Attribution Built for Shopify

There isn't a single winner in 2026. The best ROI is increasingly coming from matching channels to buying intent, rather than trying to find the cheapest CPM.

Here's how I'd think about it:

Channel	Best for	Typical strengths	Typical weaknesses
TikTok	Consumer products, apps, ecommerce	Low CPMs, excellent creative discovery, strong UGC	Requires constant creative production
Meta	Ecommerce, DTC, local services	Still one of the best conversion platforms despite rising costs	Higher CAC, creative fatigue
Google Search	High-intent demand capture	Highest purchase intent	Expensive in competitive categories
YouTube	Education, SaaS, considered purchases	Strong for building trust before conversion	Longer attribution window
LinkedIn	B2B	High-quality leads and ABM	Highest CPCs of major platforms
Reddit	Technical products, software, gaming	High trust in niche communities	Requires authentic messaging
Programmatic	Large brands, remarketing, scale	Massive reach	Quality varies greatly without good targeting

What's working particularly well in 2026

1. Creator whitelisting / partnership ads

Rather than running polished brand ads, many advertisers are getting better results by turning creator content into paid ads.

This tends to outperform traditional creative because it feels native to the platform.

---

2. YouTube is underrated

For products above roughly $100—or anything requiring explanation—YouTube often produces a lower blended CAC than short-form video.

People are willing to spend 5–15 minutes learning before buying.

---

3. Reddit continues to improve

For categories like:

• AI
• Developer tools
• SaaS
• Finance
• Consumer tech
• Health (where permitted)

Reddit often produces higher-quality users because people are already researching.

The key is avoiding overly polished ads.

---

4. Retail media

If you're selling physical products, marketplaces and retailers have become major advertising ecosystems.

Examples include:

• Amazon Ads
• Walmart Connect
• Instacart Ads
• DoorDash Ads

These audiences have extremely high purchase intent.

---

5. First-party audiences

The highest ROI often isn't a new acquisition channel at all.

Brands investing in:

• email
• SMS
• loyalty programs
• customer data platforms
• server-side tracking
• lifecycle automation

are often seeing the biggest improvements in profitability because retaining and reactivating customers is usually cheaper than acquiring new ones.

LinkedIn: expensive but often worth it

Many marketers dismiss LinkedIn because CPCs can exceed $15–30.

However, if your customer is:

• VP
• Director
• CIO
• CFO
• Enterprise buyer

paying $300 for a qualified lead can be a bargain compared with other channels.

For B2B enterprise, LinkedIn frequently delivers the highest lead quality even if it doesn't have the lowest CPL.

Programmatic has changed

Open-exchange display advertising is rarely where I see standout ROI today.

Instead, better-performing approaches include:

• connected TV (CTV)
• premium publisher marketplaces
• contextual targeting
• retail media networks
• first-party audience activation

Programmatic is strongest when it's integrated with solid customer data rather than relying on third-party audience segments alone.

If I were allocating a fresh performance budget

For consumer ecommerce:

• 35% Meta
• 25% Google Search & Shopping
• 20% TikTok
• 10% YouTube
• 10% testing (Reddit, Pinterest, retail media, creators)

For B2B SaaS:

• 30% Google Search
• 25% LinkedIn
• 20% YouTube
• 15% Reddit
• 10% experimentation (podcasts, newsletters, niche communities)

The biggest shift

The advertisers seeing the strongest returns are generally not the ones chasing the newest ad platform. They're the ones that:

• produce a high volume of creative and test it rapidly,
• use first-party data to improve targeting and measurement,
• optimize for incrementality and lifetime value rather than just last-click ROAS, and
• diversify acquisition so they're not overly dependent on a single platform.

In many cases, improving creative quality, landing pages, and post-purchase retention has a larger impact on ROI than moving budget from one major ad platform to another.

If you're willing to share your industry (ecommerce, SaaS, local services, mobile apps, etc.) and approximate monthly ad spend, I can suggest a more tailored channel mix and testing roadmap.

The short answer is: no single approach is reliably correct anymore. The most reliable organizations have shifted away from trying to find one "source of truth" and instead use different measurement methods for different decisions.

Here's how I'd think about each approach today.

