Executive summary
Agents can transact with the store, but AI search cannot yet describe or recommend it
The sandbox is unusually advanced on agent infrastructure, with live MCP and UCP endpoints. The foundations that AI search platforms rely on to cite and recommend products (metadata, structured data, detailed content and trust signals) are largely missing.
| Pillar | Weight | Score | Rating | Headline finding |
|---|---|---|---|---|
| 1. Accessibility and discoverability | 25% | 58 | Partial | All AI bots receive a 200 response and server-rendered content, but there is no robots.txt or XML sitemap. |
| 2. Understandability and structure | 20% | 34 | Weak | Every URL shares one title and description; no structured data is present. |
| 3. Content and information quality | 20% | 31 | Weak | Product descriptions average 12 words and do not answer fit, care, sizing or suitability questions. |
| 4. Authority and trust | 15% | 22 | Weak | No reviews, policies, contact details or third-party references to corroborate the store. |
| 5. Data and agent interoperability | 20% | 82 | Strong | MCP server (7 tools), UCP discovery and a filterable product API are live and publicly discoverable. |
| Overall | 100% | 47 | Weak | Interoperability is ahead of the fundamentals. |
What this means
An agent that already knows about the store can search, add to cart and check out without visiting the website. An AI assistant answering an open question such as "a warm hoodie gift for an SEO colleague" has very little to go on: no per-page metadata, no Product schema, thin descriptions and no reviews. Fixing pillars 2 to 4 is the priority, and most of the work is low effort.
Pillar 1 · Accessibility and discoverability · 58 / 100
AI crawlers can reach every page, but nothing tells them which pages exist
We requested key URLs with the user agents of six crawlers and reviewed rendering, linking and the main agent journey (product, add to cart, checkout).
Bot access test
| User agent | Platform | Status | Challenge served | Result |
|---|---|---|---|---|
GPTBot | OpenAI (training) | 200 | No | Pass |
OAI-SearchBot | ChatGPT search | 200 | No | Pass |
ClaudeBot | Anthropic | 200 | No | Pass |
PerplexityBot | Perplexity | 200 | No | Pass |
Google-Extended | Gemini / Vertex AI | 200 | No | Pass |
| Desktop browser | Baseline | 200 | No | Pass |
Findings
| Check | Status | Finding and evidence |
|---|---|---|
| Robots.txt | Fail | /robots.txt returns 404. Nothing is blocked, but there is no explicit AI crawler policy and no sitemap reference. |
| WAF / CDN / bot protection | Pass | Served by Vercel with no bot challenge for any tested AI user agent. CORS is open (access-control-allow-origin: *). |
| Authentication and consent | Pass | No login, cookie wall or consent banner in front of product content. |
| JavaScript dependence | Pass | Product listing and product detail pages are server-rendered; name, price and stock appear in the raw HTML. Cart and checkout are client-rendered, which is acceptable for transactional pages. |
| XML sitemap | Fail | /sitemap.xml returns 404. The 10 product URLs are only discoverable by following links. |
| Internal linking | Pass | Homepage → /products → all 10 product pages via standard <a href> links. Maximum click depth is 2. |
| Orphan pages | Pass | No commercially important orphan pages found. There are no category pages to orphan. |
| Redirects | Pass | No redirect chains on key URLs. |
| Key actions (identify, select variant, add to cart) | Partial | The add to cart button is a labelled <button>, but confirmation uses a blocking alert() dialog, which stalls many browser agents. There are no selectable variants: the hoodie page lists a single size (M) with no size selector. |
| Forms | Fail | In the rendered checkout form, <label> elements are not associated with their inputs (no for / id pairing), and inputs have no name or autocomplete attributes. |
| Buttons and links | Partial | Quantity controls are bare "-" and "+" buttons without aria-label. "Proceed to checkout" navigates through a JavaScript click handler rather than a link. |
| Product / category filtering | Fail | The storefront offers no category or price filters. The UCP manifest declares category, price and availability filters, but there are no equivalent crawlable URLs on the website. |
| Checkout / conversion path | Partial | Guest checkout with no account requirement. The cart is tracked in browser localStorage rather than a cookie or URL, so an agent that does not persist storage between steps loses its cart. |
Pillar 2 · Understandability and structure · 34 / 100
Every page tells AI systems the same thing, and none of it is structured
All 15 crawled URLs return the same title and meta description, and no JSON-LD is present in the HTML of any page.
