Is Walmart ready for AI agents?

Company: Walmart (WMT) URL tested: https://www.walmart.com Observed at: 2026-08-27T21:04:39.291Z Methodology: AgentReady `2026-08-23-v2`, scanner `agentready-scanner-2026-08-23-v2` Observation coverage: 48% Report status: Partial evidence; the browser interaction pass was unavailable Correction route: [Request an evidence correction](/contact?subject=index-correction)

The result in one sentence

Walmart received an applicability-aware diagnostic with 48% observation coverage. The defensible conclusion is narrow: the public homepage and protocol probes exposed several useful machine-readable and trust signals, while a large set of rendered-interface checks could not be observed in this run. That distinction is the foundation of this index. The private numeric diagnostic helps maintainers locate changes; the public article emphasizes evidence and coverage because a number without observation context invites false precision.

Walmart says its Gemini experience uses Universal Commerce Protocol and connects AI discovery with Walmart and Sam’s Club assortment, customer relationships, and fulfillment. Walmart frames agentic commerce as moving chores toward repeatable automation, which raises the bar for consent, substitution control, and reliable state. Those official claims establish strategic relevance, not proof that an arbitrary external agent can complete the audited task. This report tests the public evidence a third-party agent could discover without credentials and frames a safe, reversible customer journey for follow-up verification.

The exact journey we evaluated

The goal was: Assemble a weekly household order that satisfies budget, quantity, local inventory, and delivery constraints.

The fixed task was: Build a ten-item household replenishment basket under $120, keep sizes and quantities comparable, verify local availability and fulfillment windows, identify substitutions, then prepare the basket for order review.

The safe stopping point was: Stop before linking an account, accepting substitutions, reserving delivery, submitting payment, or placing an order. This boundary matters. AgentReady did not place an order, make a reservation, transmit private data, authorize software, create credentials, or accept terms. The URL scanner inspected public discovery files, response behavior, homepage content, security and trust signals, and optional commerce evidence. The journey contract describes what a controlled browser evaluation should attempt; it is not a claim that the current diagnostic executed every step.

Routine commerce appears simple but is sensitive to store location, membership benefits, substitutions, package size, quantity, recurring preferences, and delivery slots. A household basket tests memory and constraint integrity across many items. Silent package-size changes or unavailable substitutions can invalidate the recommendation. A useful benchmark therefore asks more than “can an agent open the site?” It asks whether the agent can preserve constraints, locate current evidence, explain uncertainty, recognize consequential transitions, and stop before authority is required.

What the AgentReady score measured

This report used methodology version `2026-08-23-v2`. The model produces 34 stable checks across public web discovery, API and MCP discovery when applicable, semantic and accessible structure, browser compatibility, security and safety, legal and trust evidence, repository signals when authorized, and optional commerce interoperability. Essential checks account for 80 base points, recommended checks account for 20, and positive bonus signals can add up to five without lifting the headline above 100. The relevant test shape here is specific: Routine commerce appears simple but is sensitive to store location, membership benefits, substitutions, package size, quantity, recurring preferences, and delivery slots.

The states are deliberately non-binary. Pass means the scanner found positive evidence for that check. Partial means some evidence was present but the condition was incomplete or ambiguous. Fail is reserved for observed negative evidence. Not applicable removes an optional surface from the denominator when it was not observed or advertised. Unobservable means the provider could not collect the necessary evidence in this run; it lowers coverage rather than the target’s score. The raw artifact retains a legacy migration value for audit continuity, but that value is not rendered or used for public comparison. For Walmart, that restraint is consequential: A household basket tests memory and constraint integrity across many items. Silent package-size changes or unavailable substitutions can invalidate the recommendation.

The scanner recorded 7 passes, 8 partials, 0 failures, 3 not-applicable checks, and 16 unobservable checks. The completed pass set included Crawler and agent access policy, Accessible landmarks, HTTPS transport, Agent bot access, No agent-blocking CAPTCHA, Structured data quality, Legal and trust pages. The partial set included Sitemap discovery, LLM-readable product description, Agent instructions, Agent card or manifest, Discoverable and well-typed API contract, Browser security headers, No exposed secrets, Commerce interoperability. The explicit failure set was: No checks were classified as fail; that is not equivalent to complete readiness because browser evidence was unavailable.

