Everyone measures whether AI can read the web. Cloudflare’s 200,000-domain study measured crawler traffic; the AI-visibility tools measure citations. We wanted the merchant-side number nobody publishes: when a shopping agent arrives at a real store, can it actually find, understand and buy the product?

So we ran our deterministic audit — score /100, public pages only, no LLM in the scoring core — across 30 direct-to-consumer supplement stores, from the biggest names in the vertical to mid-size DTC brands. 27 were scannable. Every report referenced here is live and public.

The headline numbers

  • n = 27 stores · scanned July 2026 · one vertical (DTC supplements), on purpose — cross-vertical averages hide more than they show
  • Median: 56/100 — the typical supplement store fails roughly half of what an agent needs to complete a purchase
  • 3 of 30 stores couldn’t be reached at all: two dead DNS records, and one whose TLS certificate doesn’t cover its own www subdomain — a store that agents (and some humans) simply cannot enter
  • The distribution is bimodal, and the two humps tell the whole story

The split: standard stacks win, custom stacks lose

The top hump — 11 stores between 79 and 96. Ritual (92), Transparent Labs (90), Feel (96), Hunter & Gather (95), Four Sigmatic (93), Batch (93), Olly (88), Gainful (88)… What they share is boring: a Shopify-standard stack with mature structured-data apps. Product JSON-LD, machine-readable price, stock, variants — the basics, present because the platform ships them.

The bottom hump — 16 stores at 56 or below. And here is the finding that made us publish: the biggest brands in the vertical are the least agent-ready. AG1 scored 16. Seed, 20. Hims, 20. Bulletproof, 20. NOW Foods, 4 — its robots policy appears to shut out AI crawlers entirely.

The recurring top-3 gap is identical across the bottom hump: missing Product JSON-LD, no structured price, no machine-readable variants. Not exotic protocol work — the basics.

The inversion, explained

Why do the giants lose to mid-size DTC brands? Because big brand = custom headless storefront = every machine-facing basic must be rebuilt by hand — and nobody rebuilt it. The mid-size brands never faced that choice: their platform serves the structured data by default.

It’s the pattern we keep seeing beyond supplements: readable, beautiful, high-converting human storefronts that are blank pages to a buying agent. A shopper’s assistant comparing “best greens powder, in stock, under €80” can parse Ritual’s catalog completely — and gets nothing machine-readable from the category leader that spends the most on advertising in the entire vertical.

Two caveats we insist on: scores reflect the public storefront surfaces on the scan date — stores ship changes, and a store can re-scan free at any time; and our rubric measures transaction-readiness (can an agent buy), not brand visibility in AI answers — a different layer with different tools. Merchants who believe a check misfired can dispute it: that channel exists because audits should be accountable (contact).

What this means if you run a store

  1. If you’re on a standard platform with a good structured-data setup, you are probably ahead of brands 100× your size. Verify it, then say it.
  2. If you’re headless/custom: the gap is three boring things — Product JSON-LD on product pages, structured price + availability, machine- readable variants. Weeks of work, not quarters.
  3. The door matters before the paperwork: three stores in this batch were unreachable at the network layer. Certificates, DNS, robots policy — an agent gives up in one second flat.

The free scan takes ~30 seconds per store: agentready.market. The public ranking of stores agents can actually buy from is the Agent-Ready Index.

Methodology

Deterministic, reproducible scoring — no LLM grades anything. Public pages only (homepage, product page, feeds, robots, well-known files), scanned with an honest user-agent, one vertical at a time, benchmark percentiles computed against the vertical’s own distribution. Full check list and weights: /methodology on any report. The 27 individual reports are public and linked from the Index; the batch was run July 15, 2026.

Sources

A store an agent can’t parse doesn’t lose a ranking — it loses the sale, silently, to the 92 next door.