Multi-Touch Attribution in the AI Commerce Era: Why Your Models Need to Evolve

Multi-touch attribution (MTA) was supposed to solve the attribution problem. By distributing credit across all touchpoints in a customer journey rather than just the first or last, MTA promised to reveal the true contribution of each marketing channel and enable optimal budget allocation.

For a while, it worked reasonably well. As customers moved through increasingly complex digital journeys—organic search to display ad to email to social to conversion—MTA models tracked the path and assigned credit proportionally. Marketing teams made investment decisions based on attributed performance, confident they were seeing the real picture.

Then AI commerce arrived. And multi-touch attribution started to break.

The problem isn't that MTA does anything wrong—it does exactly what it was designed to do. The problem is that AI commerce introduces touchpoints that MTA cannot see. When customers research purchases through AI assistants, those interactions shape their journey before any measurable touchpoint occurs. MTA dutifully attributes credit across observable touchpoints, systematically ignoring the invisible influence that may actually drive the outcome.

Your multi-touch attribution model isn't failing. It's becoming obsolete.

The Multi-Touch Model Gap

Every multi-touch attribution model—whether rules-based, time-decay, position-based, or algorithmic—shares a fundamental assumption: customer journeys can be observed and measured. The model's job is to look at the journey and decide how to distribute credit. But it can only distribute credit among touchpoints it knows about.

Consider how a typical MTA model works:

1. Customer journeys are reconstructed from touchpoint data
2. Each touchpoint is associated with channel and campaign information
3. The model applies rules or algorithms to assign credit
4. Channel performance is calculated based on aggregated credit

The model is entirely dependent on step one: reconstructing journeys from touchpoint data. If a touchpoint isn't in the data, it doesn't enter the model. It receives no credit. From the model's perspective, it doesn't exist.

AI commerce creates precisely this situation. A customer asks ChatGPT for product recommendations. Based on the response, they form preferences and consideration sets. Eventually they begin a measurable journey—perhaps a Google search, a website visit, an email interaction. MTA captures everything from that point forward, attributing the conversion across the observable touchpoints.

The AI conversation that shaped the entire journey? Not in the data. Not in the model. Not in the attribution.

This isn't a gap that can be closed through better implementation or more sophisticated modeling. It's an architectural limitation of the measurement infrastructure. AI conversations happen in environments that don't pass data to your analytics systems. The touchpoint is inherently invisible to any MTA model built on conventional tracking.

Where AI Fits in the Customer Journey

Understanding AI's role in customer journeys reveals why the MTA gap matters so much. AI isn't just another touchpoint to be added to the mix—it operates at a unique position in the journey that gives it outsized influence.

Discovery and need recognition. Many journeys begin with vague problems or questions. "I need something for X." "How do I solve Y?" "What's the best approach to Z?" AI assistants are increasingly the first resource consumers turn to when translating problems into potential solutions. This is often before any brand consideration—making it the most influential stage of the journey.

Consideration set formation. One of AI's most powerful effects is shaping which brands and products consumers consider at all. When ChatGPT recommends three brands in response to a category query, those three brands are the consideration set. Brands not mentioned face enormous barriers to consideration. This filtering happens entirely before measurable touchpoints.

Feature and criteria definition. AI doesn't just recommend products—it frames what matters. "When choosing a laptop, the key factors are..." These framings shape how consumers evaluate options, potentially favoring brands that excel on AI-highlighted criteria. Again, this happens before measurable comparison shopping.

Initial preference formation. AI recommendations often come with implicit or explicit preferences. "Brand A is generally considered the best for..." These early preferences influence how consumers interpret later information. The halo of AI endorsement follows the customer through subsequent touchpoints.

Comparison and validation. Even after forming initial preferences, consumers often return to AI for validation. "Should I go with Brand A or Brand B?" "Is Brand X really worth the premium?" These check-in conversations can reinforce or redirect journeys—but they're as invisible as initial queries.

The common thread is that AI's influence is concentrated in early journey stages—precisely where MTA models are weakest. Even before AI, early journey stages were harder to attribute because they're further from conversion. AI compounds this problem by creating a pre-journey stage that's completely unmeasured.

