The catalogue becomes a model
3 of 8 items
Multimodal product ranking, embedding-based retrieval at scale, and identifiers generated rather than looked up. A query lands in a learned space, not in fields and filters.
Focus / COMMERCE & MARKETPLACES
The catalogue stops
being an index.
Product search is becoming a learned model rather than a structure to query. Business value is moving inside that model. And the structural work that makes a store readable to assistive software is the same work that makes it readable to a buying agent.
ContinueRead the thesisThe thesis
MANFRED / 01
The durable work in commerce is no longer the funnel. It is making the catalogue legible: to the model that ranks it, to the regulator that audits it, and to the agent that buys from it.
Three research directions and one compliance measurement point at the same structural layer. None of them is a storefront feature. All of them determine whether a store can be read correctly by something that is not a human eye.
What Manfred is seeing
A narrow window, but a consistent direction.
Bounded observation window · not a full field snapshot and not a market-size forecast
Observation window 10 to 14 July 2026 in the commerce stream: eight items, comprising seven scientific preprints and one industry measurement circulated through Hacker News. No topic distribution is computed. At this sample size a percentage would be decoration, so the cards below report how many items sit behind each reading.
Margin moves inside the ranker
2 of 8 items
Business value encoded into document identifiers, and discount sensitivity modelled inside conversion prediction. The rule layer that used to reorder results afterwards is being absorbed.
Forgetting becomes a function
1 of 8 items
Post-purchase satiation modelled over long behavioural sequences. The least spectacular item in the window and probably the shortest path from paper to deployment.
Two circuits, one requirement
2 of 8 items · independent sources
A compliance measurement and an academic agent-readiness framework describe the same technical surface without citing each other. This is the only reading in the window that no single stream could produce.
Interpretation: The common motif is not a new model family. It is the layer that decides how a catalogue is read: how products are represented, what the ranking is optimising for, what the system remembers, and whether the interface can be parsed at all by software acting for a person. A screen reader and a buying agent need the same structural properties, because both are programs reading an interface designed for an eye.
What is changing now
The channel, the standard and the obligation arrived before the compliance did.
Public moves in products, standards, policy and measurement give external context to the window above. They are selected because they bear directly on whether a catalogue can be read by a machine.
These events do not prove the reading. They show whether the field is beginning to organise itself around the same constraint.
Company signal
Purchase moves into the assistant.
OpenAI describes Instant Checkout, in which a buyer completes a purchase inside the conversation. The seller keeps the merchant relationship but stops owning the surface on which intent is formed.
Open standardAgent access acquires a specification.
The Agentic Commerce Protocol is published as an open standard defining how agents initiate and complete checkout with a business. The signal is not that one protocol wins, but that agent-facing commerce is being treated as a shared layer.
Regulation watchThe obligation is already in force.
The European Accessibility Act covers e-commerce services and has applied since 28 June 2025. It sets requirements on how a service is presented and operated, not on what it sells.
Field measurementJust over half the market has not done the work.
AccessLumens reports that fifty two percent of EU e-commerce sites still carry a critical accessibility barrier, more than a year after the obligation took effect. Read as a measurement of machine readability, that figure describes the same population.
Company announcements are treated as market signals, not independent performance validation. The observable pattern is that a channel, a specification and a legal obligation now all require the same property from a storefront, and that just over half of the measured population does not yet have it.
What the evidence may suggest
The question is no longer how a store looks. It is how it reads.
Catalogue quality becomes model quality
Once retrieval is learned rather than matched, product data stops being database hygiene and becomes training input. Sparse attributes and inconsistent taxonomies no longer degrade a filter, they degrade a representation.
The optimisation target changes
Moving margin and discount sensitivity into the model removes the separation between what the system predicts and what the business wants. It also removes the place where a human used to notice the difference between a profitable week and a damaged price expectation.
Compliance and agent readiness share a budget
Semantic structure, labelled controls, predictable interaction order and deterministic checkout states satisfy an accessibility audit and an agent integration at the same time. Treated as two projects, this work is paid for twice.
What could challenge this view
A thesis becomes useful when it can be tested.
This reading weakens if agent-mediated purchasing stays a marginal share of transactions, in which case machine readability remains a compliance cost with no commercial return. It weakens further if agents become capable enough to parse badly structured pages reliably, which would remove the shared requirement entirely. The window also contains no patent stream and no capital stream, so nothing here establishes whether these directions are early or late. Scientific publication reports what is being studied, not what is being deployed and funded.
Evidence register
The source layer stays visible.
The interpretation on this page is Manfred's. The records below separate the internal observation window from the external sources used to test and contextualise it.
Manfred
Research
Commerce observation window
10 to 14 July 2026Eight items in the commerce stream: seven scientific preprints and one industry measurement. One further item in the window, on financial anomaly detection through multi-objective reinforcement learning, is recorded as classification noise rather than a topical signal. This window shows what the system retrieved in five days; it is not a field snapshot and not a market-size measure.
●MMRM · multiplex multimodal product ranking
13 July 2026A multimodal representation model for product ranking in e-commerce search. Used as evidence that ranking is being treated as a representation problem.
ResearchScaling and stabilising embedding-based retrieval
11 July 2026Engineering work on large-scale embedding retrieval. Used as an indicator that the architecture has settled far enough for effort to move to operation.
ResearchBeyond Semantic IDs · business value in document identifiers
13 July 2026Encodes business-value ranking into the identifiers used by generative retrieval. Used for both the retrieval and the economics readings.
ResearchCheaper is Better · discount-aware conversion prediction
14 July 2026Models discount sensitivity inside conversion rate prediction. Used as evidence that the optimisation target is shifting from behaviour to economic outcome.
ResearchLearning to Forget · satiation-aware transducers
14 July 2026Treats post-purchase redundancy as a modelling problem over long sequences rather than as an exclusion rule.
ResearchDesigning agent-ready websites
13 July 2026A framework built on machine readability, actionability and decision reliability. Used as one half of the convergence reading.
MeasurementAccessLumens · State of EU e-commerce accessibility 2026
14 July 2026Industry measurement reporting a critical accessibility barrier on fifty two percent of EU e-commerce sites. Used as the other half of the convergence reading. It is a vendor-published study and is treated as a measurement claim, not an audited statistic.
RegulationEuropean Commission · European Accessibility Act
Applies since 28 June 2025Directive (EU) 2019/882 covers e-commerce among the services in scope. Cited for timing and scope. Nothing on this page is legal advice.
BuildOpenAI · Instant Checkout in ChatGPT
Company announcementDescribes in-conversation purchase and the merchant-side requirements around it. Used as a channel-architecture signal, not as evidence of transaction volume.
StandardAgentic Commerce Protocol
Released 29 September 2025Open specification for agent-initiated checkout, maintained by OpenAI and Stripe. Used as evidence of coordination, not of universal adoption.