What does an agentic commerce audit cover?

It covers five areas: crawl access (can AI crawlers reach the site at all), product schema (does each product page carry machine readable price, availability and identifiers), discovery files (llms.txt, agents.md, a UCP manifest, a sitemap that exposes product URLs), merchant data (return policy, shipping details, FAQ and Organization as structured data), and content quality (title length, description length, image count). Crawl access and product schema carry the most weight because everything else depends on them.

How often should I re-run an agentic commerce audit?

Re-run it after any deploy that touches page templates, theme files, plugins or robots.txt. These fixes live in exactly the files that deploys overwrite, so a theme update or a plugin default can silently undo a passing check without any visible change to the storefront. A quarterly re-check is a reasonable floor if you deploy rarely.

Is an agentic commerce audit the same as an SEO audit?

No. An SEO audit optimises for a ranking algorithm serving a human who will click through and read the page. An agentic commerce audit tests whether a program can extract facts without rendering the page: exact price, currency, stock state, a product identifier it can match against other listings, and machine readable return and shipping terms. A store can rank well and still fail an agent readiness scan.

Blog

Agentic commerce audit: what it covers and why stores fail

2026-07-31

Audits

Being told to "get audited" for AI agents is not informative on its own. An agentic commerce audit is a specific thing: a test of whether an automated shopping agent can reach your store, read your product pages without rendering them, and pull out enough hard fact to compare you against a competitor. It is not an SEO audit with new vocabulary. The reader is a program, and programs fail differently from people.

The reason this became a practical concern is that the buying surface moved. Shopify turned agentic storefronts on for eligible merchants in March 2026, and since May 2026 it ships llms.txt, agents.md and UCP (Universal Commerce Protocol) support natively. AI-referred traffic and AI-attributed orders grew sharply through 2025 and 2026. The wider argument for why this shift happened when it did is in the switch was flipped; the short version is that agents now transact, and a store they cannot parse is a store they skip.

What the audit inspects

Sextant runs forty checks across five weighted categories.

Crawl access, 25 percent. Whether robots.txt blocks each AI crawler by name, whether there is a blanket disallow, whether the site is served over HTTPS, whether it is reachable at all, and whether the homepage responds in under three seconds. Five crawlers are checked individually: GPTBot (OpenAI, training and ChatGPT browsing), ClaudeBot (Anthropic, Claude retrieval), OAI-SearchBot (OpenAI, ChatGPT search results), PerplexityBot (Perplexity, answers and shopping) and Google-Extended (Google).

Product schema, 25 percent. Product JSON-LD on the page, an Offer carrying price and currency, availability, a SKU, a GTIN or MPN, AggregateRating and BreadcrumbList.

Discovery, 20 percent. llms.txt, llms-full.txt, agents.md, a UCP manifest at /.well-known/ucp, a valid sitemap, and whether that sitemap actually exposes product URLs rather than just pages.

Merchant data, 15 percent. MerchantReturnPolicy, OfferShippingDetails, FAQPage and Organization. These are the fields agents lean on when deciding between two otherwise equivalent listings, and there is a longer breakdown of them in the schema fields AI agents read.

Content quality, 15 percent. Product titles of 30 characters or more, descriptions of 500 characters or more, and three or more detectable images.

Why two categories carry half the weight

Crawl access and product schema are weighted heaviest because both are gating rather than contributory.

If a crawler cannot get in, nothing else is measurable. A perfect llms.txt behind a Disallow: / for GPTBot is not a partial pass, it is invisible. The other categories describe how well an agent can use your data; crawl access decides whether the question arises.

Product schema gates in the other direction. An agent comparing three stores is matching identifiers and reading prices. If a product page has no structured data, there is nothing to compare, and the page drops out of the shortlist before quality of description matters. This is the mechanism behind a Shopify store being invisible to AI despite ranking normally in search.

Why stores fail

The failures are rarely negligence. They are stale decisions.

Crawler blocks added during a scraping scare. Somebody added AI user agents to robots.txt when scraping was in the news, the reason was sound at the time, and nobody revisited it once those same agents started sending buyers. Sextant reports these one crawler at a time as crawl.gptbot, crawl.claudebot, crawl.oai_searchbot, crawl.perplexitybot and crawl.google_extended.

Plugin defaults. WooCommerce core emits no Product schema without a plugin, so the output depends entirely on which plugin is installed and how it was configured, often years ago by someone else.

Schema that validates but is empty where it counts. This is the hardest one to spot, because every validator passes it. An Offer node with no price. A Product with sku present and blank, and no gtin13 or mpn anywhere. A store with thousands of reviews rendering as stars in the page and no AggregateRating in the markup. Structurally correct, commercially useless.

Policy pages written for humans only. A returns page in clear prose is not MerchantReturnPolicy. The merchant.return_policy and merchant.shipping_details checks look for the structured version, because an agent quoting your returns window needs a field, not a paragraph.

Discovery files absent entirely. No llms.txt, no agents.md, no UCP manifest, and often a sitemap that lists collections but no product URLs.

What a useful output looks like

An audit that returns a number and nothing else is not actionable. A useful one returns, per check: pass or fail, the evidence found, and the fix.

Evidence matters because it distinguishes "no Product schema" from "Product schema present, offers.price missing on four of five sampled pages". Those look identical in a score and take completely different work to fix. Sextant scores 0 to 100 with an A to F grade, but the grade is a summary of the check list, not a substitute for it. If your report cannot tell you which line of which file to change, it has not audited anything.

How to run one

Start with the free scan at Sextant. It needs no signup, installs nothing, and reads only public pages: homepage, robots.txt, sitemap, a sample of product pages and your policy pages. It usually takes 15 to 30 seconds.

Then work top down by weight. Crawl access first, because it gates the rest:

User-agent: GPTBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Google-Extended
Allow: /

Then product schema. Fill the identifier and price fields, not just the node:

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Merino Base Layer, Long Sleeve",
  "sku": "ML-BASE-001",
  "gtin13": "5012345678900",
  "offers": {
    "@type": "Offer",
    "price": "89.00",
    "priceCurrency": "GBP",
    "availability": "https://schema.org/InStock"
  }
}

After that, discovery files, merchant data, then content quality. If the underlying idea is still fuzzy, what agent readiness means sets out the terms.

Re-run the scan after any deploy that touches templates or robots.txt. These fixes live in the files deploys overwrite, and a theme update can undo one without changing anything a customer would notice.

If you want to see which of these checks your store currently fails, run a free scan. It reads public pages only and takes about half a minute.