Filling the gaps: derivation, inference, and the line that must not be crossed

Suppliers omit fields. Missing values get filled either by deduction from other fields or by an experienced person looking at the product. These are three mechanisms with different trust levels, and they must never be blurred.

Derived — a rule from a standard. ISO 13006 maps water absorption to group to porcelain/ceramic; an absent PEI implies an unglazed or through-body product. Deterministic and auditable. (Noting the previous chapter's finding: ETIM's own tile class cannot express the water absorption half of this.)

Inferred — vision or category norms. Colour family, look and style (wood-look, concrete-look, terrazzo), texture, pattern, room suitability. Probabilistic. This is the tacit industry judgment.

Asked — chase the supplier. The only route for anything unsafe to guess.

The hard rule: split the fields in two

Compliance and safety fields — slip COF or R-value, fire rating, frost resistance, load ratings, PEI — are measured test results. Never infer, never publish a guess. Leave blank and chase. A wrong value here is the liability event on the risk register.

Merchandising and discovery fields — colour family, look, style, texture, room, finish family, application tags — can be inferred freely. Wrong costs a slightly worse filter.

The reframe is that merchandising fields are exactly what suppliers never provide and what stores most need for filtering, SEO and comparison. Inference targets precisely that gap, which lands the idea back in digital-marketing value rather than data entry.

Missing fields also carry signal. PEI grades the wear resistance of a glazed surface; unglazed tiles, including many through-body porcelains, receive no PEI rating at all. An absent PEI is therefore evidence about the product, not merely a hole in the data.

And there is a limit on over-deriving: water absorption, PEI and COF/slip are distinct complementary attributes. Slip rating is a measured test result and is not reliably derivable from finish or appearance — which is exactly why the liability risk stays live.

The claim this page leaned on hardest, refuted

The original position was that vision-based enrichment is shipped but not in this vertical: the vendors target fashion, electronics and automotive, so applying it to tile and lighting vocabulary was open ground. The confidence register itself flagged this as "the single most load-bearing and least verified claim on the page" and classified it as an unverifiable absence.

It was verifiable, and verifying it broke it.

SKU Launch is a product of Start With Data, a Melbourne company. Its industry list names Building Supplies and DIY & Tools, and it does AI document parsing of supplier PDFs and spec sheets. Named case studies include Huws Gray ("product data enrichment for a leading builders merchant"), MKM Building Supplies, Castorama and Kingfisher Group.

The evidence was already in hand. startwithdata.co.uk had been cited twice in the source index — including an article titled "the 5 best PIMs for builders merchants" — without the connection being made to the same company that makes SKU Launch. Worse: the Australian angle, which this research treated as unexplored home turf, is already occupied by an Australian vendor selling into builders merchants.

Two statements had to go: that building materials are absent from these vendors' industries (false), and that the application to tile and lighting vocabulary is unoccupied (too strong — partly occupied).

Image-sourced attribute extraction is likewise already shipped:

• Pimberly ImageAI: "tagging objects in photos, identifying materials, recognizing shapes or patterns, assigning accurate metadata", with ColorAI standardising colour names plus hex. Its worked examples are fashion and homewares, but its customer list is construction — including Headlam.
• Akeneo Supplier Data Manager, AI Extraction: "Image assets can also be used as sources, provided they are of type media." Shipped, no building examples found.
• Catsy: "Computer vision to extract attributes directly from your product images", aimed at industrial manufacturers with thousands of technical SKUs.
• Lucidworks (20 January 2026): multimodal generative AI over product images plus text, generating categories, keywords, synonyms and descriptions, claiming an 8.66% conversion lift and USD 25M annualised for one retailer. A Lowe's association appears only in a third-party source and is unconfirmed.
• Hypotenuse AI enriches from images, spec sheets, PDFs and supplier feeds. One unnamed case study — a home-goods marketplace, USD 120M GMV, 80,000 SKUs — records 40%+ of SKUs missing material or dimension attributes, a useful data point on how severe supplier omission is in an adjacent vertical.

The primitives are sold generically too: Google publishes a Gemini retail product-attribute-extraction notebook, and Vertex AI Search for Commerce treats "materials" as a default system attribute. No building-materials-branded example from Google, AWS or Azure — the primitive is marketed, the vertical framing is not.

A whole competing axis that had been missed

Beyond the PIM vendors sit companies already machine-processing building-product imagery, purpose-built for it.

• Renoworks (TSXV-listed): 200+ building-product manufacturer and distributor customers, claiming "the most extensive catalog of digital building products and colors", each integration adding materials, colours and configurations that power its AI visualisation. An entire public company doing this for building products.
• Floori: flooring, tile and wall only. Digital twins of physical samples; extracts UV and PBR maps from ordinary photographs; upscales wood, epoxy, LVT and ceramic textures; exports BIM-ready.
• Roomvo / Leap Tools: automated flooring and tile catalogue across 250+ brands and 7,000+ dealers — porcelain, ceramic, glass, mosaic, floor, wall, backsplash. Four US patents.
• Tile-ID apps (TilesView, imageidentifier.ai, TilesDisplay) classify from a single photo: material (ceramic/porcelain/stone/mosaic), finish (glossy/matte/textured/polished), look (stone-like, wood-look), pattern, size, grout width. Consumer and dealer facing rather than PIM facing — but this is precisely the inference task called the surviving moat, and it is commodity.
• Also Toolbx, B2Sell ("AI Product Data Enrichment for Distributors"), Ximilar (Home Decor vertical), and Domus Image Search (ML image search for tiles and finishing materials, aimed at architects).

One caution worth weighing: somebody tried an adjacent market and retreated. Lily AI ran a home vertical in 2023 built on exactly this — vision tagging Fabric/Material (glass, wood, metals), Wood/Tone, and Material/Treatment/Finish (shiny, glossy, frosted, matte), naming Arhaus and Bridge Furniture & Props. That post now 404s, and lily.ai lists only apparel, footwear, beauty and luxury, with no computer-vision mention. Understand why before assuming the economics work.

(Material Bank does not do automated extraction — its tagging is user-applied and SmartMatch is curation assistance.)

What actually survives

The refutation cannot be softened into survival, because the claim's function was to establish that the inference layer was unoccupied. It has to be replaced.

The honest restatement: the vertical is well served on the PIM/syndication axis (Syndigo, Sales Layer, Pimberly, Bluestone, inriver, Salsify, SKU Launch — all with named building-products customers) and on the visualiser axis (Renoworks with 200+ manufacturers, Floori, Roomvo/Leap Tools, tile-ID apps). Nobody has joined them.

The defensible gap is that intersection — narrow but real. No vendor was found combining vision-derived surface, finish and look tagging with a standards-grade building-product attribute taxonomy. Across every vendor taxonomy checked, slip rating (R-value/PTV), PEI abrasion class, water absorption percentage, frost resistance, rectified edge and V-shade variation returned zero hits. The deepest published building attributes found anywhere were density, fire resistance, water resistance and dimensions (Sales Layer).

This makes the ETIM depth question the most important open item rather than a nice-to-have. The surviving thesis is no longer "we infer where others cannot" but "we infer into the compliance-grade schema the specification channel needs, which merchandising-focused vendors have no reason to build" — a thesis the adoption evidence in the previous chapter puts under direct threat.

Two positioning consequences follow. The visualiser companies are the more dangerous competitor set, not the PIM vendors: Renoworks owns the largest digital building-product library and Floori already extracts material properties from photographs. And Pimberly is the acute threat — it ships ImageAI and sells to Headlam, so it holds both halves and has merely not combined them.