The Invisible Touchpoint Problem

Let's walk through a specific example to illustrate how invisible touchpoints distort attribution.

The actual journey:

1. Customer asks Claude about the best home security systems
2. Claude describes key features and recommends three brands including Brand A
3. Customer forms preference for Brand A based on the response
4. Two days later, customer searches "Brand A home security" on Google
5. Customer clicks organic search result and visits Brand A website
6. Customer signs up for email list
7. Customer receives promotional email
8. Customer clicks email and purchases

What MTA sees:

1. Google organic search (First touch)
2. Direct website visit
3. Email signup
4. Email click
5. Conversion (Last touch)

What MTA attributes: Depending on the model, credit is distributed across Google organic, direct, and email. Perhaps linear attribution gives each 25%. Perhaps position-based attribution gives 40% to first touch (Google), 40% to last touch (email), and 20% distributed across the middle.

The reality: The Claude conversation was the actual driver of the purchase. It put Brand A on the customer's radar and created the preference that drove all subsequent behavior. Google organic captured intent that AI created. Email converted a customer that AI had already sold.

If we could see the true journey, AI would deserve substantial credit—perhaps the majority of credit. Instead, it receives zero, and that credit is redistributed to observable channels that look more effective than they actually are.

Now multiply this by thousands of customers. The distortion becomes systematic. Channels that capture AI-influenced intent appear to drive conversions. The invisible AI influence that creates that intent gets no credit. Budget flows toward demand capture and away from demand creation.

Attribution Model Limitations

Different MTA model types have different strengths, but all share the same fundamental limitation in AI commerce.

Last-touch and first-touch attribution. These simple models credit either the final or initial observable touchpoint. But if AI influenced the customer before any observable touchpoint, both models credit the wrong moment. The actual journey started earlier, in an invisible conversation.

Linear attribution. Linear models distribute credit equally across all touchpoints. But "all touchpoints" means "all observable touchpoints." If AI is invisible, it's excluded from equal distribution, and its credit inflates observed channels proportionally.

Time-decay attribution. Time-decay gives more credit to touchpoints closer to conversion. This actively undervalues early-journey influence—including AI. Time-decay makes the AI blind spot worse by systematically downweighting the journey stage where AI operates.

Algorithmic attribution. Sophisticated algorithmic models use machine learning to determine optimal credit distribution. But they learn from historical data containing the same blind spot. They optimize credit distribution among observable touchpoints, with no way to learn about channels they cannot see.

The problem isn't that these models are bad at their job. They're excellent at distributing credit among observable touchpoints. But when a growing portion of customer influence is unobservable, even perfect models working with incomplete data produce misleading results.

Emerging Thinking on Attribution Evolution

While there's no solved approach to AI commerce attribution, practitioners and researchers are exploring several directions. These represent emerging thinking, not proven solutions.

Visibility-adjusted attribution. One approach involves adjusting attributed credit based on visibility metrics. If your brand appears in AI recommendations for queries that precede typical purchase journeys, you might discount credit to subsequent touchpoints, acknowledging that AI influence likely played a role. This is imprecise but directionally helpful.

Research-phase modeling. Some organizations are developing separate models for the research phase of customer journeys—the phase where AI influence is strongest. By modeling research behavior separately from transaction behavior, they can at least bound the potential AI influence even without direct attribution.

Mixed-method attribution. Combining quantitative attribution with qualitative research may provide a more complete picture. Survey data revealing AI influence can inform adjustments to quantitative models, even if the adjustment is somewhat subjective.

Attribution humility. Perhaps the most important shift is attitudinal. Organizations are increasingly treating attribution as directional rather than definitive. Acknowledging that models are incomplete encourages decision-makers to supplement attributed data with strategic reasoning and competitive logic.

Incrementality testing. Incrementality approaches—testing whether marketing activities actually cause conversions versus simply capturing them—may reveal more about true impact than attribution models. If your paid search captures AI-influenced demand rather than creating it, incrementality testing might reveal that its true value is lower than attributed.

Probabilistic attribution. Rather than assigning definite credit, probabilistic approaches might assign probability distributions across touchpoints including hypothesized invisible ones. This explicitly acknowledges uncertainty while providing usable estimates.