| Check | Status | Finding and evidence |
|---|---|---|
| Semantic HTML | Partial | Correct <header>, <main>, <nav> and <footer> landmarks, one <h1> per page and product attributes in a <dl>. Product cards on /products skip from <h1> to <h3>. |
| Structured data | Fail | No application/ld+json on any page. Product pages should carry Product with Offer (price, currency, availability), plus BreadcrumbList; the homepage should carry Organization. |
| Page titles and descriptions | Fail | The product page for the Algorithm Update Hoodie has the title "Melt Digital AI Commerce Sandbox". No product is named in any title or meta description. The site runs on Next.js (identified from response headers), so this can be generated per page from existing product data. |
| Entity information | Partial | The homepage explains what the site is, but there is no About page, location, company details or sameAs links tying the store to Melt Digital's other profiles. |
| Consistent information | Fail | The merchant is named "Melt Digital AI Commerce Sandbox" on the website, "AI Commerce Sandbox" in /.well-known/ucp and "Melt Digital AI Commerce" in /.well-known/mcp.json. |
| URL structure | Partial | Product URLs use internal product codes (/products/hoodie-001) rather than descriptive slugs (/products/algorithm-update-hoodie). |
Why this matters. When an AI platform runs a query fan-out, it often sees only the title, description and a short extract before deciding whether to fetch a page. Fifteen identical titles give it no reason to choose one product over another.
Pillar 3 · Content and information quality · 31 / 100
Product copy is too short to answer the questions shoppers ask AI assistants
We tested each product page against five realistic conversational prompts and recorded whether the page alone contained enough information to answer.
| Example AI prompt | Relevant product | Information needed | Answerable |
|---|---|---|---|
| "A warm hoodie for a gift, my colleague is usually an XL, does it run small?" | Algorithm Update Hoodie | Size range, fit guidance, fabric weight | No |
| "Is the mug dishwasher safe, and how big is it?" | Keyword Research Mug | Capacity, care instructions | No |
| "Sunglasses with UV protection under £40" | Rank Tracker Sunglasses | UV rating, price | Yes |
| "A water bottle that keeps drinks cold all day" | Hydrate & Optimize Water Bottle | Insulation, hours cold, capacity | Partly |
| "Can I get it delivered by Friday, and can I return it?" | All products | Delivery times, returns policy | No |
Findings
- Answerability: 1 of 5 test prompts answered in full. There is no delivery, returns or FAQ content anywhere on the site.
- Information clarity: Attributes are well presented in a definition list, but each product carries only 3 or 4 attributes (material, colour, style) with no dimensions, capacity or care details.
- Targeting personas: Copy is written around jokes for SEO professionals ("Stay shaded while you rank") but does not state who the product suits, what occasions it is for or why it makes a good gift.
- Content accuracy: Stock levels on product pages matched the UCP and MCP responses for all 10 products checked. Product images are referenced at
/images/*.jpgbut these return 404 on the live site, so every product page shows a broken image.
Pillar 5 · Data and agent interoperability · 82 / 100
Agent infrastructure is ahead of most production retailers
The store exposes its catalogue and checkout through three machine interfaces, each discoverable from a well-known URL. The remaining gaps are product feed coverage and richer product data.
| Interface | Discovery URL | Capabilities | Status |
|---|---|---|---|
MCP server melt-commerce |
/.well-known/mcp.json |
search_products, get_product, check_inventory, create_cart, view_cart, update_cart, checkout |
Live |
| UCP adapter | /.well-known/ucp |
Product search, details, cart management, checkout, order tracking | Live |
| Storefront JSON endpoints | Not advertised (seen in network requests) | Cart and order operations used by the website | Undocumented |
| Product feed | None | No Google Merchant Center, OpenAI product feed or llms.txt |
Missing |
Findings
- Product feeds: No feed is published. ChatGPT shopping and Google AI Mode both draw on merchant feeds, so the catalogue is invisible to those surfaces regardless of the MCP and UCP work.