Observable evidence

• Crawler and agent access policy: pass. robots.txt found: HTTP 200; AI agents are not explicitly blocked in robots.txt; robots.txt references sitemap: https://www.walmart.com/sitemap_category.xml
• Sitemap discovery: partial. sitemap.xml missing or invalid: https://www.walmart.com/sitemap.xml
• LLM-readable product description: partial. llms.txt missing: https://www.walmart.com/llms.txt
• Agent instructions: partial. Missing agent instructions (agents.md / AGENTS.md): Add one to site or repo
• Agent card or manifest: partial. No agent manifest found: Checked /.well-known/agent-card.json and ai-agent.json
• Discoverable and well-typed API contract: partial. No OpenAPI/Swagger spec found: API advertised; checked common paths
• MCP discovery card: not_applicable. No MCP capability was observed or advertised.
• Accessible landmarks: pass. Landmarks detected (fallback): 4/4
• HTTPS transport: pass. HTTP redirects to HTTPS
• Browser security headers: partial. Content-Security-Policy present; Frame embedding protection present; X-Content-Type-Options nosniff missing
• Agent bot access: pass. No edge blocking detected for agent user-agents: Probed OAI-SearchBot, ClaudeBot, PerplexityBot, ChatGPT-User
• Structured data quality: pass. JSON-LD presence (fallback): Found
• Rendered browser interaction evidence: unobservable. No rendered browser evidence was returned in the canonical run: Browser-dependent behavior remains unknown and is not classified as a failure

These observations are point-in-time facts about the returned public responses, not permanent properties of Walmart. A missing conventional path such as `/sitemap.xml` does not prove the company has no sitemap; robots.txt may advertise a different valid location. A missing `llms.txt`, agent card, MCP card, or OpenAPI file does not prove there is no private or partner integration. Conversely, the presence of a discovery file does not prove that a multi-step customer journey works. Public discovery and operational completion are different layers. The official context for this particular target is also narrower than a readiness claim: Walmart frames agentic commerce as moving chores toward repeatable automation, which raises the bar for consent, substitution control, and reliable state.

Strengths visible in this run

The strongest part of Walmart’s result is that the scanner could obtain concrete evidence instead of relying only on marketing language. HTTPS transport, crawler policy, legal discoverability, structured data, security headers, sitemap references, and agent-oriented files each answer a different operational question. Their value is cumulative: a crawler needs permission and entry points; a reasoning system needs structured, current content; an acting system needs stable controls and explicit consequence boundaries. For this journey, the operational question is: Assemble a weekly household order that satisfies budget, quantity, local inventory, and delivery constraints.

The report’s complete passes—Crawler and agent access policy, Accessible landmarks, HTTPS transport, Agent bot access, No agent-blocking CAPTCHA, Structured data quality, Legal and trust pages—provide the clearest starting assets. Teams should preserve these during redesigns and protocol launches. If a discovery surface already works, the next improvement should connect it to canonical product, policy, pricing, or developer truth rather than publish a second, divergent narrative. For Walmart, that connection is especially important because walmart frames agentic commerce as moving chores toward repeatable automation, which raises the bar for consent, substitution control, and reliable state.

Another strength is category fit. Routine commerce appears simple but is sensitive to store location, membership benefits, substitutions, package size, quantity, recurring preferences, and delivery slots. This makes the company more informative than a generic brochure site. The journey includes real constraints and a natural handoff point where an agent can summarize evidence for a person without claiming authority it does not have.

Failures, gaps, and uncertainty

The largest limitation is systematic: the canonical run returned no rendered browser evidence. As a result, rendered DOM, accessible names, form labels, interaction targets, runtime stability, primary CTA discovery, and several content checks were unobservable. The run is therefore marked partial, even where its observable checks are strong. This is missing evidence, not a failed site check, and we did not substitute a manual impression for scanner evidence. In concrete terms, the report cannot verify this Walmart task: Build a ten-item household replenishment basket under $120, keep sizes and quantities comparable, verify local availability and fulfillment windows, identify substitutions, then prepare the basket for order review.

The partial checks—Sitemap discovery, LLM-readable product description, Agent instructions, Agent card or manifest, Discoverable and well-typed API contract, Browser security headers, No exposed secrets, Commerce interoperability—represent practical opportunities, but each needs live verification before implementation. Conventional discovery paths are valuable because external agents can find them cheaply, yet conventions alone cannot establish correctness. API or MCP marketing mentions should lead to authenticated documentation, schemas, scopes, error models, and test environments. Commerce feeds and product schema should share the same price and availability source. Policies should be linked near the action they govern, not merely in a footer.

A household basket tests memory and constraint integrity across many items. Silent package-size changes or unavailable substitutions can invalidate the recommendation. That is why the safe stopping point is part of the published profile. The index rewards explainable progress toward a customer goal, not aggressive clicking. An agent should surface unresolved terms and ask for confirmation rather than treating a technically enabled action as authorized.

The correction route

The most valuable correction is reproducible evidence. Walmart can use the [index correction route](/contact?subject=index-correction) to identify the exact observation, provide a public canonical URL or response, and request a rerun under the same methodology. A correction should include the tested hostname, timestamp, check ID, expected evidence, and whether the response varies by geography, authentication, user agent, or account state. AgentReady should update an observation when the public evidence changes, not negotiate the score as a matter of opinion. The highest-value correction for this profile would reduce uncertainty around: Assemble a weekly household order that satisfies budget, quantity, local inventory, and delivery constraints.

