Capability brief · Independent · Evidence-first

Costco built the most trusted labor model in retail. It hasn't built the receipt for it.

~341,000 employees, ~93% one-year retention, top wages near $31.90/hour, members who believe the brand. Four findings below, hardest evidence first: every one re-checkable in a single command. The durable model and clean crawl hygiene are credited first, because they're real.

Method. Passive HTTP only: curl header/file GETs of standard public paths, tested both with the default client UA and with a normal browser UA, plus a soft-404 control (byte-size compare), and WebSearch/WebFetch of Costco's own sustainability pages + reported figures (labeled reported). No authentication, no scanning, no enumeration beyond standard well-knowns. Captured 2026-07-14. Not affiliated with or endorsed by Costco.

Baseline: credited first

Costco's fundamentals are genuinely strong. The labor model is the envy of retail: ~341,000 employees, ~93% one-year retention (industry turnover routinely tops 60%), top clerks near $31.90/hour under a multi-year raise agreement. On the web: a real, maintained robots.txt (200) and sitemap_index.xml (200); a serious content-security-policy with object-src 'none', base-uri 'self', and upgrade-insecure-requests; Akamai Bot Manager guarding the edge. This brief isn't "Costco is broken": it's "the most trusted brand in the warehouse has no machine-verifiable receipt to back the trust."

The bottom line, in plain English

Let a member's shopping assistant reach Costco → you stay in the cart as buying moves to AI

A growing share of shopping now runs through AI assistants, and today the site refuses them at the door; letting the good ones read the page keeps Costco reachable for the next wave of members instead of invisible to it.

Fix the agent doors that error instead of answering → Costco looks built, not half-broken

The paths an AI checks first return server errors rather than clean answers, which reads as a broken system; a small fix makes the digital front door look as solid as the warehouse floor.

A receipt behind the Kirkland promise → members can verify the trust, not just take it on faith

The palm-oil and deforestation commitments live only in a PDF a member has to believe; a per-outcome receipt lets a member, an agent, or an NGO check the claim, which protects the exact premium Costco's brand is built on.

Finding 1: the lead

Provable · Costco is invisible to non-browser agents

A plain agent request is blocked before it can read a single page.

A standard HTTP GET with the default client user-agent returns 000: the connection is refused/blocked at the edge by Akamai Bot Manager (its fingerprinting cookies _abck, bm_sz, akavpau_* appear only once a browser UA is used). Swap in a normal browser UA and the same URL returns 200. An AI assistant acting on a member's behalf, the way a growing share of shopping now happens, can't even fetch robots.txt (the file that tells search engines what to index).

$ curl -s -o /dev/null -w "%{http_code}\n" https://www.costco.com/robots.txt 000 # default UA → blocked at the edge, no response $ curl -s -A "Mozilla/5.0 ...Chrome/126 Safari/537.36" -o /dev/null -w "%{http_code}\n" \ https://www.costco.com/robots.txt 200 # browser UA → served; Set-Cookie: _abck, bm_sz, akavpau_* (Akamai Bot Manager) # server: istio-envoy
Why it matters: bot protection is legitimate, but it currently makes zero distinction between a scraper and a member's AI assistant. The result is that Costco, uniquely among these targets, isn't just missing an agent layer; it's actively unreadable to agents. Trust you can't reach is trust you can't compound.
Re-check: curl -s -o /dev/null -w '%{http_code}\n' https://www.costco.com/robots.txt (000) vs the same with -A "Mozilla/5.0…" (200)

Finding 2: broken agent well-knowns

Provable · 500s, not 404s

The agent-identity paths aren't just missing: they error out.

With a browser UA, /.well-known/agent-card.json (a file that lets other AI agents discover and call you) and /.well-known/ai-plugin.json return HTTP 500 (a server error, gzip-garbled body) rather than a clean 404. llms.txt (the file AI assistants read first), ai.txt, and security.txt return the 157,826-byte SPA 404 shell, byte-identical to a random control path, i.e. soft-404s. There is no llms.txt, no agent card, no security.txt. The platform underneath is legacy HCL/WebSphere Commerce (the robots file is full of *DisplayView, StoreCatalogDisplay, CategoryDisplay endpoints).