- APIs and integrations: Strong. Seven MCP tools cover the full journey, and the UCP manifest declares supported filters and cart operations.
- Data consistency: Price and stock matched across the website, UCP and MCP for every product checked. Merchant naming does not (see pillar 2).
- Transactional accessibility: Agents can complete a checkout through MCP or UCP. Payment is limited to a
demomethod, andsupportsAuthisfalse. - Machine-readable information: Product data returned by the UCP and MCP endpoints lacks
brand,gtin/mpn, size and colour variants, shipping and return fields, which most feed specifications require.
Recommendations
Eight fixes would lift the score from 47 to an estimated 74
Actions are ordered by impact relative to effort. Effort estimates assume the site's current Next.js platform and one developer.
| # | Action | Pillar | Effort | Impact | Score lift |
|---|---|---|---|---|---|
| 1 | Add unique titles and meta descriptions to every page, plus canonical tags | 2 | 0.5 day | High | +6 |
| 2 | Render Product, Offer, BreadcrumbList and Organization JSON-LD from existing product data |
2 | 1 day | High | +5 |
| 3 | Publish /robots.txt (explicitly allowing AI crawlers) and an XML sitemap at /sitemap.xml |
1 | 0.5 day | High | +4 |
| 4 | Rewrite product descriptions to 120+ words covering fit, size, care, dimensions and who each item suits; add delivery and returns information | 3 | 2 days | High | +5 |
| 5 | Use one merchant name across the website, /.well-known/ucp and /.well-known/mcp.json; add an About page with sameAs links |
2, 4 | 0.5 day | Medium | +3 |
| 6 | Fix checkout form labels (for/id, name, autocomplete), add aria-label to quantity buttons and replace alert() with an inline confirmation |
1 | 0.5 day | Medium | +2 |
| 7 | Publish a product feed (Google Merchant Center and OpenAI format), adding brand, GTIN, variants and shipping details to product data | 5 | 2 days | Medium | +3 |
| 8 | Collect product reviews and mark up AggregateRating; publish contact, terms and returns pages |
4 | Ongoing | Medium | +4 |
| Estimated score after all actions | ~7 days | 74 |
Recommended first sprint
Actions 1, 2, 3 and 5 take about two and a half days combined and address the fundamentals every AI platform depends on. Together they are estimated to raise the score from 47 to 65.
Methodology
How this audit was carried out
Each pillar contains six to twelve checks. A pass scores 100, a partial 50 and a fail 0, adjusted by the commercial importance of each check. Pillar scores are combined using the weights in the summary table.
- Accessibility and discoverability: 25%
- Understandability and structure: 20%
- Content and information quality: 20%
- Authority and trust: 15%
- Data and agent interoperability: 20%
- Live HTTP requests to the homepage, product listing, product pages,
robots.txt,sitemap.xml,llms.txtand well-known endpoints, using six user agents. - Raw HTML review (without JavaScript execution) to confirm what non-rendering crawlers receive.
- Rendered DOM and network request review in a desktop browser, covering product, cart and checkout templates.
- Testing of the publicly advertised MCP (
/.well-known/mcp.json) and UCP (/.well-known/ucp) endpoints, including listing available tools and running product search queries. - A manual add to cart and checkout journey, reviewed from the perspective of a browser-based agent.
This audit does not measure current citation share in AI platforms, traditional search rankings or Core Web Vitals. Off-site authority signals in pillar 4 are indicative and based on a sample of searches rather than a full backlink and mention analysis.
Each estimate re-scores the affected checks as passing once the action is complete, then recalculates the weighted total. They show relative priority and are not a forecast of traffic or revenue.
About this report. Prepared as an illustrative example of a Melt Digital AI readiness audit. The audited site is a demonstration store; some figures, particularly off-site signals and score lift estimates, are indicative.