A survey result rarely qualifies as a revelation. Most industry polls confirm what practitioners already suspect, dressed up in percentages that make the obvious feel rigorous. But the Madison Logic/Harris Poll report published in May 2025 contains a number worth pausing on: 48% of enterprise marketers admit they are guessing which activities actually drive purchasing decisions. Not inferring. Not estimating with partial data. Guessing.
This admission arrives at a moment when the martech industry is saturated with predictive AI tools promising to eliminate exactly this kind of uncertainty. Zappi, xpln.ai, Criteo, and dozens of other vendors are releasing AI-powered scoring, testing, and optimization products at a pace that suggests the guessing problem should already be solved. It is not. The reason has less to do with the quality of AI models and more to do with the condition of the data those models ingest.
1. Historical context
Marketing attribution has been a contested discipline for over two decades. In 2004, when Google Analytics launched and gave marketers their first widely accessible window into digital behaviour, the prevailing model was last-click attribution. Whichever channel a buyer touched before converting received full credit. This was wrong, but it was simple, and simplicity won.
By the early 2010s, multi-touch attribution (MTA) emerged as the corrective. Vendors like Convertro (acquired by AOL in 2014), Visual IQ (acquired by Nielsen in 2017), and Bizible (acquired by Marketo in 2018) built models that distributed credit across touchpoints. The promise was accurate. The execution was torturous. MTA required clean, unified data across channels, consistent UTM tagging, cookie-based identity stitching, and cross-platform tracking that browsers and regulators were already moving to restrict.
Apple's Intelligent Tracking Prevention (ITP) arrived in Safari in 2017. GDPR took effect in 2018. Google announced the deprecation of third-party cookies in 2020 (then delayed it repeatedly through 2024). Each of these events degraded the signals MTA depended on. Marketing mix modelling (MMM), an older statistical technique popular in CPG, experienced a revival. Meta released its open-source Robyn MMM framework in 2022. Google launched Meridian in 2024. Both acknowledged that granular, user-level attribution was becoming unreliable.
So here we are. The industry spent 15 years building increasingly sophisticated attribution tools, only to watch the data inputs those tools require erode beneath them. The 48% guessing rate from the Madison Logic/Harris Poll is not a failure of ambition. It is the accumulated debt of an industry that invested in models before it invested in data infrastructure.
"The amount of MarTech and data available to marketers has exploded, but the amount of actionable insight has barely moved."
2. Technical analysis
The Madison Logic/Harris Poll also found that 90% of marketers agree an analytics-driven approach outperforms a creativity-first strategy. This near-unanimity is striking because it coexists with the 48% guessing admission. Marketers believe in data-driven decision-making. They simply cannot practice it.
Three technical gaps explain the contradiction.
The identity fragmentation problem
Predictive models require a stable unit of analysis. In B2C, that unit is typically an individual. In B2B, it is an account, a buying committee, or sometimes an opportunity. In practice, most enterprise marketing platforms struggle to maintain a consistent identity across these levels. A single prospect might exist as three records in Salesforce (one from a webinar registration, one from a form fill, one created by sales), two in Eloqua or Marketo (one per business unit), and an anonymous cookie profile in the web analytics layer.
Predictive AI models trained on this fragmented identity graph do not produce better answers. They produce more confident wrong answers. A model might correctly identify that contacts who attend webinars convert at higher rates, but if 30% of webinar attendees are duplicate records, the model is overfitting to a data quality artefact. We explored related identity challenges in our analysis of identity resolution under agentic AI, where the tension between AI-powered personalization and data integrity grows more severe as models become more autonomous.
The signal-to-noise decay in engagement scoring
Most marketing automation platforms score leads based on engagement signals: email opens, page visits, content downloads, form submissions. These signals have degraded significantly. Apple's Mail Privacy Protection (MPP), introduced in iOS 15 in September 2021, rendered email open rates unreliable for approximately 50-60% of B2B audiences (Litmus Email Client Market Share data, Q1 2025). Bot clicks from security filters inflate click rates. Gated content downloads often reflect intent to consume content, not intent to buy.
Predictive lead scoring models trained on these signals inherit their noise. When a model reports that a prospect has an 82% likelihood of converting, the marketing operations team needs to understand whether that confidence is grounded in genuine buying behaviour or in a cascade of bot-inflated engagement metrics. Without rigorous data quality practices upstream, the prediction is a decoration.
The attribution timestamp gap
Even when identity is resolved and signals are clean, B2B buying cycles create a temporal challenge that most predictive models handle poorly. Enterprise purchases often span 6 to 18 months. A whitepaper download in January may contribute to a deal that closes in November. The CRM records the closed-won date. The marketing automation platform records the download date. But the causal link between them passes through dozens of untracked interactions: internal stakeholder conversations, analyst briefings, competitor evaluations, procurement reviews.