These approaches are exploratory. None has emerged as the industry standard for AI commerce attribution. But they represent the direction of thinking as the field grapples with fundamentally new challenges.

The Future of Commerce Attribution

Looking ahead, commerce attribution faces a fundamental crossroads. The measurement infrastructure built over two decades assumes observable, traceable customer journeys. AI commerce breaks that assumption.

Possible futures:

Attribution adapts. New data sources, partnerships with AI platforms, or technological innovations might enable AI touchpoint tracking. This would allow evolution of existing attribution approaches rather than wholesale replacement. However, the privacy and architectural challenges are substantial.

Attribution becomes contextual. Rather than precise touchpoint-level attribution, measurement might shift toward contextual understanding—knowing generally how AI commerce affects your business without attributing individual conversions. This is less satisfying but may be more realistic.

Attribution accepts limitations. Organizations might simply accept that attribution is incomplete and supplement it with other forms of evidence. Strategic logic, competitive dynamics, and qualitative research might play larger roles in investment decisions.

Measurement paradigm shifts. Entirely new approaches to marketing measurement—perhaps based on market-mix modeling, experimentation, or methods not yet developed—might replace touchpoint attribution entirely. AI commerce might be the catalyst for a broader measurement revolution.

Most likely, the future involves some combination of these paths. Attribution won't disappear, but its role may diminish as organizations develop comfort with complementary approaches.

Implications for marketing leaders:

The evolution of attribution has significant implications for how marketing teams operate:

• Investment decisions can't rely solely on attributed return. Other evidence must inform allocation.
• Team capabilities need to expand beyond analytics to include strategic reasoning and qualitative research.
• Executive communication needs to explain measurement limitations honestly while still providing actionable insight.
• Competitive monitoring becomes more important as direct impact measurement becomes less reliable.
• Experimentation may become the gold standard for understanding true marketing impact.

Adapting Your Measurement Strategy

While the future evolves, organizations need practical strategies for today. Here's how to adapt your measurement approach for the AI commerce era.

Acknowledge model limitations explicitly. Don't treat attribution as ground truth when you know it's incomplete. Communicate limitations clearly and build decision-making processes that account for uncertainty.

Supplement attribution with visibility metrics. What you can't attribute, you can at least observe. Monitoring AI visibility provides leading indicators that complement attribution's backward-looking view.

Invest in customer research. Survey data revealing AI usage can't provide conversion-level attribution, but it can inform adjustments to how you interpret attributed data.

Use multiple measurement lenses. Attribution plus visibility monitoring plus customer research plus incrementality testing plus competitive analysis—multiple perspectives provide a more complete picture than any single approach.

Build flexibility into processes. Investment allocation based solely on last quarter's attribution locks in historical biases. Build flexibility that allows for strategic adjustments based on qualitative factors including AI commerce considerations.

Maintain measurement perspective. Attribution was never perfect. It struggled with cross-device journeys, offline influence, brand effects, and more. AI commerce is a significant new challenge, but it's an expansion of existing limitations rather than an entirely new category.

Prepare for evolution. The measurement landscape will continue changing. Building adaptable measurement capabilities—rather than investing heavily in current approaches—positions you to evolve as standards emerge.

Multi-touch attribution served marketing well for years. It's not disappearing tomorrow. But its limitations in the AI commerce era are real and growing. Organizations that recognize these limitations and develop complementary approaches will make better decisions than those that continue treating attribution as complete and accurate.

---

Multi-touch attribution is one of many measurement challenges facing marketers in the AI commerce era. Explore related topics including why you can't track AI-influenced purchases and how to prove AI commerce ROI despite attribution limitations.

Ready to evolve your measurement approach? Understanding AI visibility provides insight that attribution cannot. Discover how leading platforms approach this challenge.

---

Want to see how your store scores? Run a free AI readiness scan and get your store's AI visibility report in 60 seconds.

---

About the Author: Josh is the founder of Noema, an AI commerce observability platform that helps e-commerce brands understand how AI shopping agents see their products. Noema has scanned 80,000+ Shopify stores to build the industry's most comprehensive AI readiness benchmarks.