Every new client site gets the same first move from me: an audit before a single word of new copy gets written. Not a vanity report with a score out of 100. A script I actually run, page by page, that tells me whether the site can get cited by an AI surface at all before I worry about ranking it higher on one.
I built this framework by running the pre-launch SEO/AIO/GEO/AEO/SXO/LLMO checklist on enough sites that it stopped being a one-off list and turned into a repeatable audit. This is the exact framework, the four passes, the questions each one answers, and the table I fill in by hand on every engagement.
Pass 1: can the page even get crawled and parsed
This is the pass most audits skip because it's boring. I don't. A site that writes perfect answer-first copy but blocks crawlers or buries content behind client-side rendering never gets read in the first place, so none of the later passes matter until this one clears.
- Fetch the page with JavaScript disabled. If the main content doesn't show up in the raw HTML, most AI crawlers won't see it either. I check this with a plain curl request, not just a browser render.
- Check robots.txt against the known AI crawler user agents. GPTBot, ClaudeBot, PerplexityBot, Google-Extended. A blanket disallow on one of these is usually an accident from a template default, not a deliberate choice, so I flag it either way and let the client decide.
- Confirm the sitemap is current and submitted. New pages that never made it into sitemap.xml take weeks longer to get discovered. On johnarce.com every new page goes into the sitemap in the same commit as the page itself, I hold client sites to the same rule.
- Test canonical tags for self-reference. A canonical pointing at the wrong URL silently tells crawlers to ignore the page entirely. I've found this exact bug on three separate client sites, always from a copy-pasted template.
Pass 2: entity consistency and schema validity
AI systems resolve who is saying something before they decide whether to trust it. That's entity resolution, and it happens before content quality ever gets weighed. If the name on the site, the name on LinkedIn, and the name in old press mentions don't match exactly, the model treats that as noise rather than a trust signal.
I pull every schema block on the site through a validator, not a glance. Schema that doesn't parse is often worse than no schema at all, it tells a crawler the site isn't maintained. I covered why this matters more than backlink volume now in how to dominate SEO in the age of super intelligence.
| Check | Tool | Pass condition |
|---|---|---|
| Organization/Person schema present | Rich Results Test, Schema.org validator | Valid JSON-LD, no required fields missing |
| sameAs entries match live profiles | Manual spot-check | Every linked profile resolves and matches the entity name |
| Name consistency across the web | Site search for brand name variants | One name, one spelling, everywhere it's mentioned |
| FAQPage / Article schema on content pages | Rich Results Test | Valid and matching the visible on-page content exactly |
| BreadcrumbList schema | Rich Results Test | Present on every page with a URL depth greater than one |
FAQPage schema that lists five questions in the JSON-LD block, but the visible page only shows three of them in the rendered accordion. Google's own guidance treats this as a mismatch between structured data and visible content, and it's grounds for the rich result getting suppressed entirely. It's an easy miss because the schema was written once and the copy got edited after, without anyone going back to sync the two.
Pass 3: is the content actually structured to be an answer
This is where most of the real work happens, and it's the pass clients usually expect the audit to start with. A page can be fully crawlable and schema-valid and still never get cited, because nothing on it reads like a direct answer to a real question.
- Answer-first structure. I check whether the first two sentences after any H1 or H2 directly answer the implied question, no scene-setting first. If I have to read past a paragraph of throat-clearing to find the actual answer, I flag the section.
- Question-shaped headings. Headings phrased as the actual question a reader or a model would ask, not a generic label. "What does HubSpot onboarding cost" beats "Pricing" every time I've tested it.
- Tables for anything comparative. Every AI surface I've tested extracts table rows more reliably than the equivalent comparison written as prose. If a page has three paragraphs comparing options, I convert it to a
.compare-table-style layout. - A visible, accurate freshness signal. A real "updated on" date that matches
dateModifiedin the page's schema. I've seen pages with a schema date six months newer than the actual last edit, which is worse than no date at all once a model starts checking. - Internal links into a cluster, not an island. A page with zero inbound links from related content gets cited less often than one that's clearly part of a topic cluster. I check this with a simple sitemap crawl plus a manual pass over the homepage and category pages, since sitemap-only crawls miss homepage-linked pages.
Pass 4: verification, not vibes
The first three passes produce a list of fixes. This pass is where I confirm the fixes actually moved something, instead of assuming structural changes worked because they look right on paper.
For each page flagged in passes 1-3, confirm the fix against the live
URL, not the local draft.
1. Re-fetch the page with JavaScript disabled, confirm the content
that was missing is now present in raw HTML
2. Re-run the fixed schema through the Rich Results Test, confirm
zero errors and zero warnings
3. Re-check the sitemap for the page's presence and lastmod date
4. List any fix that still fails after reapplying it, with the
specific error returned
Output a table: page URL, pass, status (fixed/still failing), and
the exact next action if still failing. I don't trust a fix until it's been re-checked against the live page. Schema that validates in a local file can still break once it goes through a CMS templating layer, and a sitemap entry added in a draft PR doesn't count until it's actually deployed.
The scorecard I hand the client
Every audit ends with one page, not a slide deck. A simple pass/fail grid across all four passes, plus a short list ranked by effort versus impact. Clients read the grid in under a minute and know exactly where their site stands before we touch a single word of new copy.
| Pass | What a fail usually means | Typical fix effort |
|---|---|---|
| Crawlability | Client-side rendering hiding content, or an accidental robots.txt block | Low, usually a config change |
| Entity & schema | Inconsistent naming, broken or missing JSON-LD | Medium, needs a content and dev pass together |
| Answer structure | Content written for humans skimming, not for direct extraction | High, this is a rewrite, not a patch |
| Verification | Fixes assumed to work but never re-checked live | Low, just discipline to close the loop |
Key takeaways
- Run crawlability first. Perfect content a crawler can't parse never gets a second chance to be read.
- Entity consistency and valid schema come before content quality in how models decide to trust a source.
- Answer-first structure, question-shaped headings, and tables over prose are the highest-leverage content changes I've tested across client sites.
- Never close an audit item on the plan. Re-check every fix against the live URL before marking it done.
- A one-page scorecard beats a long report. Clients act on what they can read in a minute.
What I'll do next
I'm turning the verification pass into a repeatable script I can run against any client domain without rebuilding it each time, so pass 4 stops being manual copy-paste between tools. I'm also adding a fifth check to pass 2 for Product schema completeness on any client site that sells something directly, since agentic shopping surfaces are starting to select off that data the same way search models select off FAQPage schema.
FAQ
How long does a full audit take? +
Do I need to fix everything the audit finds before launch?
Is this framework different from a regular technical SEO audit?
Can I run this audit on my own site without hiring someone?
What's the single most common failure you find in pass 1?
Want this built for your team?
I design AI agents and growth automation that run without babysitting. If that sounds useful, let's talk.
Get in touch โ