Logitech's Award-Winning AI Share of Voice, Mastering Agentic Commerce with Typesense

In the third episode of Enterprise Aligned AI, host Aparna Sinha sits down with Deepa Shekhar, Director of E-commerce Technologies at Logitech, and Jason Bosco, CEO and Co-founder of Typesense, for a two-sided look at agentic commerce, the brand-side operator and the infrastructure founder.

Deepa’s Three Tracks of Agentic Commerce

Deepa offers a clean way to think about where this is heading. The three tracks are distinguished by two questions: where does the transaction actually happen, and who owns the consumer experience?

“I see agentic commerce evolving across three tracks, distinguished by where the commerce transaction takes place and who owns the consumer experience. We must execute on all three tracks to be successful.” — Deepa Shekhar

• Track 1: LLM Platforms as Strategic Engines of Commerce. Discovery and transaction journeys are moving from retailer websites to third-party LLMs like ChatGPT, Perplexity, Gemini, and Claude. The new KPIs are visibility and conversion inside LLM platforms, not website click-through rates.
• Track 2: Proprietary Brand Agents Grounded in Your Data. Building your own AI agents (in your app, your website, messaging channels) is more important than ever, in your unique brand voice enriched with your user and product data.
• Track 3: Pure Algorithmic Buying. Soon, personal consumer AI agents will shop on behalf of humans, and enterprise systems need to be ready. Agents rely on structured data and select products objectively based on operational parameters: granular specifications, shipping velocity, and historical fulfillment reliability.

AI Visibility Is the New Top of Funnel

Logitech won the “Growth Engine” award in Semrush’s inaugural 2025 AI Visibility Awards for Consumer Electronics, measured across thousands of real prompts in ChatGPT and Google AI Mode. Logitech.com was also named PCMag Reader’s Choice 2025 Best Manufacturer Online Store. The work to show up inside agentic surfaces is already producing a measurable funnel. This isn’t future state.

To win mindshare inside LLM answer windows, Deepa explains that marketing and technology teams must partner closely and jettison the keyword-based SEO mindset. Shoppers in LLMs ask intent-driven questions: “What is a good mouse for long working hours?” “What is a good mouse that works cleanly on a glass surface?” Logitech rewrote their content for those questions, and engineered it for machine readability with standards like llms.txt.

Composable Architecture Is the Prerequisite

Logitech’s shift to composable architecture wasn’t an AI play. It’s the reason agentic commerce is available to Deepa’s team today. When the front end, the search layer, the cart, and the content can each be composed and called by an agent, you can show up wherever the customer is, whether that’s the brand site, an LLM, or something that doesn’t exist yet. The infrastructure decisions made to “modernize” the stack now look like AI-readiness decisions. The work compounds.

Jason’s View from Underneath the Stack

Jason has a slightly counter-narrative point of view. The conversation has gravitated to LLMs and vector databases, but whether an agent can find the right product, in the right context, at sub-50ms latency, is decided in the search and retrieval layer. Typesense ships as a single C++ binary, open-source, revenue-funded, with 25,000+ GitHub stars and 12M+ Docker pulls behind it, and runs 10B+ queries a month on Typesense Cloud as the open-source alternative to Algolia and Pinecone. Jason explains what most teams get wrong when they reach for a vector database before they’ve gotten lexical retrieval right.

In Deepa and Jason’s recommended architecture, the LLM does intent detection (translating a query about “repetitive strain injury” into an ergonomic design match), then passes the extracted intent to a high-performance search engine that runs dozens of query combinations across millions of records in milliseconds. The result is more queries per second, better answers, and higher AI visibility.

The Missing Product Catalog Interface

One roadblock facing enterprises trying to plug into Agentic Commerce is that there is no unified mechanism to share product catalogs with LLMs. Deepa walks through the fragmented protocol landscape:

• OpenAI’s Agentic Commerce Protocol (ACP) provides structured capabilities for syncing catalogs, pricing models, and active promotions, but lacks native payment capabilities.
• Universal Commerce Protocol (UCP) defines standards for checkout, order updates, and payment handoffs, but only checks inventory at the final moment of checkout, which forces merchants to build parallel low-latency inventory APIs to avoid cart abandonment.
• Model Context Protocol (MCP) is not yet intuitive for scaled retail and functions more like a sandboxed app extension inside the LLM container.
• Order Exchange Protocol (OnX), championed by a consortium of modern OMS and e-commerce vendors, is aiming to introduce standardized, real-time eventing layers over legacy backends.

The industry needs a unified, cross-platform protocol that binds real-time inventory, batch product attributes, and multi-vendor financial transactions into a single standard. Until then, engineering teams face significant architectural friction.

Agents Are Already Making Product Decisions

Perhaps the best moment of the discussion: Jason recounts an autonomous agent (built to construct a personal knowledge repository) that was given autonomy to choose its own technology stack. It rejected a vector-only database, discovered Typesense, systematically tested its hybrid search and typo tolerance, and deployed it into production, documenting the whole evaluation in its own log files. After seeing this, Jason updated how Typesense exposes its technical assets, serving documentation as pure markdown alongside HTML, and embedding metadata that points crawling AI bots at the markdown versions so they don’t waste tokens parsing decorative web formatting.

Guardrails for the Buy Click

Agents executing financial transactions require financial integrity and system auditability. The primary defense against AI-driven transactional risk is automated, real-time evaluation layers, specialized LLM evaluators in production to dynamically audit conversation flows, monitor financial thresholds, and catch prompt injection vectors before anomalies register on the financial ledger. This is a domain that’s still early; the conversation barely scratches the surface of what’s coming.

Why This Matters Now

If you’re running e-commerce, the question isn’t “should we have an agentic strategy?” anymore. The question is which of Deepa’s three tracks you’re investing in, and whether the layer underneath your store (the part Jason cares about) can actually serve those agents at the latency and relevance bar they expect.