For the browser gap, the correction is first internal: obtain a functioning configured browser provider, rerun the same journey and URL, and publish the new timestamp and coverage. The prior artifact should remain available so readers can see why the evidence changed. If Walmart publishes new protocol files or changes bot rules, the next report should identify the changed response rather than silently overwrite history. Any interactive rerun must retain this company-specific boundary: Stop before linking an account, accepting substitutions, reserving delivery, submitting payment, or placing an order.

A practical improvement sequence

1. Make canonical discovery cheap. Keep robots directives, sitemap locations, agent guidance, and machine-readable descriptions current and mutually consistent.
2. Bind claims to operational truth. Product, pricing, availability, policy, API, and protocol statements should come from authoritative systems with timestamps and stable identifiers.
3. Expose structured constraints. Agents need dimensions, eligibility, dates, totals, cancellation terms, fulfillment options, scopes, and consequence metadata—not just persuasive copy.
4. Design the review boundary. Before any purchase, reservation, signature, integration, or payment, show the exact subject, amount, terms, identity, and resulting effect.
5. Support recovery. Offer idempotency, back navigation, cancellation, correction, and a human escalation path. The happy path is not enough.
6. Verify the same task after changes. Improvements should be judged against the fixed journey above with the same methodology and a new production timestamp.
7. Resolve this target’s dominant uncertainty. For Walmart, publish or expose the authoritative fields needed to assemble a weekly household order that satisfies budget, quantity, local inventory, and delivery constraints. Then test the exact task—Build a ten-item household replenishment basket under $120, keep sizes and quantities comparable, verify local availability and fulfillment windows, identify substitutions, then prepare the basket for order review.—without crossing the safe stopping point.

For Walmart, the most commercially meaningful next test is not another homepage scan. It is a controlled browser run that attempts this exact task: Build a ten-item household replenishment basket under $120, keep sizes and quantities comparable, verify local availability and fulfillment windows, identify substitutions, then prepare the basket for order review. The run should record every decision-relevant field and stop at: Stop before linking an account, accepting substitutions, reserving delivery, submitting payment, or placing an order. The resulting report should distinguish target facts, agent inferences, missing data, changed data, and user choices. That creates a useful product artifact for Walmart and a credible benchmark for readers.

Patterns this profile contributes to the index

This company illustrates three wider trends. First, public companies are announcing agentic partnerships faster than their public web surfaces are converging on common discovery conventions. A UCP, MCP, AI assistant, marketplace integration, or internal agent can be strategically important without making the conventional homepage easy for arbitrary agents to interpret. Second, discovery is separating from action: product and content data may be portable while checkout, booking, signing, or payment remains governed by proprietary identity and policy systems. Third, safety is becoming a product feature. The most trustworthy journey is not the one that reaches commitment fastest; it is the one that preserves constraints and asks for authority at the right moment. Walmart makes that pattern concrete because walmart says its gemini experience uses universal commerce protocol and connects ai discovery with walmart and sam’s club assortment, customer relationships, and fulfillment.

The report should therefore be read as a diagnostic entry point. Walmart’s evidence at 48% coverage is useful for locating observable strengths and gaps under one frozen model. It is not a certification, accessibility determination, security audit, legal opinion, search-placement guarantee, or claim that every agent can use the site. Comparisons are responsible only when methodology version, task, timestamp, target, and coverage are shown together.

Sources and reproducibility

• Walmart and Google Turn AI Discovery Into Effortless Shopping Experiences — Walmart, checked August 27, 2026.
• The Future of Shopping Is Agentic. Meet Sparky. — Walmart, checked August 27, 2026.

The raw scanner output is stored with this article at `raw/walmart.json`. It contains the complete check list, evidence messages, phase results, score metadata, timestamps, and the browser limitation. The official sources above provide company context; they did not replace scanner observations. This separation lets a reader reproduce public HTTP evidence, challenge a finding, and understand which conclusions are measured versus editorial.

Bottom line: Walmart belongs in the Agentic Customer Journey Index because routine commerce appears simple but is sensitive to store location, membership benefits, substitutions, package size, quantity, recurring preferences, and delivery slots. This run found enough public evidence to identify concrete strengths and correction opportunities, but the incomplete browser coverage prevents a final judgment about task completion. The honest next move is a same-task rerun focused on assemble a weekly household order that satisfies budget, quantity, local inventory, and delivery constraints., followed by a reviewed report that stops here: Stop before linking an account, accepting substitutions, reserving delivery, submitting payment, or placing an order.