Marketing attribution is a problem that multi-touch models have been failing to fully solve for 20 years. LLMs do not solve it either, but they solve a different problem alongside it — making sense of the unstructured signals (reviews, social, support conversations, competitive activity) that attribution models historically ignore. The combination produces more honest answers than either alone.
What LLMs actually add to marketing analytics
Voice-of-customer at scale
Reading 10,000 product reviews, support tickets, or survey comments to identify themes — LLMs do this well. A marketer asking 'what are customers saying about our new feature' gets a categorized summary in minutes instead of a week. The output quality depends on prompt discipline (specific categories, explicit request for examples, citations back to source rows).
Competitive landscape synthesis
Competitor news, product updates, pricing changes, customer sentiment shifts. LLMs can ingest these feeds, synthesize positioning maps, and flag changes in competitive posture. Higher quality than keyword-alert dashboards most marketing teams have lived with.
Creative performance analysis
Ad creative at scale (thousands of variants, multiple channels) is hard to reason about. LLMs can describe creative (image + copy + targeting), cluster similar concepts, correlate with performance. Teams discover that 'casual testimonial' outperforms 'polished founder shot' — a pattern a human can act on but would take days to identify from data alone.
Marketing briefs and reports
Weekly performance reports, campaign retrospectives, brand-health narratives. Pure productivity — what was a two-hour writing task becomes a 20-minute review-and-edit.
What classical methods still own
Paid channel attribution. Media mix models (MMM) using regression on spend and outcomes remain the most defensible approach to 'what should our paid mix look like?' Bayesian MMM with proper priors, incrementality testing, and saturation curves is the state of the art and is not being displaced by LLMs — the problem is fundamentally causal, not about text.
Incrementality testing. Geographic holdouts, lift studies, causal-forest designs. LLMs have no role in the estimation; they can help with narrative explanation of results.
Customer lifetime value prediction. Supervised ML on transaction histories. LLMs can enrich CLV models with signals from unstructured data (support ticket sentiment, review activity) but don't replace the core model.
The combined pattern in weekly review
Marketing leaders want answers to 'what should we do differently this week'. The answer almost always requires both: what the numbers say (MMM, performance metrics) and why they say it (customer sentiment shifts, competitor moves, creative fatigue). Combining these is an LLM workflow: ingest the numbers, ingest the narratives, produce the synthesis.
Example output from a real client weekly: 'Paid social CAC rose 22% week-over-week. The MMM attribution shows saturation on the primary campaign; incrementality held. Support tickets show a 40% spike in competitor-X comparison questions, and the last three weeks of competitor-X ad creative has shifted toward feature-parity messaging. Recommendation: investment shift from paid social to comparison-landing-page SEO and refresh comparison messaging.' This is the level of answer that actually drives decisions, and neither the pure numbers nor the pure narrative gets there alone.
Common mistakes
Trying to replace MMM with LLMs. Doesn't work; the problem structure doesn't suit LLMs and the regulated-context requirements (CFO defensibility) demand interpretable models.
Trusting LLM outputs without source citations. Every qualitative claim should cite specific reviews, tickets, or signals. Unsourced summaries feel confident and drive wrong decisions.
Skipping the human in the loop for strategy decisions. LLMs synthesize well; they don't judge strategic fit. A CMO reviews the synthesis before it drives budget.