$ UA="Mozilla/5.0 ...Chrome/126"; for p in llms.txt ai.txt .well-known/agent-card.json .well-known/ai-plugin.json; do \ curl -s -A "$UA" -o /dev/null -w "$p %{http_code} %{size_download}\n" https://www.costco.com/$p; done llms.txt 404 157826 # soft-404 SPA shell ai.txt 404 157826 # == control size → soft-404 .well-known/agent-card.json 500 792 # server error, not a 404 .well-known/ai-plugin.json 500 792 $ curl -s -A "$UA" -o /dev/null -w "control %{http_code} %{size_download}\n" \ https://www.costco.com/this-is-a-control-xyz123.txt → 404 157826
Why it matters: a 500 on the exact path an A2A/MCP client checks first is worse than a 404: it signals a route that half-exists and half-fails. Costco keeps its crawl surface tidy (robots + sitemap are real); the agent surface was never built, and where a stub exists it errors.
Re-check: curl -s -A "Mozilla/5.0…" -o /dev/null -w '%{http_code}\n' https://www.costco.com/.well-known/agent-card.json

Finding 3: the commitment gap

Provable (reported) · Kirkland commitments are PDF-bound + self-attested

The sustainability promise members are asked to trust has no verifiable object.

Costco's Kirkland Signature commitments are real and public: RSPO membership with a goal of 100% RSPO-certified (or equivalent) palm oil, support for zero-net-deforestation (HCV/HCS forest protection), and a stated plan to release a deforestation assessment and action plan for beef, palm oil, soy, and wood. But the substance lives in HTML narrative pages plus a literal PDF policy, self-attested: there is no per-supplier or per-product proof an agent, a member, or an NGO can independently verify.

# public commitment surfaces (HTML + PDF): https://www.costco.com/sustainability-kirkland-signature.html https://www.costco.com/sustainability-environment.html https://www.costco.com/wcsstore/CostcoUSBCCatalogAssetStore/feature-pages/16w0604-sustainability-palm-oil.pdf # PDF policy # reported: RSPO 100% target · zero-net-deforestation (HCV/HCS) · deforestation action plan
Why it matters: Costco's entire advantage is that members believe it. A PDF asks for belief; a receipt earns it. Proof-of-outcome, funded, proof-verified, settled on-chain, is the one upgrade that fits a trust brand better than it fits anyone else.
Re-check: open the three URLs above · WebSearch "Costco RSPO palm oil deforestation action plan Kirkland Signature"

Finding 4: hygiene (minor)

Provable · staging hosts + dead robots cruft

Two small housekeeping notes on an otherwise serious build.

(a) The production content-security-policy allowlists lower-environment hosts: mobilecontent-qa.costco.com, gdx-npd.np.api.cc-costco.com, api-tst.np.gdx.cc-costco.com, stg.api.bazaarvoice.com, network-stg-a.bazaarvoice.com, alongside production origins. (b) The top of robots.txt carries a large block of commented-out legacy Disallow rules (dead cruft) above the active ruleset.

$ curl -sI -A "Mozilla/5.0 ..." https://www.costco.com/ | grep -io 'content-security-policy[^\n]*' \ | grep -oE '(mobilecontent-qa|gdx-npd|api-tst|stg\.api\.bazaarvoice|network-stg-a)[^ ]*' mobilecontent-qa.costco.com gdx-npd.np.api.cc-costco.com api-tst.np.gdx.cc-costco.com stg.api.bazaarvoice.com/
Why it matters: low stakes, but QA/NPD API hosts in a production CSP and a wall of commented rules are exactly the kind of thing a legibility pass sweeps up alongside the missing agent layer.
Re-check: curl -sI -A "Mozilla/5.0…" https://www.costco.com/ | grep -i content-security-policy

The ask

One warehouse's community. One verified receipt. Under $25.

Fund one place-based Restore packet, a stormwater or urban-canopy survey in a single warehouse's community, for under $25. It settles to a public, permanent receipt a member or the sustainability page can cite: this warehouse, this community, this outcome, verified.

Or have any AI agent make one keyless call over x402 (a web-native way for AI agents to pay per use): x402 POST vealth.net/.well-known/vealth/work-packet ($0.10), to mint a Kirkland supply-chain proof packet. Deliverable: the on-chain proof receipt.

Why this and not a meeting: it ties straight to Costco's own stated commitments (RSPO, zero-net-deforestation, the deforestation action plan) and hands the trust brand its missing verifiable object. Proof first, ask second.

← Hub · Run the demo · Agent Kit →

Provisioned by EcoWealth · Digital Provisioning Protocol · provenance‑sealed ↗