Structured product data for independent makers, published so the open web — and the agents that read it — can find them without going through a marketplace.
Independent retailers don't need better marketplaces. They need their inventory to be legible to systems that don't belong to a marketplace. Structured data, owned by the seller, will route more honest demand than algorithmic surfacing inside a closed marketplace.
SPF predates the publication and its framework, but the test it runs is one the framework would have proposed. The structural argument is general: any system whose value flows through a closed legibility layer will, over time, extract rent from the participants on both sides of the layer. Marketplaces are one specific instance. They are not unique.
For independent retailers, the closed legibility layer is the marketplace itself — Amazon, Etsy, Shopify’s surfaces. The marketplace reduces a seller’s catalog into a schema it controls, then surfaces that catalog through ranking the seller does not control. Two layers of control, both held by an entity whose interests are not aligned with the seller’s.
The structural alternative is to make a seller’s data publicly legible at the source, in formats every reader — search engines, AI agents, comparison tools, third-party guides, the next generation of agentic commerce — can parse without paying the marketplace. The data sits at a URL the seller owns. There is no platform; therefore no rent.
62 makers currently active, in three product categories: handmade furniture, specialty pantry goods, and small-batch beauty. All operate independently and have a marketplace presence to baseline against.
Each maker maintains a structured-data feed at their own domain, conforming to a public schema designed to be read by search, AI agents, and third-party tools without authentication.
Identical SKUs listed on at least one marketplace, with identical pricing and copy. Discovery and conversion are measured side by side over rolling 12-week windows.
The design is intentionally narrow. SPF doesn’t claim that every seller benefits from leaving a marketplace; it claims that describably-distinct sellers — those whose differentiation can be expressed in attributes rather than imagery alone — can outperform on a per-unit-effort basis. The test holds that variable constant by limiting participation to sellers in that population.
Four measurements form the test:
Discovery-to-purchase conversion, for identical products published in both places. Time-from-publish to first agent reference, where an agent is any external system (search engine, AI assistant, third-party tool) that cites the product. Per-seller take-home margin on the same product, after marketplace fees on one side and SPF’s flat per-seller cost on the other. Time-to-first-sale on a new SKU, which proxies how quickly the data routes new inventory to the right buyers.
None of the measurements is a vanity metric — none is “traffic” or “engagement” or “follower count.” Each is tied directly to whether the seller ended up with more money, with the same effort.
The numbers above are eight months in, across the 62 active makers. They are real but the sample is narrow. They should not be treated as a generalized claim yet.
When the experiment started, the assumed primary reader of an SPF feed was search engines. Search is still important. The surprise is how quickly AI assistants — Claude, ChatGPT, Perplexity — have become substantial referrers. The feeds were already structured the way they want to read; we did not redesign for them.
I expected the strongest results in categories where the seller has the most to say about each product — meaning furniture, with its long copy and material distinctions. Pantry goods, where the copy is shorter, has outperformed. Hypothesis: structured data favors items where attributes are discriminating, not items where they are merely numerous.
Not because SPF didn’t work, but because the marketplace they were leaving sued them — and the cost of fighting the suit was higher than the SPF margin uplift. Filed under: the closed layer extracts rent even on the way out. Predictable in retrospect; should have been planned for.
Three months in, two large marketplaces began offering “structured catalog export” features that look unmistakably like SPF’s schema, in a clear attempt to make leaving the marketplace less attractive. This is, in a perverse way, the strongest signal yet — they are now treating the absence of structured exports as a competitive vulnerability.
Three makers, three categories, a hand-rolled schema, and an unreasonably long Saturday. The hypothesis is on paper. The test will run for at least two years.
The findings below are provisional. The experiment is not concluded. Each of the lessons here is what I currently believe; each is also something I would revise without much resistance if a year of additional data disagreed.
Three moves are queued, in this order. First: widen the categories. The current three are too narrow to support generalization. Adding two more in the second half of 2026 — leaning toward apparel and small-format consumer electronics, both of which have notably distinct attribute structures.
Second: build the exit-cost model. The maker who left taught us that leaving a marketplace is itself a separate cost the experiment should model. Future intake will price that cost into the SPF onboarding decision rather than discover it after the fact.
Third: publish the schema. SPF’s structured-data format has been refined through eight months of use. It is ready to be opened to anyone, including makers who never join the program, on the theory that the schema’s value is in being adopted, not in being controlled.
First three makers onboarded with a hand-rolled structured-data schema across furniture, pantry goods, and small-batch beauty. Two-year hypothesis test begins: factory-direct schema for makers, owned by the seller, can outperform marketplace surfacing on per-effort discoverability for describably-distinct categories.
3 models for this experiment, side by side. Each is a complete business model in its own right.