Every projection of $15 trillion in agent-mediated B2B commerce assumes catalogs are machine-readable. They aren’t. In a live census of Catalist’s own ingestion telemetry, 41.9% of supplier catalogs arrived as PDFs, 43 distinct column names referred to wholesale cost, and across 17,863 brand websites only 71 — 0.4% — published a wholesale page. The buy side has shipped AI procurement agents. The sell side still ships spreadsheets. The arbitrage of the next five years sits in the layer that translates between them.
The Supply-Side Internet Does Not Exist
Most coverage of agentic commerce describes the buyer-side stack — procurement agents that anticipate need, query suppliers, compare options, place orders. Gartner forecasts $15 trillion in B2B spend flowing through these agents by 2028. Kearney calls it agentic commerce. McKinsey puts the addressable revenue at $3 to $5 trillion globally by 2030.
Catalist sits on the other side of that interaction. Of 17,863 brand websites Catalist has audited with automated analysis, 11,796 have a public consumer site. 71 have a public wholesale page where a buyer can see pricing, request a catalog, or place an order online without first sending an email. That is 0.4 percent.
The other 99.6 percent of suppliers transact the way they did in 2005: a wholesale inquiry form gets a sales rep, the sales rep sends a price sheet attached to an email, the buyer parses the sheet manually, and the order is placed by reply-all. No API exists. No public catalog exists. The agent the buyer just deployed has nothing to query.
This is the data problem that every agentic-commerce thesis quietly assumes away. The buy-side has shipped. The sell-side has not.
0.4 percent. The supply-side internet doesn’t exist yet.
Catalist supplier-directory AI analysis, May 2026. Sample: 17,863 brand websites in Catalist's outreach pipeline.Inside 191 Catalogs From 128 Suppliers
When suppliers do send a catalog, the question becomes what an agent receives. Catalist’s ingestion pipeline has captured 191 raw catalog files from 128 distinct suppliers — first-hand, brand-direct deliveries, not scraped or licensed from a data vendor. The format mix:
Source: Catalist supplier_documents table, 2026-05-22. Includes catalogs delivered via email attachment, supplier portal upload, and direct file share.
41.9 percent PDFs. Functionally opaque to a procurement agent without OCR plus LLM extraction — and even then, prone to mis-parsed prices, missed page footnotes, and concatenated unit fields.
51.8 percent Excel. Better than PDF, but Excel files routinely arrive with multi-row headers, embedded merged cells, color-coded pricing tiers, hidden formula references, and free-text product descriptions in numeric columns. Catalist’s parser hint table records suppliers whose catalogs require skipping the first ten rows of branding before the data begins.
4.7 percent CSV. Zero percent JSON, XML, or API. Across 191 first-hand catalog deliveries, no supplier shipped a structured machine-first format. None published an API endpoint a procurement agent could call.
Every entry below appeared on a real supplier’s catalog header in Catalist’s ingestion log. Catalist maps each one to a canonical role — the price a buyer pays (cost), the MSRP, or the MAP enforcement floor.
Source: Catalist supplier_column_overrides table, 2026-05-22. Includes only catalogs where automated parsing required a manual override — the hard-tail subset of the corpus.
43 distinct column names across 47 suppliers — all referring to the same field. "Base." "Net." "WHL." "DEALER." "Level 3 SWIM." "Regular RSC Purchase Price Ea." A buyer’s agent receiving these catalogs has to decide which one means cost-to-buyer, which one is MSRP, and which one is the MAP enforcement price. Get it wrong and the order ships at retail.
The same pattern holds for retail pricing — 18 distinct column names ("MSRP," "SRP," "Reg Price," "Suggested Retail," "Retail USD") — and for MAP enforcement, where five different headers ("MAP," "MAP Pricing," "US MAP," "2026 MAP," "MAP 03/2026") all signal the same advertising floor.
Step back and the schema variance compounds. In Catalist’s 47 hardest catalogs, the ingestion pipeline recorded 53 distinct header signatures — the unique combination of columns, order, and naming a given supplier shipped. That is more than one unique schema per supplier. The supply side is not a market with a few formats. It is a market where every catalog is bespoke.
What This Costs a Procurement Agent
The visible cost of catalog fragmentation is onboarding latency. Distribution Strategy Group puts the industry norm for onboarding a new B2B supplier at four to six weeks — most of that time spent on catalog ingestion, column mapping, unit reconciliation, and data validation. Catalist’s normalized pipeline collapses that window to three to five days because the parsing layer is already built.
The invisible cost is everything an agent has to do, per supplier, to make a single quote-and-order round-trip work:
Format ingestion
Detect whether the file is PDF, Excel, or CSV. PDFs need OCR plus structured-extraction prompts. Excel needs row-offset detection, merged-cell handling, and multi-tab traversal.
Column inference
Map "WHL," "DEALER," "Net," or "Level 3 SWIM" to the same canonical cost field. Distinguish cost from MSRP from MAP — three different prices on the same row, sometimes labeled identically across files.
Unit reconciliation
Determine whether the listed price is per unit, per case, per pack, or per pallet — usually inferable only from a separate "case pack" column, a parenthetical in the description, or a footnote. Get this wrong and margins move twenty points.
