From keyword commerce to agentic commerce
The commentary by James Taylor describes a structural shift in how retail discovery and monetisation are built: from keyword‑driven, page‑centric ecommerce to agentic commerce, where AI systems act on behalf of users and interface directly with product data and commercial logic.
The core claim is that Amazon‑level personalisation and discovery no longer require Amazon‑scale budgets. Instead, retailers can assemble a modular stack centred on transformer‑based semantic search, a decisioning layer that governs relevance and monetisation, and Model Context Protocol (MCP) as the standard way to connect AI models with first‑party catalogues, feeds and tools. In this architecture, ads in AI search become functional shopping experiences, and the retailer’s governance layer – not a third‑party model – decides what is shown, on what terms, and with what economic outcome.
This vision appears against a broader backdrop: major AI providers are moving towards tool‑using, “agentic” models that can call external APIs, transact, and optimise on user intent rather than on isolated queries. Industry discussions increasingly treat product catalogues, retail media networks and MCP‑style interfaces as the primary surface through which ecommerce will be exposed to these agents. The debate is shifting from “how to get traffic from AI search” to “how to control what AI agents can do with my inventory and margins”.
Why the shift matters: intent over keywords
Traditional ecommerce search has been built around exact‑match or close‑match keyword logic. Taylor points out that this architecture systematically misses intent: a query like “how to reduce wrinkles” may not match any product titles or attributes, even though it clearly refers to anti‑ageing skincare. In his estimate, exact‑match keywords miss roughly three‑quarters of true intent.
Agentic commerce assumes a different primitive: semantic understanding of user goals. Transformer‑based, vector search models map queries and catalogues into the same embedding space, allowing them to recognise that “anti-ageing cream”, “aging cream” and “reduce wrinkles” point to overlapping solution sets, even when the wording differs.
At the same time, recommendations move from segment‑based heuristics to behavioural intent. Rather than assuming that all shoppers in a demographic cohort want similar things, the system looks at session‑level signals: search terms, products viewed, sequence of interactions before adding to basket, and the downstream behaviour of comparable users. Collaborative filtering and real‑time event streams allow the system to respond to “what this person is trying to do right now”, not to “who this person is in general”.
This change is critical for AI agents. Agents interpret free‑form user instructions (“find me a cruelty‑free anti‑wrinkle routine under $100”, “rebuild my running kit for a marathon in a cold climate”) and expect the underlying commerce layer to resolve them into actual items and offers. Exact‑match search cannot deliver robust coverage for such open‑ended, long‑tail requests; transformer‑based semantic search can, provided the underlying data and governance are in place.
MCP as connective tissue between AI and commerce
A central element in Taylor’s argument is Model Context Protocol (MCP), described as an open‑source standard for connecting AI models with external tools and data sources. In practice, MCP plays three roles:
- It defines how large language models discover, authenticate and call external capabilities (“apps”) such as search, pricing, inventory or cart APIs.
- It structures how product data, attributes and commercial rules are exposed to AI systems in a controlled, machine‑readable way.
- It standardises a “handshake” – a secure, auditable negotiation of what the model may access and do on behalf of a user.
By building MCP‑compliant integrations, retailers can allow AI search interfaces to open what are essentially interactive shopfronts, not static ads. An “ad” in an AI environment becomes an embedded shopping flow: invoke the retailer’s MCP app, run a transformer‑based search with governance rules, display options, and potentially transact, all without leaving the AI interface.
Crucially, Taylor emphasises that retailers should own this layer themselves. If a retailer simply dumps a product feed into a third‑party LLM without an intervening decision engine, the model – and the platform operating it – effectively becomes the auctioneer. In that scenario, the retailer’s catalogue is just one more piece of inventory bidding for attention in someone else’s marketplace. Owning the MCP layer means that every AI‑initiated discovery or transaction still passes through the retailer’s own relevance, merchandising, and margin logic.
The decisioning layer: governance for AI‑driven discovery
Transformers and vector search improve relevance, but they do not, by themselves, align outcomes with commercial strategy. Taylor positions the decisioning layer as the missing link: a governance surface that combines several streams of logic:
- Semantic ranking and retrieval, based on transformer embeddings.
- Merchandising rules (boosting, pinning, exclusion, seasonal logic).
- Margin and yield considerations (prioritising higher‑margin items when relevance is comparable).
- Sponsored placements and retail media campaigns.
- Personalisation signals derived from user behaviour and context.
In practice, this means that each AI‑driven request produces a ranked list of products shaped by both user intent and business rules. The “relevance curve” that results must be robust enough to stand up to AI scrutiny, because agents will test and compare outcomes over many sessions and users, optimising towards their own goals such as price, quality or delivery time.
The same layer underpins retail media. Keyword bidding in its manual form becomes untenable when queries are free-form, multi-constraint and long-tail. Taylor cites a test where an Australian pet supplies retail media network replaced manual keyword bidding with transformer‑driven search bidding; the share of “performant” search queries with monetisation coverage reportedly increased four‑fold. This suggests that semantic search can surface monetisable intent much more broadly than human‑maintained keyword lists can.
Implications for product feeds and cataloguing standards
The agentic commerce stack assumes that product data is both machine‑readable and semantically rich. MCP can only expose what lives in the retailer’s systems, and transformer search can only interpret what is encoded in the catalogue. This has several concrete implications for content infrastructure:
Attribute quality becomes foundational. Descriptive, normalised attributes – ingredients, materials, sizes, fit, function, use cases, certifications, compatibility, and so on – allow models to map products into embedding spaces that reflect real‑world properties. Sparse, inconsistent attributes limit the model’s ability to match complex queries or to respect constraints (for example, “fragrance‑free, paraben‑free moisturiser”).
