How to Audit Your Site's AEO: The Step-by-Step Checklist

If you want to know whether your site is set up to get cited by AI answer engines, you don't need a tool or a vendor to start — you need to run seven checks in order and write down where each one breaks. That ordered audit is below. I build answer engine optimization for a living, so this is the exact sequence I run against a real site, not a theoretical checklist. It's the diagnostic companion to the full answer engine optimization playbook; work through it top to bottom, because each step depends on the one before it. There's no point perfecting your schema if a firewall is blocking the crawler at the gate.

1. Can AI crawlers reach and render your pages?

Everything downstream is wasted if the bots never see your content, so this is check one. Three failure points, in order of how often they bite:

• robots.txt. Open yoursite.com/robots.txt and look for rules that block AI user agents — GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot. A disallow here locks you out of that engine entirely.
• Firewall and bot management. This is the silent one. Cloudflare, WAF rules, and "block AI scrapers" toggles reject crawlers without ever touching robots.txt, so the file looks clean while the bot gets a 403. Check your firewall's bot settings directly.
• Rendering. Fetch a page with curl -A "GPTBot" yoursite.com/page and read the raw HTML. If your actual answer content only appears after JavaScript runs, many crawlers never see it — they read the empty shell. The answer has to be in the server-rendered source.

While you're here, confirm you have an llms.txt file pointing these systems at your best content. It's not a ranking factor, but it makes your important pages trivially findable — the full implementation is in llms.txt and AI crawlers. If you fail step one, stop and fix it. Nothing else counts until crawlers get in.

2. Do you have a page that answers each real question?

Next, map coverage. Write down the real questions your customers ask — in their words, not your marketing language — and for each one ask a blunt question: is there a page on my site whose job is to answer that? Not a page that mentions the topic. A page that answers the question.

Most sites have a big gap here. They have a homepage, a services page, and a blog, but no page that directly answers "how much does X cost in [city]" or "is X safe for Y." Answer engines retrieve at the passage level, so a question with no dedicated page and heading almost never gets you cited. Build the grid: question in one column, the URL that answers it in the next, blank where nothing exists. Every blank row is a query you can't win.

3. Is the answer extractable?

Having a page isn't enough — the answer has to be liftable. Open each page from step two and check it against three tests:

• Answer-first. Does the true, specific answer appear in the first sentence or two under a heading that matches the question? Or is it buried in paragraph nineteen after a long windup? The passage selector rewards the chunk that is the answer, not the chunk that promises one is coming.
• Specific. "Our service is fast" isn't quotable. "Most installs are finished same-day, usually under four hours" is. Numbers, named methods, and concrete comparisons extract cleanly; vague claims get paraphrased from a competitor instead.
• Self-contained. Read the key sentence out loud with no surrounding context. Does it still make sense and stay true? If it only works in situ, the model won't risk quoting it.

A hard rule while you audit: never fix an extractability problem by inventing a specific. A made-up number that gets cited is a liability that detonates the moment someone checks. Use real figures you can stand behind. The deeper mechanics are in how to get cited by AI search.

4. Is schema present and valid?

Now check the machine-readable layer. For each key page, run it through Google's Rich Results Test and the Schema.org validator and confirm three things: the JSON-LD parses without errors, the type is correct for the page (FAQPage for Q&A, Product for products, Article for guides, Organization for your entity), and — most important — every value in the markup matches what a human actually sees on the page. Schema that contradicts the visible content gets distrusted, not rewarded. Structured data won't rescue a thin page, but on a genuinely useful one it removes friction between your answer and the engine extracting it. Full detail lives in schema and JSON-LD for AI search.

5. Run the coverage test across engines

This is the step most people skip, and it's the one that tells the truth. Take your fixed prompt list from step two, ask each question in ChatGPT, Perplexity, and Google's AI mode, and log a simple grid: prompt down the side, engine across the top, and in each cell whether you were cited, who was cited instead, and whether the citation represented you accurately.

That prompt-coverage rate — the share of real questions where an engine names you — is your actual AEO scoreboard, far more honest than "AI traffic," which attribution still badly undercounts. Re-run it monthly so you're reading a trend, not a snapshot. The full measurement method is in track your AI visibility and how to measure AEO.

6. Check your corroboration

When you're not cited in step five, look at who is — and then look at why. Answer engines are consensus machines; they lean toward businesses that multiple trusted sources describe the same way. So the audit here is off-site: do reviews, directories, comparisons, and third-party mentions of you exist on the sources these models already read? Are they consistent and specific? A great page with no corroboration around it often loses to a weaker page whose claims are echoed across the web. Note every query where the winner's edge is clearly corroboration, not content — that's your off-site work list.

7. Check entity consistency

Last, audit how your business is described everywhere a machine might read it. Is your name, category, location, and core offering stated the same way on your site, your schema, your social profiles, and third-party listings? Contradictions — a different business name here, a stale address there, an inconsistent description of what you do — make the model unsure who you are and less willing to name you confidently. One clean, consistent entity is easier to cite than a fuzzy one.

How to prioritize the fixes

You'll finish this audit with a list of breaks. Fix them in this order:

1. Crawler blocks (step 1) first, always. They zero out everything else. One firewall rule can be your entire problem.
2. Missing answer pages (step 2) for your highest-value questions next — you can't be cited for a question you don't answer anywhere.
3. Extractability (step 3) on the pages you already have — often the cheapest, highest-leverage fix, because the content exists and just needs restructuring.
4. Schema (step 4), then corroboration and entity consistency (steps 6–7), which compound more slowly.

Order by blast radius: fix the break that's blocking the most queries at the lowest cost first.

Running this by hand on ten pages is a good afternoon. The catch is that a real site isn't ten pages — it's hundreds, and the questions your customers ask map to the long tail, not your five bestsellers. Auditing and fixing all seven layers across an entire catalog, then re-testing every month, is the actual work, and it's the problem I built RunOctopus to solve at scale. But you don't need it to start. Run the seven checks, write down where each breaks, and fix the highest-leverage break first.