An AI agent website audit tests whether an automated visitor can complete the tasks a human can: find your pricing, confirm a product is in stock, read your returns policy, and get far enough into checkout to hand off. Most sites fail at least two of those. The usual culprits are pricing hidden behind a "contact us" form, stock status rendered only by JavaScript, policies trapped in PDFs or click-to-open accordions, and checkout flows gated by CAPTCHA or account creation. Agent traffic is currently small, this is preparation, not an emergency, but the audit is cheap, takes an afternoon, and most of the fixes improve your human experience and your search visibility at the same time.
Key Takeaways
- Fetch your own pages the way an agent does: no JavaScript execution, no cookies, no session. What survives is what the agent sees.
- The four most common failures: gated pricing, JS-only stock status, unextractable policies, and checkout walls.
- Structured data is not optional plumbing any more, for an agent it is often the only reliable source of price and availability.
- CAPTCHA and forced account creation stop an agent completely. That is a deliberate business decision, not a bug, but make it deliberately.
- Agent traffic is small as of September 2026. Fix in priority order; do not rebuild your stack.
- Almost every fix here also helps humans, search crawlers and AI answer engines. That is why it is worth doing now.
With JavaScript disabled, most modern commerce pages lose exactly the information an agent came for.
Why This Audit, Why Now
Two things changed in September 2026. Meta launched Muse (8 September), a consumer agent that browses inside a sandboxed VM and can fill forms and make purchases. OpenAI launched the Agents API (10 September), which makes it trivially cheap for anyone to build an agent that visits websites on a user's behalf.
Neither of those means agents are a meaningful share of your traffic today. They are not. But the cost of the audit is an afternoon, and OpenAI's own retreat from Instant Checkout in spring 2026, partly on product-data-quality grounds, tells you what happens when agents meet messy sites: the agent picks someone else.
How an Agent Actually Sees Your Site
An agent is not a browser with a person attached. Its constraints are specific.
It may not run your JavaScript
Some agents render fully; many fetch raw HTML for speed and cost reasons. If your price is injected client-side, you are relying on the agent being in the expensive mode. Do not.
It reads text, not layout
Visual hierarchy carries no meaning. A price in a big bold box and a price in a footnote are identical to an agent unless the markup distinguishes them. That is what structured data is for.
It cannot see images
Your shipping timeline in a graphic, your size chart as a JPEG, your returns window in an infographic, all invisible. Alt text is the entire channel.
It gives up quickly
Agents operate under token and time budgets. A three-click journey to find a price will often just end with the agent choosing a competitor whose price was on the page.
The Audit Method
Four passes. Do them in this order.
Pass 1: Fetch raw HTML
Use curl or any fetch tool on your key pages: homepage, a category page, three product pages, pricing, shipping, returns. Save the output. Search it for price, currency, stock status and delivery estimate. If those strings are absent, an agent that does not render JS cannot find them.
Pass 2: Browse with JavaScript disabled
Turn JS off in your browser and use the site as a customer. This shows you the human-readable version of what pass 1 found. Note every place the page is functionally empty.
Pass 3: Validate structured data
Run your product and policy pages through schema validation. Check that Product, Offer, price, priceCurrency, availability and shipping details are present and accurate. Reference the vocabulary at schema.org. A schema block that is present but wrong is worse than one that is absent.
Pass 4: Attempt the task as an agent would
Pick three concrete tasks: "find the price of X", "confirm Y ships to Bengaluru within a week", "find out how many days I have to return Z", and try to complete each using only the raw HTML from pass 1. Count how many you can answer. That number is your score.
The Common Failure Points
Pricing behind "contact us"
For B2B and many edtech products this is a deliberate lead-gen choice, and I am not going to tell you to abandon it casually. But understand the trade: an agent comparing three vendors will report the two with published pricing and describe yours as "price not disclosed." In an agent-mediated comparison, that reads as a disqualification rather than an invitation.
The middle path is publishing a starting price, a range, or a worked example, and keeping the form for bespoke quotes.
Stock status rendered only in JavaScript
Extremely common on modern storefronts. The product page HTML contains the product, and availability arrives from a separate client-side call. An agent sees a product with no availability signal, which it will typically treat as unknown, and unknown loses to a competitor marked in stock.
Fix: put availability in the initial HTML via Offer schema, updated server-side.
Policies that are not extractable text
Three variants of the same problem: policy in a PDF, policy inside an accordion that only loads content on click, policy only available via chat widget. All three are opaque. Shipping, returns and delivery-time policies should exist as plain HTML text at stable URLs.