Predictive models can identify correlations across this timeline, but they cannot observe the untracked middle. This is why so many marketers are guessing. The signals they can measure (digital touchpoints) account for a fraction of the buying process. The signals they cannot measure (internal consensus-building, executive sponsorship shifts, budget reallocation) account for the rest.
3. Strategic implications
The 90/48 split in the Madison Logic/Harris Poll data (90% believing in analytics, 48% guessing) has a name in organizational theory. Chris Argyris, the late Harvard Business School professor, called it the gap between espoused theory and theory-in-use. Organizations say they value data-driven marketing. Their operational reality tells a different story.
For enterprise marketing operations leaders, this gap creates three strategic risks.
Risk 1: AI investment without data readiness
Gartner's 2024 CMO Spend Survey found that martech accounted for 23.8% of total marketing budgets, the largest single category. A growing share of that spend is directed at AI-powered tools for predictive analytics, content generation, and campaign optimization. But AI tools are accelerants, not foundations. Deploying predictive lead scoring on a database with 25% duplicate records and inconsistent field mappings produces forecasts that feel scientific and are operationally useless.
The implication is that data management work (deduplication, normalization, enrichment) needs to precede or at minimum run parallel to AI deployment. This is unsexy work. It does not generate conference keynotes. But it determines whether predictive models produce actionable intelligence or plausible fiction. As we noted in our examination of the CRM-email convergence and its data quality gap, the unglamorous discipline of data hygiene is the actual prerequisite for every AI use case the industry is chasing.
Risk 2: The automation of bias
When 48% of marketers are guessing, they are making assumptions about what works. Those assumptions become encoded in campaign structures, content calendars, and channel allocations. When predictive AI models are trained on the outcomes of those assumption-driven campaigns, they learn to replicate the assumptions. A team that has historically over-invested in top-of-funnel content because it is easy to measure will train a model that recommends more top-of-funnel content. The model is not wrong given the data. It is faithfully reproducing a biased strategy.
Breaking this cycle requires what statisticians call an exploration-exploitation tradeoff. Some percentage of campaign activity must be deliberately experimental, testing channels, messages, and audiences that the existing model would not recommend. Without this, predictive AI becomes a lock-in mechanism for existing biases.
Risk 3: Vendor consolidation obscuring attribution
The martech industry is consolidating around platform ecosystems. Salesforce, Adobe, Oracle, and HubSpot each offer integrated suites spanning CRM, marketing automation, analytics, and increasingly, AI. These suites simplify operations but also create attribution blind spots. When the analytics layer, the campaign execution engine, and the CRM are all owned by the same vendor, the platform has an incentive to attribute success to its own channels.
Enterprise teams need independent attribution logic that sits outside any single vendor's ecosystem. This might mean a dedicated data warehouse (Snowflake, BigQuery, Databricks) with custom attribution models, or it might mean a purpose-built CDP with transparent methodology. Either way, the attribution function should not be owned by the same vendor that is being evaluated.
Source: Gartner CMO Spend Survey 2024
"Good data beats clever algorithms. Every single time."
4. Practical application
Closing the guessing gap requires interventions at three levels: data infrastructure, model governance, and organizational process.
Step 1: Audit your identity graph before deploying predictive models
Before investing in any AI-powered scoring or attribution tool, run a data quality audit focused specifically on identity resolution. How many unique contacts exist in your marketing automation platform versus your CRM? What is your duplicate rate? How are anonymous web visitors matched to known contacts? What percentage of your database has complete firmographic and demographic fields?
If your duplicate rate exceeds 15% or your field completeness rate is below 70%, predictive model outputs will be unreliable. Invest in data deduplication and data normalization first.
Step 2: Build a signal integrity scorecard
Not all engagement signals are created equal. Create a tiered signal framework that classifies every trackable interaction by reliability:
- Tier 1 (high reliability): Demo requests, pricing page visits, sales-accepted meetings, direct reply emails
- Tier 2 (moderate reliability): Content downloads behind progressive profiling forms, webinar attendance (confirmed, not registered), repeat visits to solution pages
- Tier 3 (low reliability): Email opens (post-MPP), single page visits, social media impressions, generic form submissions
Weight your predictive models and lead scoring rules accordingly. A model that treats all signals equally will produce scores that feel precise and mean little.
Step 3: Implement controlled experimentation
Allocate 10-15% of your campaign budget to structured experiments that test assumptions your current models cannot challenge. If your predictive model says webinars are your highest-converting channel, run a quarter where you redirect webinar spend to targeted direct mail or executive briefing dinners and measure the impact. If the model says mid-funnel content performs best, test whether a concierge-led nurture strategy that bypasses content entirely produces better pipeline.
