Adobe Summit 2026 · Session OS956 · MARKETO · Watch on Adobe.com
Attribution models fail not because of methodology, but because of dirty upstream data. Razvan Paraschiv, Head of Marketing Systems and Operations at Airwallex, makes the case that Marketo is the right anchor for a complete marketing touch record — including the offline and third-party activities.
The Four-Layer Marketing Touch: Who, What, When, and Context
Paraschiv frames the marketing touch as having four distinct layers: who interacted (lead/contact identity), what channel generated it (UTM medium, UTM source), when it happened (timestamp), and contextual detail about what the interaction was about (UTM campaign, content, and term). Most MOps teams capture two of these four reasonably well. The UTM framework does the heavy lifting across all four layers, but only if the naming conventions are enforced team-wide and consistently.
The practical implication is that your attribution and scoring models are only as good as the touch records feeding them. Paraschiv is explicit: it doesn't matter how sophisticated your attribution model or ML scoring is — without a strong data foundation, both lose accuracy and fail to produce business impact. That's not a caveat; that's the primary design constraint every MOps practitioner should be building around.
Extending Marketo's Activity Log to Capture Every Touch Type
Marketo's native activities — email clicks, opens, web visits, form submissions — are table stakes. Paraschiv's architecture goes further: custom activities logged via API integrations pull in web and mobile push notifications, event platform registrations and attendances, WhatsApp marketing messages, and product usage data. Any marketing tool that touches a lead should be writing back into Marketo's activity log, making it the single source of truth for engagement history.
The harder category is offline conversions: trade show badge scans, partner-organized events, content syndication lead lists. These arrive as flat imports with no automatic activity timestamp. Paraschiv's solution is a Self-Service Flow Step in Marketo connected to Workato (though he notes UiPath or similar RPA/API automation tools work as well). When a list import occurs for a known marketing event, the flow step triggers an API call that retrieves the lead's last UTM details, appends them to the payload, and logs the custom activity against the person record. The result is a complete, queryable touch history in Marketo regardless of channel or delivery mechanism.
The Data Architecture: Marketo to Data Lake to Models
Paraschiv's recommended stack flows from Marketo as the activity source into a data lake, with a CDP as an optional layer depending on how your data platform is structured. Once the engagement data is in the lake alongside CRM data — opportunity stage, closed/won outcomes, account-level contact aggregation — you have the joined dataset needed to build attribution models, ML-based lead scoring, and sales insight feeds back into the CRM.
The CDP is explicitly optional here, which is a pragmatic acknowledgment that not every team has one, and that the modeling work can happen directly in the data lake if your data engineering supports it. What's non-negotiable is getting the CRM opportunity data and Marketo activity data into the same modeling environment so you can link touches to pipeline outcomes at the account level.
Why Milestone-Weighted MTA Beats AI Attribution in B2B
Paraschiv makes a pointed argument against defaulting to AI-driven attribution in B2B: touch volume is often too low for AI models to reach statistical reliability, unlike B2C where data density makes machine learning viable. His preferred approach is a multi-touch attribution model weighted around the most important milestones in the customer journey — lead creation, MQL conversion, deal creation, and deal close.
The mechanic is straightforward: identify which marketing touches occurred in proximity to each milestone, apply business-defined weights to those milestone windows, and assign a fractional dollar value to each touch. The weights should reflect what you know about your specific sales motion — there's no universal configuration. What Paraschiv is arguing against is both the oversimplification of first-touch or last-touch models and the overcomplexity of AI attribution when your data volume doesn't support it. The output should be a dollar value assigned to each marketing touch, which is the whole point of building the foundation in the first place.
Key takeaways
- Standardize your UTM framework across the entire marketing team first — it's the mechanism that populates three of the four layers of a valid marketing touch record.
- Use Marketo custom activities via API to log every non-native channel (push notifications, event platforms, WhatsApp, product usage) into the central activity log, not just the tools Marketo natively integrates with.
- For offline imports (trade shows, content syndication), build a Self-Service Flow Step connected to Workato or equivalent RPA tooling to auto-log custom activities with retrieved UTM context at import time.
- Don't force AI attribution onto B2B data — if touch volume is insufficient, a milestone-weighted multi-touch model will outperform it. Define your own milestone windows based on your sales motion.
- The CDP layer in your data stack is optional; the non-negotiable join is Marketo activity data plus CRM opportunity data in the same data lake environment.
Bottom line
Paraschiv's session is a practical systems design talk, not a strategy pitch — and that's exactly what it should be. If your attribution model is underperforming, the fix almost certainly starts in Marketo's activity log, not in your attribution tool.