Checkout an agent cannot complete
CAPTCHA stops agents by design. So does mandatory account creation, so does OTP verification, so does a multi-step flow with client-side-only state.
Here is the honest framing: given the "discover in AI, buy on site" consensus emerging after OpenAI's Instant Checkout retreat, an agent completing your checkout may not be the goal at all. The realistic goal is a clean handoff: the agent gets far enough to pass a fully-formed, deep-linked cart or product URL to the human, who finishes. Test that handoff works.
Content behind interaction
Tabs, "read more" toggles, hover states, infinite scroll. If content only exists after an interaction, assume it does not exist.
robots.txt blocking agents you wanted
Worth an actual look. Some sites block agent user-agents through inherited bot rules without anyone deciding to. Whether you allow them is a business call: but make it a call, not an accident.
Fix in priority order. Most of this list helps humans and search crawlers too.
Priority Fixes, In Order
First, server-render price and availability. Highest impact, and it also removes a class of SEO problems. If you only do one thing, do this.
Second, publish policies as plain HTML. Shipping times, returns window and conditions, warranty. Stable URLs, extractable text, linked from product pages.
Third, complete your Product and Offer schema. Price, currency, availability, SKU, variants. Validate it and keep it in sync with reality: stale schema is a trust problem, not just a technical one.
Fourth, publish some pricing signal if you currently publish none. A range or a starting price is enough to stay in an agent's comparison set.
Fifth, put key facts in alt text. Size charts, shipping graphics, spec images.
Sixth, test the handoff path. Can an agent construct a URL that lands a human directly on the right product, right variant, ready to buy?
Seventh, instrument agent traffic. Log and segment agent user-agents now so you have a baseline before the numbers matter.
What Not to Do Yet
Do not remove CAPTCHA from checkout to accommodate agents. Do not build a bespoke agent API. Do not rebuild your site as a headless architecture on the strength of two September 2026 launches.
Agent traffic is small. One major agentic checkout effort has already been wound down. The correct level of investment is hygiene, not architecture.
The Boring Upside
Nearly everything in this audit is also standard technical SEO and accessibility work. Server-rendered prices help crawlers. Extractable policies help AI answer engines quote you correctly. Complete structured data helps rich results. Good alt text helps screen readers.
That is what makes the audit worth doing now: the downside case is that agent traffic never materialises and you are left with a faster, more accessible, better-structured site.
FAQ
What is an AI agent website audit?
A check of whether an automated AI visitor can complete the same core tasks a human can on your site: find pricing, confirm availability, read policies, and reach or hand off checkout, usually tested by fetching pages without JavaScript and inspecting structured data.
How do I test my site as an agent would see it?
Fetch key pages with curl to get raw HTML, browse with JavaScript disabled, validate your structured data, then try to answer three concrete customer questions using only the raw HTML.
Does hiding pricing behind a form hurt me with agents?
Yes, in comparison contexts. An agent will report competitors' prices and mark yours as undisclosed, which functions as a disqualification. Publishing a starting price or range while keeping the form for custom quotes is usually the better trade.
Is JavaScript rendering a real problem for agents?
Some agents render fully, many do not, fetching raw HTML is faster and cheaper. Relying on client-side rendering for price and stock means relying on the agent being in its expensive mode.
Should I remove CAPTCHA so agents can check out?
Not as of September 2026. CAPTCHA exists for fraud and abuse reasons that have not gone away, agent traffic is small, and the emerging consensus is "discover in AI, buy on site" rather than agent-completed checkout. Optimise the handoff instead.
How much agent traffic should I expect?
Currently a small fraction of total traffic for most sites. That is precisely why this is preparation work and why you should start logging it now, to have a baseline.
What structured data matters most?
Product and Offer, with price, priceCurrency, availability, sku, and shipping and returns details. See schema.org for the vocabulary.
Can an agent read my PDF policies?
Sometimes, unreliably, and at extra cost. Assume no. Publish policies as HTML at stable URLs and keep the PDF as a secondary copy if you need one.
Will fixing this improve my SEO too?
Almost certainly. Server-rendered content, extractable policies, valid structured data and useful alt text are standard technical SEO and accessibility improvements.
How long does this audit take?
An afternoon for a small site; a few days for a large catalogue if you sample pages properly rather than checking everything.
If you would like someone to run this audit with you and sort the real problems from the theoretical ones, take a look at my work and get in touch via the contact form at younusfardeen.com. I have 4+ years of marketing experience helping brands grow organically, and I will tell you plainly which of these fixes is worth your next sprint.