The goal is not to invalidate the model. It is to feed the model new data patterns that prevent it from calcifying around historical assumptions.
Step 4: Separate attribution from execution
Build your attribution logic in a platform-agnostic environment. If you run Eloqua, Marketo, or SFMC, export touchpoint data to a data warehouse and build attribution models there. This ensures that your understanding of what works is not filtered through the same vendor whose tools you are evaluating. A campaign maturity assessment can help identify whether your current campaign reporting infrastructure is capable of supporting independent attribution.
Step 5: Establish a buying behaviour baseline
Before any predictive model can tell you what influences a purchase decision, you need a qualitative baseline. Interview 15-20 recent buyers (closed-won) and 10-15 recent losses (closed-lost). Ask them to reconstruct their decision process. Which touchpoints did they remember? Which did they ignore? What happened internally that no digital tracking could capture? Use these interviews to build a narrative map of the buying process that your quantitative models can be tested against.
5. Future scenarios
The next 18 to 24 months will determine whether predictive AI narrows the guessing gap or widens it. Three scenarios are plausible.
Scenario A: The data readiness reckoning (most likely)
As more enterprise teams deploy predictive AI tools and discover that model outputs conflict with each other and with reality, a wave of investment in data infrastructure will follow. This pattern has precedent. The CRM adoption boom of the early 2000s was followed by a data quality remediation wave in the late 2000s. The marketing automation boom of the early 2010s was followed by the rise of CDPs in the late 2010s. Predictive AI will trigger a similar corrective cycle, with demand surging for data enrichment, identity resolution, and cross-platform data unification.
In this scenario, the vendors that win are not the ones with the best models. They are the ones with the best data preparation tooling. Companies like Snowflake, Fivetran, and Census (reverse ETL) will benefit more than pure-play AI vendors.
Scenario B: The agentic attribution agent
AI agents that can autonomously query multiple data sources, run attribution analyses, and recommend budget reallocations are already being prototyped. Salesforce's Agentforce, HubSpot's agent ecosystem, and startups like 11x and Relevance AI are building towards this. By late 2026, it is plausible that enterprise marketing teams will have AI agents that continuously run attribution experiments, reallocate spend in real-time, and flag when a model's confidence exceeds what the underlying data supports.
The risk in this scenario is governance. An AI agent that reallocates budget autonomously without human review could optimize for metrics that look good in a dashboard but destroy brand equity or customer relationships. As we explored in our analysis of the 88% daily AI usage claim, the speed of AI deployment is outpacing the organizational structures needed to govern it.
Scenario C: The measurement retreat
In this less optimistic scenario, the complexity of attribution in a privacy-constrained, multi-channel, long-cycle B2B environment proves intractable. Enterprise teams give up on granular attribution and retreat to simpler heuristics: brand awareness surveys, pipeline velocity metrics, and win/loss analysis. Predictive AI is repurposed from attribution to execution (optimizing send times, subject lines, and audience segments) rather than answering the harder question of what actually influences a purchase.
This scenario is not as bleak as it sounds. Execution-level optimization delivers real value. But it concedes the strategic question. And as long as 48% of marketers are guessing about what drives purchases, marketing will struggle to command the boardroom credibility that finance and product functions already enjoy.
6. Takeaways
-
The Madison Logic/Harris Poll finding that 48% of marketers guess what drives purchases is a data infrastructure problem, not an analytics sophistication problem. Predictive AI cannot compensate for fragmented identity graphs, degraded engagement signals, and inconsistent CRM data.
-
The 90% of marketers who endorse analytics-driven strategies are correct in principle but operationally unprepared. The gap between aspiration and execution is widening as AI tools proliferate faster than data foundations improve.
-
Signal quality matters more than signal volume. Email opens, single page views, and generic form fills are increasingly unreliable inputs for predictive models. Enterprise teams should tier their engagement signals by reliability and weight models accordingly.
-
Attribution logic should be platform-agnostic. Building attribution inside the same vendor ecosystem being evaluated creates structural conflicts of interest. Export touchpoint data to a neutral data warehouse for independent analysis.
-
Controlled experimentation (allocating 10-15% of budget to assumption-testing campaigns) is the only reliable way to prevent predictive models from calcifying around historical biases.
-
The most likely near-term outcome is a data readiness reckoning: enterprise teams will invest in data deduplication, normalization, and enrichment not because it is exciting, but because their AI tools will force the issue by producing visibly contradictory results.
-
Qualitative buyer interviews remain the most underused tool in the enterprise marketing operations arsenal. No model can capture the internal consensus-building process that determines most B2B purchases. Talk to your buyers.