Identifier resolution
Match the supplier’s internal SKU, UPC, or part number to a canonical product. Suppliers reuse the same identifier for case packs and singles. UPCs go stale. MPNs drift across model years.
Build that pipeline once and you get an inert dataset. Run it continuously across every supplier’s next-quarter catalog and you get a live one. Catalist’s ingestion layer currently keeps 62,963 of 64,044 SKUs refreshed inside a thirty-day window — possible only because the normalization runs as infrastructure, not as a one-off integration per supplier.
The implication for a buyer deploying a procurement agent today is direct: the agent is only as good as the supply-side data it can reach. Either the agent builds the catalog-normalization layer itself, in which case the agent is a data-engineering project that occasionally places orders, or it queries a connector that has already done that work.
"Every founder pitching an agentic-commerce platform shows the same demo: an agent that researches a product, compares suppliers, places an order. The demo always uses three suppliers with clean APIs. The actual long tail of B2B procurement is 17,863 brand websites where 0.4 percent have a wholesale page. That gap is the whole game."
Devin Sardana · Founder, Catalist AICatalist has normalized catalog data across 1,343 brands and 64,000+ SKUs into a single query layer for B2B procurement.
The Bet
Two predictions follow from the data above.
One: agentic procurement at the long tail of B2B SKUs is supply-side-constrained, not buy-side-constrained. The model improvements forecast for 2026 and 2027 will not change the fact that 99.6 percent of suppliers have no public wholesale endpoint and 96 percent of catalogs arrive as Excel or PDF. The bottleneck moves from "can the agent reason about a purchase" to "can the agent see the inventory at all."
Two: the supply-side connector becomes a structural position. Whoever owns the normalization layer between heterogeneous supplier data and the buyer-side procurement agents earns a margin that compounds with every new supplier onboarded. Catalist currently runs that layer for 1,343 brands and 64,000+ live SKUs. The corpus grows weekly. The connector position is not theoretical — it is operating.
The agentic-commerce projections describe a world where AI agents do most of the buying. The data from the supply side suggests the rate-limiting step is not the agent. It is the catalog.
Methodology
All Catalist numbers in this report are queried from production data on 2026-05-22. The raw counts and SQL queries are reproducible from the snapshot in docs/research/catalog-fragmentation/raw-stats.json in the Catalist website repository.
| Evidence | Claim 01 Only 0.4% of 17,863 audited B2B brand websites publish a public wholesale page. | Claim 02 41.9% of supplier catalogs arrive as PDFs and 0% arrive as JSON, XML, or API. | Claim 03 43 distinct column-header names map to the canonical wholesale-cost field, with 53 distinct full header layouts across the manual-override subset. |
|---|---|---|---|
| Methodology | Automated AI analysis of supplier websites in Catalist’s outreach corpus. Each domain was crawled, classified for the presence of a wholesale, dealer, or trade portal, and tagged in the suppliers.has_wholesale_page boolean field. | File-extension and MIME-type analysis of every catalog file delivered to Catalist by email attachment, supplier-portal upload, or direct file share between January and May 2026. Classification by file extension first; falls back to declared MIME-type for ambiguous cases. | Inspection of the supplier_column_overrides table, which records the manual mapping a Catalist engineer applied when automated header detection failed for a given supplier’s catalog. Distinct counts of source_column_name and header_signature within the cost role. |
| Source | Catalist proprietary supplier-directory analysis | Catalist supplier_documents ingestion log | Catalist supplier_column_overrides table |
| Sample | 17,863 brand websites, of which 11,796 had a public consumer site | 191 raw catalog files from 128 distinct suppliers | 103 column-mapping records across 47 suppliers (the hard-tail subset whose catalogs required manual intervention) |
| Limitations | Counts only public, crawlable wholesale endpoints. Some suppliers have private wholesale portals gated by login that this analysis cannot detect; the underlying claim — buyer cannot self-serve — is preserved either way. | Excludes catalogs that Catalist accesses via supplier-portal scraping rather than direct file delivery. Including portal-scraped data would not materially change the JSON / API rate, which remains 0%. | Counts the harder cases where automated parsing failed; the full universe of header variance across all 191 ingested catalogs is larger. Numbers shown should be read as a floor, not a ceiling. |
External research referenced:
- Gartner Strategic Predictions 2026 — $15 trillion AI-agent B2B forecast.
- McKinsey — "The Agentic Commerce Opportunity" (2025).
- Kearney — Agentic commerce framework and buyer segmentation.
- Distribution Strategy Group — Industry benchmark for B2B supplier onboarding latency.
Frequently Asked Questions
What is catalog fragmentation in B2B distribution?
Why can’t AI procurement agents read B2B supplier catalogs today?
How many distinct B2B catalog formats exist in practice?
What is a "supply-side connector" for agentic commerce?
How big is the supplier-side data problem?
Further Reading
The State of AI in B2B Distribution
Industry-wide view of agentic commerce and how distributors should prepare.
What is an AI-Native B2B Marketplace?
How AI-native architecture differs from AI-enabled bolt-ons.
The Future of B2B Sourcing
How AI is transforming procurement strategy.
Certificates of Analysis
Enterprise documentation for compliance and provenance.