Enter your domain and get an answer engine optimisation score out of 100 in about ten seconds, with every failing check explained. What ChatGPT, Perplexity, Claude and Google AI Overviews can read on your site, and what they cannot. Free, no signup, no email required.
Answer engine optimisation is the work of making a site usable by the systems that answer a question directly instead of returning a list of links. An AEO checker scores how much of that work is done. Where SEO competes for a position on a results page, AEO competes to be the source the answer is built from, which needs three things a ranking does not prove you have: content a model can parse without inferring, an identity it can tell apart from companies with similar names, and claims it can quote without guessing.
GEO, generative engine optimisation, is the same discipline under a narrower name, usually meaning citation inside generated answers. Nothing in the scan changes depending on which label you use.
| SEO | AEO | |
|---|---|---|
| What it competes for | A position on a page of ten blue links. | Being one of the sources the answer is assembled from. |
| Who consumes the page | A crawler that indexes, then a human who clicks and reads. | A model that extracts facts and never renders the page to anyone. |
| What it reads | Rendered HTML, links, titles, headings, page speed. | Structured data, discovery files, machine-readable pricing, callable tools. |
| Where you lose | Outranked by a page with more authority on the same term. | Summarised wrongly, or left out of a comparison because a number was not parseable. |
| How you measure it | Rankings, impressions, clicks. | Whether the checks exist at all, and what a model says when asked about you. |
AEO is a phrase nobody can act on until someone says what is actually being looked at. Here is the whole of it, in the order an answer engine hits them. Every check below is scored, and every failure is named on your result page rather than rolled into a number.
Can an agent find you and be confident it is you?
Before an agent can recommend you it has to know you exist and tell you apart from companies with similar names. This is where most sites lose points first, and it is the layer least related to conventional SEO: ranking for your own brand name is necessary but nowhere near sufficient.
Can it read your content and understand your business?
An agent extracting facts does not read the way a person does. It needs your claims in a form it can parse without inferring, and it needs to reach them without executing JavaScript. Pricing is the single most common gap: if an agent cannot find what you charge, it cannot put you in a comparison.
Can it do something, or only look?
This is the layer almost nobody has, and it is where the category is heading. An agent that can read your prices but cannot start an enquiry has to hand the user back to a browser. Tools exposed on the page, through the W3C WebMCP draft, let it act instead. Chrome and the ChatGPT desktop browser already discover them.
Can it transact?
Only scored where it applies. Agent-initiated payment protocols are early and most businesses correctly have nothing here. It is listed so you know it was considered rather than missed.
We ran this on our own site, scored 74, and spent a day closing the gaps to reach 83. These are the checks that moved it, what each is worth, and how long it actually took. Ordered by return, not by difficulty.
| Check | Points | Effort |
|---|---|---|
| llms.txt missingPublish a markdown file at /llms.txt: what the business is, what it sells, the canonical URLs, and when an agent should reach for you. Use real markdown links rather than bare URLs, because at least one scanner checks for them. This is the single cheapest thing on the list. | 1 to 2 | Minutes |
| No resource catalogPublish an Agentic Resource Discovery catalog at /.well-known/ard.json listing every machine-readable thing you expose. Identifiers must be domain-anchored urn:air URNs. We had ours on urn:ai and every entry failed validation silently, costing three separate checks. Add a trustManifest to each entry with publisher, contact, licence and canonical. | 1 to 5 | An afternoon |
| No markdown twin of pagesServe a markdown version of each page, reachable by appending .md, by sending Accept: text/markdown, or by detecting a known AI crawler user agent. Open each document with a frontmatter block carrying title, description, canonical and last-updated, so an agent gets metadata without scraping. Advertise it with an RFC 8288 Link header. | up to 5 | An afternoon |
| robots.txt says nothing about AIA blanket User-agent: * tells an agent nothing about intent. Name the AI crawlers you want individually and allow them explicitly, so the access reads as deliberate. Skip the Schemamap directive: validators report the whole file as invalid on seeing it, and /.well-known/ard.json does the same job on a path agents already check. | 1 to 2 | Minutes |
| No auth.mdPublish /auth.md, the WorkOS discovery path, stating how an agent gets access. If everything you expose is open and keyless, say exactly that. The file scores because it saves an agent from probing, not because it grants anything. | up to 4 | Minutes |
| Pricing an agent cannot readIf an agent cannot find what you charge, it cannot put you in a comparison. Put pricing on a page it can parse, and publish a machine-readable /pricing.md. This is a commercial decision as much as a technical one, and worth making deliberately either way. | up to 5 | Minutes |
| No A2A agent cardPublish /.well-known/agent-card.json: who you are, what you can do, how to be reached. Declare your skills with clear descriptions of when to use each one. | 2 | Minutes |
| No WebMCP toolsThe largest single item, and the one that changes what an agent can do rather than what it can read. Register tools on document.modelContext, not navigator.modelContext, which is the mistake most implementations make. No separate server is needed; the tools live on the page. Write the descriptions for the agent: say when to call each one, and what not to do. | up to 5 | A project |
| Content needs JavaScriptMany agents fetch without executing JavaScript. If your content only exists after hydration, they see an empty shell. Server-render the substance. | 1 to 3 | A project |
| Weak entity dataJSON-LD identifying the organisation, with url, logo, address and sameAs pointing at profiles that actually resolve. Check them: ours pointed at a LinkedIn URL that returned 404, which is worse than claiming nothing. | 1 to 2 | Minutes |
Each of these has a full write-up in the agent web reference, including the parts we got wrong first. Nothing here is proprietary and none of it requires us. If you would rather not spend the day, that is what the audit is for.
5 working days, fixed scope. It starts from your scan result and goes through to the process behind the site. Yours to keep and act on with anyone, including a competitor of ours.
The scan gives you a score and the names of what failed. The audit tells you what each one actually costs you: which agents cannot read what, which questions about your business they will get wrong, and which of the gaps are worth closing versus which are noise at your scale.
What to change first, what it takes, and what it moves. Some of these are an afternoon of work, a discovery file or a structured data block. Others are architectural. We separate them so you can do the cheap half immediately without waiting for a project.
Being readable is not the same as being understood. We check what an answer engine actually reports when asked about your business, your pricing and your category, and where it is drawing that from. This is usually the most uncomfortable page in the report.
A site is where an enquiry arrives; the cost sits in what happens next. We map one process end to end, including the exceptions people handle without thinking about them, and mark which steps a machine can take, which it can propose for approval, and which it must never touch.
Expressed as hours per month against the process as it runs today, so there is a baseline to measure against afterwards rather than an impression. If automating it will not pay, the report says so.
Which rules govern this workflow in your market and sector, and what they require of the design. For UK work that now means Articles 22A to 22D, which replaced UK GDPR Article 22 on 5 February 2026. Plus the parts we would not automate, and why.