What is Answer Engine Optimization? How it works Evidence Blog Offers Contact Get a free audit
Checklist

The AEO checklist

Twenty-four checks, grouped by gate and ordered by what failing each one actually costs you.

Work the gates in order: retrieval first, because a page a model cannot fetch fails everything downstream; then extraction, then entity, then corroboration, then measurement. Fixing gate three while gate one is broken changes nothing.

Most published AEO checklists are essays with numbers in front of the paragraphs. This is the actual list, in working order. Nothing here is gated, and none of it requires hiring anyone.

Gate 1 — Retrieval: can a model fetch the page?

Binary, cheap to fix, and the most expensive thing to get wrong. Everything below is irrelevant until this passes.

  1. Name the retrieval agents in robots.txt. OAI-SearchBot, ChatGPT-User, PerplexityBot are fetched during a live query. Blocking these is the costly mistake.
  2. Decide separately on training crawlers. GPTBot, ClaudeBot, Google-Extended, CCBot affect long-term familiarity, not live citation. Blocking them is a defensible choice; blocking them by accident is not.
  3. Check what your CDN does by default. Managed “AI bot” rules block agents your robots.txt allows. The file says yes and the edge says no.
  4. Verify status codes per agent, not in a browser. A 200 for you can be a 403 for a named agent.
  5. Remove the JavaScript dependency from anything you want quoted. If the answer only exists after hydration, assume it is not read.
  6. Check the sitemap is complete and current, with lastmod generated from real change dates rather than maintained by hand.
  7. Ship an llms.txt — low cost, contested value, and it is read by some agents. Treat it as cheap insurance, not a strategy.
  8. Confirm pages are not orphaned. Anything reachable only from a search box may as well not exist.

Gate 2 — Extraction: is there anything worth quoting?

  1. Write headings as the literal question a buyer asks. Not “Our approach”.
  2. Answer in the first forty words under the heading, in a self-contained sentence that survives being lifted out of the page.
  3. Replace narrative with structure where the content is comparative — tables are lifted far more readily than paragraphs.
  4. Define your terms on the page, once, plainly, rather than assuming the reader arrived with context.
  5. State numbers with their source and date inline. Unattributed figures are weak quotes.
  6. Cut hedging. “Can help to potentially improve” gives a model nothing to assert.
  7. Keep one question per section. Two questions in one block means neither is cleanly extractable.
  8. Add an FAQ block with the objections, not the softballs.

Gate 3 — Entity: does the model know who you are?

  1. Organization schema with a stable @id, identical across every page.
  2. A real Person node for whoever is named as author or founder, with sameAs pointing at profiles that resolve.
  3. Check your own name for collisions. If your brand shares a word with something else entirely, say what you are not — disambiguatingDescription exists for this.
  4. Make the description identical everywhere — schema, llms.txt, footer, third-party profiles. Models quote these strings verbatim.
  5. Audit the description you already publish. When an assistant describes you wrongly, it is usually repeating your own structured data back at you.

Gate 4 — Corroboration, and Gate 5 — Measurement

  1. Find the sources models already cite for your category — roundups, directories, trade publications — and get named in them. This is outreach, not on-page work.
  2. Freeze a prompt set of the questions your buyers actually ask, and do not change it when results are inconvenient.
  3. Run every prompt more than once and record the spread, not the average. A single run cannot be distinguished from noise, and a score built on one is a number with no error bar.

The ordering matters more than the list. A site that fixes schema and content while a retrieval agent is blocked at the CDN has spent weeks on gates two and three while gate one is still closed. Work top to bottom.

How to use this against a provider

Hand this list to anyone selling you AEO and ask which of the twenty-four they check, and how. The answer to check 24 is the one that separates measurement from theatre.

FAQ

Related questions

Crawler access, per named agent. A page an assistant cannot fetch fails every later check automatically, and access is the cheapest thing on the list to fix. Schema and content work done while a retrieval agent is blocked changes nothing.

It is cheap and read by some agents, and its value is genuinely contested — Google has said publicly it is not used for AI features while other vendors recommend it. Ship it as low-cost insurance, not as a strategy, and do not let it displace crawler access or schema work.

More than once, always. Assistant output is non-deterministic, so a single run produces a number that cannot be distinguished from session variance. Record the spread across runs rather than the average, and keep the prompt set frozen between measurements.

It overlaps on crawlability and structure and diverges everywhere else. SEO checks whether a page can rank in a list of links. This checks whether an assistant can fetch it during a live query, lift a self-contained statement from it, and identify you as the source worth naming.

Most of them, yes — robots.txt, schema, heading structure and the prompt test need no tooling beyond a browser and patience. The parts that are hard to do by hand are per-agent status verification and repeated multi-engine runs, which is what the free check automates.

Further reading

Or have it run for you.

Send your domain and the check runs these against your site. Report in 48 hours, no call.

Get a free audit See offers