Agentic commerce is AI assistants completing purchases on a buyer’s behalf: ChatGPT checks out in-chat via ACP (the OpenAI/Stripe Agentic Commerce Protocol), Google AI Mode and Gemini read product catalogs via UCP, and agent fetchers like OAI-SearchBot and PerplexityBot parse open product feeds and Product JSON-LD directly. The question for every store in 2026 is no longer only "does Google rank me" — it is "can a buying agent find my catalog, trust my data, and complete a purchase".
On 2026-07-23 we scored 857 leading Shopify DTC storefronts (drawn from public best-of-Shopify and DTC lists, 955 candidates scanned, platform verified during the scan) with the Agentic Commerce Score — an open, deterministic 0-100 rubric computed only from public signals: robots.txt AI-crawler access, llms.txt, sitemap, no-JavaScript rendering, the open product feed, Product JSON-LD completeness, checkout-rail prerequisites, and policy discoverability. The scanner is open source (MIT); every number below is reproducible with one command.
The headline numbers
- 544 of 857 stores (63%) clear the agent-buyable bar: AI crawlers allowed + a machine-readable catalog + ≥80% required-field coverage + a native/partial agentic-checkout rail + readable price and availability + verified shipping and returns policies with a stated return window.
- 313 stores (37%) fail that bar — more than one in three, and not the long tail. The list includes Princess Polly (40/100), Lovevery (42), Brooklinen (46), Gymshark (56), SKIMS (67), Tatcha (71), Casper (76), Kylie Cosmetics (76), Allbirds (78), Ruggable (78) and Mejuri (81).
- 105 stores score a grade A (85+ on the 0-100 scale) yet still miss the agent-buyable bar — a high score with one hard blocker still means a lost agent sale.
- Mean score 84, median 87. On this elite, Shopify-platform sample the floor is high because the platform ships most prerequisites by default; the gaps that remain are merchant-owned.
Household names an AI agent cannot buy from today
Scores below are point-in-time (scanned 2026-07-23, rubric v0.1) and public-signal only; each store has a full check-by-check breakdown page, and scores change as soon as the store fixes the gap.
| Store | ACS score | Grade | Primary gap at scan time |
|---|---|---|---|
| Princess Polly | 40/100 | D | No sitemap; no machine-readable catalog |
| Lovevery | 42/100 | D | No machine-readable catalog; required fields missing |
| Brooklinen | 46/100 | D | No machine-readable catalog; no Product JSON-LD |
| Gymshark | 56/100 | C | Homepage refuses declared bot user-agents; no Product JSON-LD |
| SKIMS | 67/100 | C | No product identifier; no rating schema |
| Tatcha | 71/100 | B | No product identifier; image alt text missing |
| Chomps | 75/100 | B | No product identifier; thin descriptions |
| Vuori | 75/100 | B | No product identifier; no rating schema |
| Casper | 76/100 | B | No product identifier; no rating schema |
| Kylie Cosmetics | 76/100 | B | No product identifier; no rating schema |
| Allbirds | 78/100 | B | No product identifier; no rating schema |
| Ruggable | 78/100 | B | No product identifier; no rating schema |
| OLIPOP | 81/100 | B | No product identifier; no rating schema |
| Mejuri | 81/100 | B | Image alt text missing; thin descriptions |
Two distinct failure shapes show up here. A handful of stores (Princess Polly, Lovevery, Brooklinen) serve no parseable catalog at all — bot protection or a deliberate lockdown closed the surfaces agent fetchers read, so what used to be a scraping defense is now a checkout blocker. Far more common is the second shape: a technically fine store whose product data is missing the fields an agent needs to compare it against a rival — no identifier to match on, no ratings to rank by.
Why the 313 fail
- Policies an agent cannot verify — 270 of the 313. Either the shipping/returns pages could not be fetched as substantive pages at stable URLs, or the returns page never states an explicit window ("within 30 days"). Agentic checkout surfaces merchant policies before completing a purchase.
