Mixed signal. Agents can reach the page, but guess from there.

Glossier scores 61 out of 100, a mixed signal that undersells how strange the profile is. The site does the hard infrastructure work well: every AI crawler from GPTBot to ClaudeBot to PerplexityBot is explicitly allowed in robots.txt, an llms.txt file exists at the well-known path, all nine sampled product pages carry Product plus Offer schema in both JSON-LD and microdata, prices render in raw HTML via schema payloads, and Shop Pay, Apple Pay, and PayPal all sit on cart and homepage with a Shopify /cart/add endpoint exposed. Then the site turns around and serves reCAPTCHA and hCaptcha on the cart page. An agent can find the product, parse the price, and get blocked at the moment of purchase.

That CAPTCHA finding is the single most expensive item on the report, a zero out of four with transaction_blocking flagged true. Detecting both reCAPTCHA and hCaptcha on cart (and homepage) while offering machine-native rails like Shop Pay is a contradictory posture: the rails invite agent commerce, the challenge widgets refuse it. Whatever fraud logic gates that widget should be reviewed against the reality that agent traffic is now the traffic Glossier said it wanted when it opened robots.txt to everyone.

Structured product attributes are the next hole, another zero, this time out of six. Across all nine sampled URLs the scanner found zero structured attribute hits and zero labeled hits. Glossier writes strong human copy (the Body Hero mist description name-checks Sea Fennel Extract, Sunflower Seed, Grapeseed, and Oat), but that content lives in prose only. There is no attribute table, no key-value labeling, no schema additionalProperty. An agent trying to answer "is this fragrance-free" has to read paragraphs when it should be reading fields.

Checkout structure and guest checkout together give up four more points. The classifier could not even render /checkout without an initialized cart, and the cart page surfaces a "Log in" button with no visible guest path. The one raw 403 the scanner captured across 32 fetches was, tellingly, https://glossier.com/checkout. Combined with the CAPTCHA on cart, the transaction surface is effectively closed to non-human sessions despite everything upstream being open.

Two smaller but revealing gaps: the sitemap at glossier.com/sitemap.xml returns 757 URLs but contains no product URLs and no lastmod timestamps anywhere, costing two points on discoverability hygiene; and stated agent policy is a flat zero because the scanner found no terms or privacy content to classify. For a brand that clearly thought about llms.txt, the absence of any written agent policy is a miss, not an oversight worth defending.

Price stability and price history also fail (zero of two each). Sale markup was detected on all nine products but no schema priceValidUntil or dated validity window accompanies it, so an agent sees a discount with no expiry and no provenance. Delivery dates are similarly unstructured: five of nine pages mention shipping speed in prose, zero use structured schema.

The single highest-leverage move is removing CAPTCHA from the cart path and adding structured product attributes; Glossier has already done the discoverability work, and the transaction layer is what is currently wasting it.