In our previous blogs we explored how unifying different modalities creates simpler, better, and more customisable user experiences. A shopper can search with text, refine with images, and get personalised results without hitting the walls that separate traditional search, recommendations, and browse experiences.
But at Solenya we think the effects of this unification go far beyond just user experience and ease of integration, both for us, our customers, and the industry as a whole.
Invisible Walls
It may seem like the idea of a unified model is just a technical detail, an implementation choice hidden behind APIs and abstractions, but the model is everything. PageRank, an algorithm for measuring a website’s relevance on the internet, didn’t just improve search results; it shaped the web, and had profound effects on every business and industry. MP3 Compression didn’t just make music files smaller; it changed how we consume, distribute, and make music. The point is: we should stop thinking about this as just a new model which improves your search, but as a fundamental shift in e-commerce.
Product discovery in e-commerce has simply not kept up with the advances in AI. Over the last few decades we’ve built some “invisible walls”, the effects of which can be felt throughout e-commerce, but not necessarily understood.
A Unified Model is Unified Discovery
The current paradigm in e-commerce is fragmented. Different use-cases, like text search, visual search, and recommendations are handled by different models, often from different providers. Each model has its own architecture, its own API, and its own data requirements, but at the end of the day, it’s all one problem: ranking. Whether it’s ranking products based on how well they align with a text query, or a text query and an image, or a user’s preferences, it’s all about finding the most relevant products for a given combination of inputs. So why not use one model?
This fragmentation creates invisible walls that limit what’s possible. It not only creates a confusing set of integrations for customers, but also hinders advances in the model training and research required to push the state-of-the-art. It also opens up new possibilities downstream in analytics and marketing, and makes new features like AI shopping assistants possible.
The Ripple Effects
A Composable Interface
A unified model allows for a composable API where use-cases aren’t predefined and constrained. Instead, users can mix and match modalities in ways that weren’t possible before. The same interface can be used for everything from search and recommendations, to MCP tools, to customer understanding and analytics. There may be use-cases in e-commerce we haven’t even thought of yet which are naturally handled by this unified approach.
R&D
Changes to model weights or architecture benefit every use-case simultaneously, rather than requiring coordinated updates across multiple systems. This massively reduces the test surface, focuses research energy, simplifies model training, and makes evaluation more straightforward, all leading to faster iteration cycles.
A unified model also means a unified team. In a fragmented architecture, different parts of the stack often have different owners. The search team owns the search system, the recommendations team owns the recommendations engine, and when a use case spans both, you need coordination meetings and shared roadmaps. With a unified model, there’s one system to improve, which means clearer ownership and better alignment.
Infrastructure and operations become simpler too. You’re deploying one model instead of many, which means one set of serving infrastructure to optimize, one monitoring dashboard to watch, one set of performance metrics to track. When you need to scale, you’re scaling a single service rather than coordinating capacity across multiple vendors. Caching strategies become more effective because product embeddings are reused across all discovery surfaces rather than being specific to one use case.
Marketing, Sales, Analytics
This is where you can really start to see the ripple effects. Marketing, sales, and analytics have always been divorced from search and recommendation. But if the model understands user preference, what opportunities does that open up for marketing teams to create personalised campaigns and get new, deeper insights into customer behaviour?
Traditional solutions create data silos that are hard to integrate, but a unified model breaks them down. Analytics teams can start to answer questions that were previously impossible, analysing how users interact with the entire discovery experience instead of looking at search and recommendation data in isolation. Which combinations of modalities lead to higher conversion rates? How do different user segments respond to different types of inputs? This holistic view leads to better segmentation, more targeted marketing, and ultimately, higher conversion rates.
A/B testing becomes more coherent as well. Instead of testing “the search system” versus “the recommendations system,” you’re testing variations of the unified model or changes to how modalities are weighted. Your experiments aren’t fragmented across different vendors with different testing methodologies, they’re all happening in the same framework with consistent metrics.
AI Assistants and Beyond
The idea of conversational shopping assistants has been toyed with for years, but has only recently become feasible with advances in Large Language Models (LLMs) and the rise of MCP, a protocol allowing AI agents to interact with existing tools (like search). OpenAI recently announced their Agentic Commerce Protocol (ACP), enabling users to shop via ChatGPT itself. Algolia also announced Ask AI earlier this year, allowing agents to interact with existing indexes. Daydream provides a conversational shopping assistant which searches across retailers.
Access to state-of-the-art LLMs is no longer the bottleneck; the challenge now is integrating these models with existing fragmented e-commerce architectures. In the past, a complex and fragmented API was manageable because integrations were static and could be rigorously tested before deployment. However, with AI agents making dynamic calls to your API, a simple, unified interface becomes crucial. Complex schemas with many different endpoints and authentication methods become a nightmare for agents to navigate, leading to brittle integrations, frequent failures, and ultimately poor user experiences.
Similarly, current e-commerce frontends are built around the established limitations of fragmented architectures. The entire layout is designed to separate search, recommendations, and browse experiences. The interface guides users to use endless filters and facets to narrow down their options because the underlying architecture can’t handle more complex queries that describe user intent. Latency has to be incredibly low, because users need to repeatedly engineer the perfect search term or combination of filters to find what they’re looking for. But with a unified model, these limitations disappear. The frontend can become more conversational and exploratory, mixing modalities and allowing for instruction-based searching. Latency is less critical if users can express their intent more clearly and trust that the underlying model understands it.
The Future
A unified model is more than just a technical improvement; it’s a fundamental shift in how e-commerce operates. It breaks down invisible walls put up by fragmented architectures, enabling new possibilities in R&D, marketing, analytics, and user experience.
The benefits mentioned here are just a few, limited by our current understanding of what’s possible. As the e-commerce industry continues to evolve, we believe the shift to unified models will have profound effects that we can’t yet predict. The question is: will you be ready to take advantage of them?