ATT prompt hit every market. Vietnam got a harder version of the same problem for three local reasons.
First, the buying journey lives inside apps that do not talk to each other. A customer sees a TikTok video, asks a question in a Zalo OA chat, checks the price on Shopee, and buys in a physical store in District 1. No pixel spans that. Second, cash and COD are still normal, so a meaningful share of revenue never touches a trackable payment rail. Third, the phone number is the real identity key in Vietnam — not an email, not a cookie — and the phone number sits in your CRM, Zalo OA and the delivery partner’s system, not in the ad platform.
Directional estimates from SMK Vietnam client work and public benchmark studies — treat them as planning ranges, not audited figures. Your own numbers will differ by category.
Layer 1 — What does a first-party identity spine look like in Vietnam?
Before any modelling, you need one place where a customer is one customer. In Vietnam that spine is built on the normalised, hashed phone number, with email and Zalo user ID as secondary keys. Three practical moves:
Normalise the phone number
Convert every 0-prefixed and +84 variant to one canonical format before hashing. This single step usually collapses 15–25% of “new” customers into existing ones.
Move to server-side events
Meta CAPI, TikTok Events API and GA4 Measurement Protocol, fired from your server with hashed identifiers. Browser-only pixels now under-report badly on iOS traffic.
Log the un-trackable
A “how did you hear about us?” field at checkout, a unique promo code per KOC, a distinct Zalo entry link per campaign. Crude self-reported data beats no data.
This is also where your wider AI marketing stack pays off: models are only as good as the joined data underneath them. And do the privacy homework first — Vietnam’s PDPD obligations around consent and cross-border transfer apply to exactly this kind of identifier matching.
“Last-click attribution in Vietnam is not slightly wrong — it is systematically wrong in one direction. It over-credits branded search and retargeting, and it makes Zalo, KOC seeding and livestream look like cost centres. Teams then cut the exact channels that were creating the demand.”
Layer 2 — How does AI make marketing mix modelling practical for a Vietnamese team?
MMM is old technology — regression on aggregate weekly spend versus weekly revenue, with adjustments for adstock (advertising’s decay over time), saturation (diminishing returns) and seasonality. It ignores cookies entirely, which is precisely why it came back. Its historic problem was cost: you needed an econometrician.
That is the part AI genuinely fixed. Open-source libraries — Meta’s Robyn, Google’s Meridian, PyMC-Marketing — are free, and an assistant like Claude or ChatGPT can write the fitting script, reshape your exported Facebook/TikTok/Google/Shopee spend into the wide weekly format the model expects, and explain the response curves in Vietnamese for your board deck. Where AI helps most, in order:
- Data wrangling — reconciling four platform exports with inconsistent date grains, currencies and campaign naming into one clean weekly table. This is 70% of the real work.
- Model scaffolding — generating the Robyn/Meridian config, priors and adstock ranges, then debugging the errors you hit.
- Vietnam-specific controls — reminding you to add Tết, Mid-Autumn, and the 9.9/10.10/11.11/12.12 mega-sale weeks as variables. Miss those and the model blames your ads for a seasonal spike.
- Interpretation — turning a saturation curve into “another 200 million VND/month into TikTok returns roughly half what the first 200 million did.”
One honest limit: AI will confidently produce a model that fits beautifully and means nothing. A high R² on 30 weeks of data with eight channels is overfitting, not insight. Somebody on the team has to hold the statistical judgment. Use Perplexity to pull the source papers if you need to check an assumption.
Layer 3 — What is incrementality testing and why is it the tie-breaker?
MMM tells you what correlated with revenue. An incrementality test tells you what caused it, by deliberately withholding advertising from part of your audience and measuring the gap. Three formats that work in Vietnam:
Geo holdout
Run in HCMC and Hà Nội, go dark in Đà Nẵng and Cần Thơ for 4 weeks, compare. Best for offline-heavy and multi-outlet brands.
Platform lift study
Meta and Google run randomised holdouts natively. Free, statistically clean — but the platform is grading its own homework, so triangulate.
Scaled spend test
Cut one channel’s budget 50% for a month and watch total revenue, not that channel’s ROAS. Cheapest test to start with.
Run two or three of these a year and feed the measured lift back in as a prior for your MMM. That loop — model, test, recalibrate — is what separates a real measurement practice from a dashboard.
