Skip to content

AI readiness audit sampling: choose representative pages

Build an AI readiness audit sample that covers templates, markets, access rules and important URLs, with clear limits on what the results establish.

By Crawlito · 6 min read

In short

  • Sample by template and delivery conditions, then add important URLs and known exceptions.
  • A clean sample does not establish that every URL on a website passes.
  • Record both the tested URL count and the page families those tests represent.
On this page

An AI readiness audit sample should cover the different ways your website produces and delivers content. Start with page templates, then include important commercial pages, languages, access conditions and known exceptions. A random handful of popular URLs is easy to collect but may miss the one template that sends crawlers an empty response.

Sampling is a way to allocate inspection time. It is not permission to describe unchecked pages as verified. This guide explains how to build a defensible sample and how to state its limits, whether you are reviewing ten pages manually or inspecting a larger crawl export.

Define the question before choosing URLs

Decide whether you are checking an entire publishing system, a recent release or a particular business journey. Those questions need different samples. A pricing-page investigation should follow pricing and checkout-adjacent routes. A template review should inspect every materially different template even when some templates receive little traffic.

Write the scope in one sentence: “Review public English and Arabic product information served from the main domain, excluding customer accounts.” That statement identifies coverage and boundaries. It also prevents a discovery crawl from quietly expanding into private documents or unrelated subdomains.

Keep the technical objective separate from visibility. At the time of writing, Google documents indexing and snippet eligibility as requirements for supporting links in its AI features. A sampled audit can inspect some prerequisites; it cannot establish future selection by a search system. Our explanation of what an AI readiness audit covers provides the broader service context.

Identify the available evidence before promising a scope. You might have a sitemap but no CMS export, or public pages but no firewall logs. Record those limitations at intake, because they affect what you can discover and verify.

Build a page-family matrix

Group URLs by how they are built, not only by directory. Two routes under /products/ might use different templates because one is a configurable product and the other is a simple item. Conversely, several directories can share a single article template.

Use a worksheet like this hypothetical example:

FamilyRepresentative pageVariation to includeReason
Main landing pagesHomepageCampaign landing pageDifferent layout system
ProductsSimple productConfigurable productDifferent content loading
ArticlesRecent guideOlder migrated guideDifferent stored markup
DocumentationOverviewDeep reference pageDifferent navigation and rendering
Regional pagesEnglish service pageArabic equivalentDifferent locale handling

Add columns for template owner, CMS, host, language, authentication expectations and delivery layer. You are looking for combinations that could alter the response. A second CDN or a separate documentation deployment deserves attention even when the design looks identical.

Do not multiply every dimension mechanically. Testing every template in every market on every device can create a large grid with little additional information. Instead, identify which dimensions actually change the implementation, and record the reasoning behind each exclusion.

Add risk-based and business-critical pages

A representative template sample is only the beginning. Include pages whose failure would matter disproportionately: your main service page, key documentation entry point, high-value product family and principal conversion landing page. Business importance justifies inspecting an individual URL even when its template already appears in the sample.

Then add unusual cases. Useful candidates include recently migrated URLs, pages with redirects, long articles, pages using embedded tools, discontinued products and routes previously reported as blocked. These are deliberately selected exceptions, not a random estimate of overall error prevalence.

Distinguish three labels in your worksheet:

  1. Representative: chosen to inspect a shared implementation.
  2. Critical: chosen because the specific page matters commercially or operationally.
  3. Exception: chosen because there is a plausible failure condition.

A single URL may carry more than one label. Keeping the labels makes the sample explainable to a client or developer and prevents a biased exception-heavy sample from becoming a misleading sitewide failure percentage.

For example, eight failures among twelve deliberately suspicious pages do not mean two-thirds of the website is broken. The valid conclusion is that eight selected cases failed and the affected templates need further investigation.

Run consistent checks and preserve evidence

Use the same core checklist across the sample. Capture the final status, redirect destination, applicable crawl rules, indexing directives, canonical target and presence of the main content. Compare raw and rendered HTML when the browser shows facts that are absent from the initial response.

Save the tested URL exactly, including query parameters. Note the timestamp, tool, request conditions and environment. If a test uses a crawler-like user agent, label it a simulation; it does not prove that a request from the operator's real infrastructure receives the same response.

Avoid reducing each page to a single pass or fail. A page can be reachable but have incomplete text, or readable while pointing its canonical at the wrong destination. Record outcomes by check so that a rendering issue does not conceal an access-policy issue.

For each failure, retain enough evidence to reproduce it without collecting unnecessary personal data. A relevant response header and a short HTML excerpt are usually more useful than an unannotated screenshot. The guide to verifying real AI crawlers explains why identity evidence needs its own treatment.

Expand the sample when findings suggest a pattern

Sampling should respond to evidence. If one old article lacks its body in raw HTML, inspect other articles created by the same migration path. If a regional product route redirects to the homepage, test other routes sharing that locale middleware. Expansion follows a plausible shared cause.

Use a stop rule before the investigation grows indefinitely. For example, after identifying a template defect, inspect a few additional instances to confirm the pattern, then ask engineering to query the full affected population. Automated inventory checks can establish the URL count more reliably than continuing manual inspection.

Keep direct evidence and inferred coverage separate:

Observed: 4 of 4 tested legacy articles omit the main body from raw HTML.
Shared implementation: legacy-article template.
Potential population: 186 URLs from the CMS inventory.
Not established: whether all 186 URLs have the same omission.
Next step: check the full template population after the fix.

This is a hypothetical reporting example, not a result from a real customer. Its value is the distinction between observation and extrapolation. That distinction also helps decide whether a narrow patch or a shared-template fix is appropriate.

Report coverage and schedule the recheck

Give readers a coverage table alongside findings. State the total known inventory, tested URL count, tested families, excluded areas and discovery limitations. If you do not know the full inventory, say so instead of presenting the crawl's discovered count as the website's actual size.

A useful completion statement is: “All selected page families were inspected; two access variants remain untested because regional logs were unavailable.” It says more than “audit complete” and tells the next reviewer exactly what remains uncertain.

For the recheck, reuse the original failing URLs and add neighboring pages from the same family. Preserve at least one unaffected control page to detect an unintended broad change. Replacing the original sample with only newly created pages makes before-and-after comparison difficult.

Assign someone to maintain the matrix. A new template, language or hosting arrangement should trigger a sample update. Otherwise a previously representative sample gradually becomes a record of an older website, even while every listed URL still returns successfully.

Frequently asked questions

How many pages should an audit sample contain?

There is no universal number. Choose enough pages to cover materially different templates, delivery conditions and critical journeys, then expand when findings reveal a shared problem.

Is a random sample better than a template sample?

Random sampling can help estimate prevalence when the population and method are well defined. A template-based sample is often more useful for discovering implementation defects, but its failure rate should not be presented as a statistical sitewide estimate.

No. It establishes results for the checks and conditions tested, while untested pages and future search-system decisions remain outside that evidence.

Start with a free single-page check to identify the evidence you will collect consistently across your sample.

Filed under Strategy and audits. Topics: audits, ai-crawlers, rendering