Method	Good for	Weaknesses	Reliability today
Multi-touch attribution (MTA)	Optimizing within digital channels	Privacy restrictions, walled gardens, cookie loss	★★☆☆☆
Platform attribution (Meta, Google, TikTok)	Campaign optimization	Self-attribution bias, limited cross-channel visibility	★★★☆☆
Incrementality testing	Measuring causal lift	Expensive, requires experimental design	★★★★★
Media Mix Modeling (MMM, Robyn, Meridian, etc.)	Budget allocation across channels	Less granular, requires historical data	★★★★☆
Marketing experiments	Creative, pricing, landing pages	Doesn't scale to every question	★★★★★

Multi-touch attribution isn't dead—but it's no longer the decision engine

Five years ago, many teams built elaborate attribution models:

• First-touch
• Last-touch
• Linear
• Time decay
• Position-based
• Algorithmic attribution

Today those models have major blind spots because of:

• iOS App Tracking Transparency
• Third-party cookie deprecation
• Cross-device journeys
• Walled gardens (Meta, Amazon, TikTok, LinkedIn)
• Consent requirements

Even if your attribution model is mathematically elegant, it's often assigning credit over an incomplete customer journey.

That doesn't make MTA useless—it just makes it better suited for operational optimization than for answering "what actually caused growth?"

Incrementality testing is the closest thing to ground truth

If the CEO asks:

"If we stop spending $500k on YouTube, what happens?"

Attribution can't answer that.

An incrementality test can.

Typical approaches include:

• Geographic holdouts
• Audience holdouts
• PSA/control campaigns
• Time-based experiments
• Conversion lift studies

These estimate the causal effect rather than just observed correlations.

That's why many sophisticated growth teams increasingly treat experimentation as the gold standard.

MMM has made a comeback

Media Mix Modeling used to be associated with large consumer brands spending tens of millions on TV.

Privacy changes have made it relevant again because it doesn't rely on user-level tracking.

Modern open-source implementations like:

• Robyn (Meta)
• Meridian (Google)
• Lightweight Bayesian MMM frameworks

can estimate:

• diminishing returns
• channel saturation
• carryover effects (adstock)
• optimal budget allocation
• ROI by channel

MMM works especially well when:

• multiple offline and online channels interact
• attribution data is incomplete
• budgets are significant
• executive budget planning matters

Its weakness is that it generally isn't useful for deciding which Facebook ad to pause tomorrow.

What high-performing marketing organizations often do

Rather than choosing one method, they layer them.

Daily optimization

• Platform reporting
• Limited MTA
• Conversion APIs
• First-party analytics

↓

Weekly optimization

• Channel attribution
• Creative testing
• Funnel analysis

↓

Monthly/quarterly validation

• Incrementality tests

↓

Quarterly/annual planning

• Media Mix Modeling

Each layer answers a different question.

A practical framework

Think about matching the measurement method to the decision you're making.

Decision	Best method
Which creative wins?	A/B test
Which landing page converts?	Experiment
Which Meta campaign to pause tomorrow?	Platform attribution + MTA
Should we spend more on YouTube vs. Search?	MMM + incrementality
What's the true ROI of TV?	Incrementality + MMM
How should next year's budget be allocated?	MMM validated with experiments

If I were building a measurement stack today

I'd prioritize something like this:

1. Strong first-party event collection (server-side where practical).
2. Platform conversion APIs to improve optimization signals.
3. Limited multi-touch attribution for operational reporting, while recognizing its gaps.
4. A regular cadence of incrementality tests for major channels and campaigns.
5. An MMM (such as Robyn, Meridian, or another Bayesian approach) to guide strategic budget allocation.
6. A process where MMM and incrementality results are compared periodically to recalibrate assumptions.

The biggest shift in the last few years is philosophical: marketers are moving from asking "Which touchpoint gets credit?" to asking "What actually caused additional business outcomes?" Attribution models estimate credit; experiments estimate causality. As privacy constraints have reduced visibility into individual user journeys, the industry has increasingly favored a combination of experimentation and media mix modeling for strategic decisions, while using attribution primarily as a tactical optimization tool.

For a medium-sized e-commerce dataset, I would almost never start by training a transformer. I'd build the system in stages, optimizing for time-to-production, maintainability, and measurable business impact rather than model sophistication.

Here's how I'd think about it in 2026.

