SOMIN

It is a pleasure for SOMIN to present in London about the future frontiers of AI in Marketing and the implications of Agents doing so much work for companies instead of humans! Thanks for helping to organize DEFY Group

The ad industry spent a decade debating third-party cookies. AI just ended that debate in about 18 months. Here is what the data actually shows: 71% of brands, agencies, and publishers are now actively growing their first-party data sets, nearly double the rate from two years ago. US retail media spend is projected at $69.33 billion in 2026. These numbers are not coincidental. They are the market responding to a structural constraint that AI systems have imposed. The constraint is this: agentic AI cannot optimize toward outcomes — sales, retention, incremental lift — without deterministic identity, clean feedback loops, and auditable data lineage. Probabilistic, third-party signals are not just less accurate. They are architecturally incompatible with how AI allocation engines actually work. Brian O'Kelley put it precisely: OpenRTB is a protocol for day trading. AdCP is a protocol for investing. The shift from impression trading to outcome allocation is not a trend. It is a systems-level requirement. What this means technically is that the intelligence layer — the data layer that AI agents query when making allocation decisions — must be first-party, permissioned, and governable. Not because regulators say so, but because the model cannot safely optimize without it. Governance is no longer overhead. It is performance infrastructure. At SOMIN, we see this shift playing out directly in how marketing agencies and brand teams approach audience intelligence. The platforms that win in the agentic era are not those with the most data. They are those with the most structured, contextually rich, and governable data. SOMIN's approach — connecting psychographic audience signals, content performance patterns, and creative intelligence into a unified layer — is precisely the kind of first-party intelligence architecture that AI allocation systems can actually use. Our work with clients like Citibank (43% cheaper CPL, 53% higher CTR) and Audi (43% lower CPC, 45% ad budget savings) reflects what happens when the intelligence layer is clean, structured, and connected to outcomes. Key takeaways: - AI agents require deterministic identity and auditable data lineage to function reliably at scale — probabilistic signals are not a fallback, they are a failure mode - The budget shift toward retail media and premium first-party environments is architectural, not cyclical — it will not reverse when macro conditions improve - Governance is now a competitive differentiator, not a compliance cost — platforms that can prove data provenance and consent will capture disproportionate share of agentic ad spend The question worth sitting with: if your current data architecture cannot answer "what signal was used, under what permission, and what outcome did it drive" — how confident are you that your AI-assisted campaigns are actually optimizing toward the right thing? @SOMIN #MarTech #FirstPartyData #AgenticAI #DigitalAdvertising #DataGovernance

Most LLM-based product search systems are still functionally blind. That is the core finding that motivated the HMPPS framework published at ACM Multimedia 2025. The paper, "Harnessing Multimodal Large Language Models for Personalized Product Search with Query-aware Refinement," by Beibei Zhang et al., addresses a gap that anyone building intelligent commerce or content discovery systems should take seriously: current LLM-driven personalized search pipelines process only text, while the actual purchase decision — and the user's behavioral history — is deeply multimodal. The architecture is technically precise. HMPPS introduces a perspective-guided summarization module that conditions product description generation on the search query itself, stripping out noise and redundancy before the MLLM ever sees the input. Paired with a two-stage training paradigm that filters user history using multimodal representations, the system reduces both inference cost and the semantic drift that plagues naive retrieval approaches. The result is not just a benchmark improvement across four public datasets — it is a system that was deployed on a production search platform with billions of daily active users and demonstrated measurable gains in A/B testing. That is the kind of validation that moves a research contribution from interesting to consequential. From where I sit at SOMIN, this resonates directly with work we have been doing on multi-source user profiling and personality-driven content recommendation. Our 2015 ICMR work on harvesting multiple data sources for user profile learning, and the subsequent SIGIR 2017 work on cross-domain recommendation via multi-layer graphs, were early signals that unimodal user representations leave significant signal on the table. What HMPPS formalizes is the next logical step: not just fusing modalities at the representation level, but making the query itself an active participant in how multimodal context is selected and compressed. For MarTech teams building recommendation or search layers, this is the architectural shift worth studying. Key takeaways: - Query-aware refinement is not optional. Feeding raw multimodal user histories into an MLLM without query conditioning introduces noise that degrades both accuracy and latency. The perspective-guided summarization module in HMPPS is a practical blueprint for how to fix this. - Multimodal user history filtering changes the cost equation. The two-stage training paradigm reduces inference overhead by pruning irrelevant history items using multimodal similarity before the expensive MLLM forward pass. This matters at scale. - Production validation is the real benchmark. A/B testing on a billion-user platform is a stronger signal than any offline metric. The gap between academic benchmarks and production behavior is where most recommendation research quietly fails. What does your current personalized search or recommendation stack look like — are you conditioning on query

Join Hendrik Schwartz presenting SOMIN at INSEAD now!!! Drop by to learn more about our new integration with Manus AI , Zapier , and Anthropic to building marketing strategy for each of your GEO markets in 13 min and with just a click!

The era of constant martech churn is over. And most marketing leaders haven't adjusted their strategy yet. New data from the 2025 MarTech Replacement Survey tells a story that should reshape how every CMO thinks about their stack. Marketing automation replacement dropped from 31% in 2024 to just 19% in 2025. CRM fell to 9.7% — the lowest ever recorded. Email platforms nearly halved. This is not a blip. It is a structural shift. For years, the martech industry ran on a feature-driven replacement cycle. A new tool promised better capabilities, and teams switched. That logic no longer holds. The dominant driver today is efficiency and ROI — not novelty. Teams are asking "what more can we extract from what we already have?" rather than "what should we replace next?" The AI factor is particularly telling. 37% of respondents cited AI capabilities as important in replacement decisions — yet that interest is not triggering immediate action. It is doing the opposite: creating a wait-and-see posture. If AI-native platforms are coming, why replace now? This is rational behavior, but it also means many organizations are sitting on stacks that are underperforming while waiting for a wave that may arrive unevenly. What this means in practice: - The argument for any new tool must now be built on outcomes, not features. "Better" is no longer enough. "Measurably better" is the new bar. - Marketing ops teams are being asked to do more with existing infrastructure — which puts a premium on intelligence layers that sit above the stack, not inside it. - The brands that will win the next cycle are those investing now in understanding what their current tools are actually producing — creatively, commercially, and competitively. At SOMIN, we see this shift playing out directly with our clients. As stack replacement slows, the question becomes: how do you improve performance without changing the foundation? The answer increasingly lies in better intelligence — understanding which content concepts are working, which audience segments are underserved, and where creative quality is leaking budget. That is exactly the layer SOMIN operates in: above the stack, across channels, connecting creative decisions to commercial outcomes. The pause in replacement is not a sign of stagnation. It is a sign of maturity. The next competitive advantage will not come from the tool you buy — it will come from how deeply you understand what your tools are producing. What is your team doing differently in this efficiency-first environment? #MarTech #MarketingStrategy #CMO #MarketingOperations #SOMIN @SOMIN