- Price or availability not machine-readable — 57 of the 313. An agent will not guess stock.
- No machine-readable catalog at all — 51 of the 313: no open feed and no complete Product JSON-LD on the product pages.
- robots.txt root-blocks AI shopping crawlers (GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot…) — 15 of the 313. These stores are invisible to agent-led commerce by their own configuration.
The three gaps almost everyone has
- No machine-readable review ratings — 73% of stores fail (n=806). Reviews exist on the page for humans, but no AggregateRating in Product JSON-LD, so an agent ranking three products by rating simply cannot see yours.
- No product identifier — 43% fail. Without brand plus a sku / gtin / mpn, an agent cannot match, dedupe or price-compare the product; the listing is an orphan.
- No image alt text — 32% fail outright and another 31% are partial. Alt text is the only thing a text-mode agent can read about product photography.
These three are the difference between being technically buyable and actually being chosen. They are also why the same catalogue that looks flawless to a shopper reads as thin to a machine.
The quiet 2026 shift: llms.txt went from novelty to default
91% of scanned storefronts now serve an /llms.txt agent-instructions document — not because 800 merchants each adopted an emerging convention, but because Shopify now auto-generates one for storefronts. Shopify likewise serves a sitemap and /products.json by default, and 100% of these storefronts sit on a native agentic-checkout rail. The platform has taken the floor; every remaining gap — ratings, identifiers, alt text, policy terms — is merchant-owned product data.
Check your own store in 10 seconds
npx agentic-commerce-score your-store.comFree, open source (MIT), zero dependencies, ~8 GET requests, nothing leaves your machine. The full rubric — three pillars (Discover 35 / Evaluate 40 / Transact 25), twelve checks, every threshold — is versioned in SCORE.md, and the complete 857-store dataset ships in the repo. Per-store breakdown pages for every scanned store are at arenza.ai/agentic-commerce-score/report.
- Add AggregateRating (ratingValue + reviewCount) to Product JSON-LD — the single most-failed check, at 73%.
- Publish brand plus a product identifier (sku / gtin / mpn) in Product JSON-LD so agents can match and price-compare you.
- Ship a complete Offer: price, priceCurrency and availability — and expose stock per variant, not just per product.
- State the return window in words on the returns page ("returns accepted within 30 days"), and keep both policy pages at stable URLs (on Shopify: /policies/shipping-policy and /policies/refund-policy).
- Write descriptive alt text on product images, and keep a machine-readable catalog open (/products.json) with title, image, price and description on every product.
- Unblock AI shopping crawlers in robots.txt (GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended) and serve real HTML without JavaScript.
Agent-buyable is the floor, not the win
The Agentic Commerce Score measures prerequisites: whether agents CAN find, evaluate and buy from you. Its 15 checks deliberately do not measure the outcome — whether ChatGPT, Gemini or Perplexity actually RECOMMEND you when a buyer asks. That is a separate share-of-voice, citation and answer-accuracy measurement, per brand and per market, which is what Arenza does. Fix the floor with the checklist above; then measure the outcome — that number is usually the uncomfortable one.
Methodology
Candidates: 955 unique domains compiled from public best-of-Shopify / DTC roundups (Fastbundle, Omnisend, Wisepops, PipiAds, Charle, Eastside Co, TrendTrack and others) plus a canonical set of widely-known DTC storefronts; 857 verified as Shopify storefronts during the scan and included in headline stats. Scanner: agentic-commerce-score v0.2, honest bot user-agent, no JavaScript execution, ~12 fixed GET requests per store, three product pages sampled per store (from the feed, or from the sitemap when the feed is closed). Scores are deterministic and reproducible; the raw NDJSON is committed in the repository. Rubric changes only ship with a version bump, and comparing scores across rubric versions is invalid — an earlier v0.1 pass over these same domains reported higher scores because its 0-100 was dominated by a single feed-open/closed signal, which is exactly what v0.2 fixes.