Taxonomy and ontology design matter more than ever. Category trees, product types and relationship structures (variants, bundles, accessories, substitutes) need to be coherent and stable. While transformers can compensate for inconsistent naming, they cannot invent a hierarchy that does not exist.
Unstructured content needs structure. Descriptions, FAQs and reviews carry rich signals but are often noisy. Retailers increasingly use AI to extract attributes and normalise terminology from this content into structured fields, which then feed semantic search and MCP apps.
Media assets become part of the semantic layer. Images and videos are now routinely embedded by multimodal models; clear alt text, captions and tagging increase their utility for search and recommendations, and for agents that want to verify visual aspects of products. In practice, investing in cataloguing standards is less about SEO in the narrow, page-ranking sense and more about making the catalogue intelligible to a growing ecosystem of AI agents. The same structured feeds that power ads and marketplaces must now be able to power conversational, task-oriented interactions.
Product cards: completeness as a precondition for prediction
Taylor frames personalisation as “just good prediction”. For prediction to work at scale, product cards must be complete, consistent and kept up to date. The pressure here is two‑sided:
On the discovery side, missing attributes, outdated images or ambiguous titles reduce the probability that semantic search will retrieve the product for relevant queries. If the system cannot distinguish between similar items, it may default to safer, better‑described alternatives.
On the monetisation side, incomplete commercial metadata – margin, promotion status, co‑op funding eligibility, stock thresholds – weakens the decisioning layer. The engine cannot reliably identify optimal candidates for sponsored or high‑margin placements.
Agentic commerce adds an additional constraint: AI agents will increasingly benchmark results across sources. If one retailer systematically offers clearer, richer product representations – ingredients lists, sizing guidance, compatibility data, environmental or ethical indicators – agents have more evidence to justify recommending its inventory. Thin or templated content, once a tolerable compromise, becomes a competitive liability.
This dynamic is amplifying investment in content operations: automated attribute enrichment from supplier data, large‑scale image standardisation, AI-assisted copywriting with human review, and continuous quality monitoring. The goal is not just to “have a product page” but to produce a machine‑optimised representation that supports reliable inference.
Speed to market: automation across the catalogue lifecycle
The article indirectly highlights another effect of agentic commerce: speed of assortment expansion becomes constrained by the slowest manual step in the catalogue pipeline. To fully exploit transformer search and MCP integrations, new products must be onboarded with high‑quality data from day one.
Retailers are therefore re‑architecting catalogue workflows around automation:
Supplier feeds are normalised and validated automatically, with AI models mapping disparate attribute schemes into a unified schema.
Gaps in mandatory attributes are flagged in real time for suppliers or internal teams, often with suggested values generated from packaging images, spec sheets or similar items.
Initial titles, bullet points and descriptions are drafted by models and reviewed by editors, cutting time‑to‑publish while keeping editorial oversight.
Category assignment and variant grouping are semi‑automated using clustering and similarity models, reducing misclassification and orphaned products.
When such pipelines are in place, MCP‑exposed apps can instantly incorporate new SKUs into AI‑driven discovery and advertising. Without them, there is a lag during which the retailer’s own catalogue is invisible to agents for many high‑intent queries – a direct loss of revenue and of training signal. To understand how to properly prepare and upload the the product information, you may consider reading our article, "How to upload product cards".
No‑code, AI and the democratisation of retail media architecture
Taylor stresses that “Amazon‑style” personalisation is available to retailers without Amazon‑level budgets, provided they adopt modular retail media platforms and standards like MCP. This reflects a broader trend: many components of the agentic commerce stack are now accessible as services or no-code modules rather than as bespoke in‑house builds.
In practice, this means:
Vector search and recommendation engines can be integrated via APIs, tuned by configuration rather than custom research.
MCP adapters and connectors can be implemented once and reused across multiple AI partners, reducing integration overhead.
Business teams can define merchandising rules, margin priorities and campaign logic through graphical interfaces, with changes propagated to the decisioning layer without code deployments.
Predictive bidding and budget allocation for retail media can be automated using models that optimise towards ROAS or other KPIs, freeing specialists to focus on strategy and creative.
The constraint shifts from engineering capacity to data discipline and governance. Retailers that can maintain clean feeds, coherent taxonomies and clear commercial rules will be able to plug into agentic ecosystems with relatively modest technical effort. Those that cannot will find that no amount of no-code tooling can compensate for poor underlying data.
Strategic implications for ecommerce and content infrastructure
Taken together, the developments described in Taylor’s piece outline a new reference architecture for ecommerce in the AI era:
- Discovery is mediated by transformers and agents rather than by static SERPs and exact‑match search boxes.
- The retailer’s primary asset is not just inventory, but the decisioning layer that controls how that inventory is exposed to AI systems.
- Product data and content are redefined as inputs to machine prediction, not only as human‑readable marketing materials.
- Retail media becomes inseparable from search and recommendation; monetisation logic is embedded directly into relevance algorithms.
- Standards like MCP ensure that, as AI interfaces proliferate, retailers can connect once and distribute many times without ceding control.
For content and catalogue teams, this raises the bar. Their work now underpins not only brand experience and conversion, but also the retailer’s ability to be “understood” – and chosen – by a growing class of autonomous agents. In this environment, investing in structured data, semantic search, and a robust decisioning layer is less an optimisation and more an operating requirement for participating in agentic commerce at all.
The shift towards agentic commerce, as highlighted in the article, underscores the critical importance of high-quality, structured product data. At NotPIM, we recognize this as the foundation for success in the evolving e-commerce landscape. Our platform empowers retailers to address these challenges head-on by streamlining data transformation, enrichment, and catalog management – enabling them to provide AI agents with the detailed, consistent product information they need to effectively drive discovery and sales.