Which method answers which question?
| Method | Best question | Watch out for |
|---|---|---|
| Platform-reported | Is this ad set beating that ad set, today? | Never sum across platforms. Double-counting is guaranteed. |
| First-party / CRM | Who actually bought, and what are they worth over 12 months? | Only as good as your phone-number hygiene and consent record. |
| MMM | How should I split next quarter’s budget across all channels? | Needs 2 years of weekly data; useless for daily decisions. |
| Incrementality test | Would this revenue have happened anyway? | Costs real money in foregone sales; needs enough volume to detect a signal. |
✓ What AI genuinely does well here
Cleaning and joining messy multi-platform exports · writing and debugging modelling code · flagging Vietnamese seasonality you forgot · drafting the board-ready explanation in two languages · designing test geographies and sample sizes.
⚠ What it will get wrong unsupervised
Declaring an overfit model valid · inventing plausible-looking coefficients · confusing correlation with lift · ignoring PDPD consent constraints · assuming US channel mix applies to a Zalo- and TikTok-heavy market.
The 6-week rollout: from broken dashboard to a defensible number
Week 1 — Audit and admit the gap
Put platform-reported conversions next to finance-booked revenue for the last 12 months. Quantify the over-claim. Get leadership to accept the number before you propose a fix.
Week 2 — Fix the identity spine
Normalise phone numbers, deduplicate the CRM, add the self-reported source field at checkout, confirm consent language meets PDPD.
Week 3 — Go server-side
Meta CAPI, TikTok Events API, GA4 Measurement Protocol with hashed identifiers. Verify match rates before you trust anything downstream.
Week 4 — Build the weekly spend table
Two years of weekly spend by channel, weekly revenue, plus Tết, Mid-Autumn, mega-sale and price-change columns. Have AI reshape the exports; you check the totals by hand.
Week 5 — Fit the first MMM
Robyn, Meridian or PyMC-Marketing. Treat the first output as a hypothesis, not a verdict. Sanity-check it against what the team already believes.
Week 6 — Launch one incrementality test
Pick the channel the MMM and the platform disagree about most. Run a 4-week geo holdout or a 50% budget cut. Feed the result back into the model.
Five mistakes I keep seeing in Vietnamese measurement projects
| Mistake | Do this instead |
|---|---|
| Summing Meta + TikTok + Google conversions into one ROAS | Compare each platform only to itself; use MMM for the cross-channel view |
| Excluding Zalo OA and KOC seeding because they are “unmeasurable” | Put their spend in the MMM as a channel — that is exactly what MMM is for |
| Fitting an MMM on 6 months of data | Wait for ~2 years of weekly data, or start with incrementality tests instead |
| Forgetting Tết as a control variable | Add lunar-calendar seasonality explicitly; it moves every year |
| Letting AI both build and validate the model | A named human signs off on model validity before any budget moves |
Key takeaways
- Platform-reported conversions in Vietnam typically over-claim revenue by roughly 2–3× when summed; never add them together.
- The Vietnamese identity key is the normalised, hashed phone number — not the cookie, not the email.
- MMM needs no user-level tracking and finally lets you value Zalo OA, KOC seeding, livestream and offline stores.
- Incrementality tests are the only method that measures causation; run two or three a year and feed results back into the MMM.
- AI’s real contribution is speed and accessibility — data cleaning, code, interpretation — not statistical judgment.
- Add Tết, Mid-Autumn and the 9.9–12.12 mega-sale weeks as explicit model variables or your results will be wrong.
Related reading from SMK Vietnam
The Complete AI Marketing Stack for Vietnam (pillar)
Claude AI for Marketing in Vietnam
ChatGPT for Marketing in Vietnam
Gemini in Google Workspace for Marketing in Vietnam
Perplexity AI for Marketing in Vietnam
Not sure which of your channels is actually working?
SMK Vietnam builds first-party data foundations, AI-assisted MMM and incrementality test plans for Vietnamese companies and foreign brands operating in Vietnam. Tell us your channel mix and we will tell you what your numbers are hiding.
Talk to SMK Vietnam →About SMK Vietnam
SMK Vietnam (smkvietnam.com) is a marketing hub based in Vietnam serving both Vietnamese companies and foreign brands from the US, Japan, Korea, Thailand and Europe that market in Vietnam. We specialise in AI-assisted SEO, content, performance marketing and marketing measurement for the Vietnamese market — including Zalo, TikTok, Shopee, Facebook and Google.
Leave a Reply