Approach	Start here?	Best for	Downsides
Matrix factorization / LightFM	✅ Yes	Fast baseline, sparse interaction data	Limited feature usage, weaker cold start
Two-tower retrieval	✅ Usually the first neural model	Large catalogs, rich features, scalable retrieval	More engineering complexity
Transformer recommenders	❌ Usually later	Sequential behavior, session-aware recommendations	Expensive, harder to tune, often marginal gains

Phase 1: Build a strong classical baseline

I'd begin with something like:

• Implicit matrix factorization (ALS or BPR)
• LightFM if side features matter
• Item-item collaborative filtering
• Popularity + business rules fallback

Why?

Because you'll quickly answer questions like:

• Is the data good enough?
• What's the achievable offline metric?
• How sparse are interactions?
• Is cold start the real bottleneck?

Many recommendation projects fail because the data pipeline is poor—not because the model isn't sophisticated.

A surprisingly large number of production recommenders still rely heavily on collaborative filtering plus heuristics.

---

Phase 2: Move to a two-tower model

This is where I'd invest engineering effort.

A two-tower architecture has become something of the default retrieval model because it separates:

User tower

Inputs might include:

• viewed items
• purchases
• categories
• brands
• demographics
• embeddings
• session signals

↓

Produces a user embedding.

---

Item tower

Uses

• title
• category
• price
• brand
• image embedding
• text embedding
• popularity
• metadata

↓

Produces an item embedding.

Recommendation becomes

nearest_neighbors(user_embedding)

This scales extremely well.

Advantages:

• ANN indexes (HNSW, ScaNN, FAISS, etc.)
• precompute item vectors
• millisecond retrieval
• supports millions of items
• handles new items better than pure collaborative filtering

This architecture has become fairly standard in industry because it's efficient and flexible.

---

Where transformers shine

Transformers become compelling when order matters.

For example:

Laptop

↓

Mouse

↓

Keyboard

↓

Monitor

↓

Dock

The sequence itself carries information.

Or:

Viewed:
A
B
A
C
D

Purchase:
E

Attention models capture these dependencies better than latent-factor models.

They're useful for:

• session recommendation
• next-item prediction
• homepage ranking
• basket completion
• personalized feeds

Less useful when you only need:

"Given this customer, retrieve 500 candidate products."

---

Retrieval vs. ranking

Modern production systems are usually multi-stage rather than relying on a single model.

User

↓

Candidate Retrieval
(two-tower)

↓

500 candidates

↓

Ranking model

↓

Top 20

The ranking model can be:

• Gradient-boosted trees
• DeepFM
• Wide & Deep
• DLRM-style models
• Small transformers (if sequential context is important)

This is where transformers often provide more value than using them for retrieval.

---

Cold start

This is where LightFM starts to struggle.

A two-tower model naturally incorporates:

• text embeddings
• image embeddings
• metadata
• category
• brand

so a brand-new item can immediately receive a useful embedding before any clicks occur.

That's a major production advantage.

---

Cost

Approximate engineering complexity:

Collaborative filtering

⭐

Can be production-ready in days.

---

Two-tower

⭐⭐⭐

Requires:

• feature store
• embedding training
• ANN index
• online serving

Still very manageable.

---

Transformer

⭐⭐⭐⭐⭐

Now you also need:

• sequence preprocessing
• masking
• longer training
• larger GPUs
• online sequence features
• latency optimization

The engineering burden increases substantially.

---

What I would build in 2026

For a medium-sized retailer:

1. Collaborative filtering baseline (ALS/BPR/LightFM).
2. Item-item similarity for robustness and explainability.
3. Two-tower retrieval model using user history and rich item features.
4. Learning-to-rank model on the retrieved candidates.
5. Only introduce transformer-based sequence models if user behavior indicates that sequential context materially improves business metrics (e.g., session recommendations or next-item prediction).

This progression lets you validate each layer's incremental value before taking on additional complexity.

One additional point that's become more important

In practice, the biggest gains often come from feature quality and serving architecture rather than from replacing one deep architecture with another. Better signals (recency, dwell time, returns, inventory, price changes, promotions, and rich item metadata), well-designed negative sampling, and low-latency feature freshness frequently produce larger improvements than moving from a two-tower model to a transformer.

For a medium-sized e-commerce business in 2026, I'd generally expect the best return on investment from a two-stage system: a two-tower retriever followed by a strong ranker, with transformer models introduced only if sequential behavior is central to the recommendation task and offline gains translate into online improvements through A/B testing.