We read your site the way a language model reads it and ask it the six questions every prospect asks before buying. You see which ones your copy answers and which it does not.
One check per day and device, no sign-up. It takes about twenty seconds because we read up to six subpages. The address is not stored.
Reading your pages …
Built from your own material. Read it through and put your numbers in, then it can go on the page.
Every other tool checks signals and stops there. Signals tell you whether a model can reach your copy. They say nothing about whether an answer stands in it.
Six measurements on the source: crawlers, raw HTML, llms.txt, sitemap, structured data, title and description. This is the question whether the door is open. It can be settled in a morning, mostly in files, not in copy.
A language model reads your pages and answers six questions from them alone. No prior knowledge, no benefit of the doubt. This is the question whether anything stands behind the open door. It costs an afternoon of writing and decides whether you get quoted.
A website can invite every crawler and still say nothing worth quoting. That is why the two numbers stay apart.
They are not invented. They are the questions a prospect brings to any vendor site. A model asks them on their behalf.
All of them on the publicly available source. None of them needs access to your website.
Your robots.txt is checked for six names: GPTBot, ClaudeBot, PerplexityBot, Google-Extended, CCBot and Bytespider. Blocking is usually accidental, inherited from a boilerplate.
Your copy has to sit in the HTML the server sends. If the browser assembles it first, some models see an empty page. The check counts the words and looks for the telltale empty container.
A file that states in note form what your website is about. Young standard, ten minutes of work. We have a generator for it.
The list of all your pages, either declared in robots.txt or sitting at the usual address. Without it a crawler finds only what is linked.
JSON-LD tells a machine unambiguously what is name, address, price and opening hours. The check names the types it found instead of merely counting them.
Both get quoted in answers as the short version. The check looks at whether they exist and whether the length carries: title up to 65 characters, description from 80.
Reference
AEO stands for Answer Engine Optimization. The difference from a search engine is small and consequential: a search engine shows your page, an answer engine talks about it. What it says is decided by your text.
Language models with web search cite sources. They pick paragraphs that answer a question completely, without needing the rest of the page. A paragraph starting with “as described above” never gets quoted, because it does not stand on its own.
Of the six questions we ask, price fails most often. The reason is almost always on request. A model can do nothing with that and answers “no prices are given”, which a prospect reads as expensive. An entry price, a range or a worked example all fix it.
We flag wording such as bespoke, tailored and on request separately. They are not mistakes, they simply sit where a number could have stood. Wherever such a word appears, the information ends for a model.
We measure the word count in the delivered HTML, so in what the server sends before any script runs. Some crawlers execute JavaScript, many do not, and none guarantees it. Below sixty words of raw text a page counts as an empty shell.
We check robots.txt against GPTBot and OAI-SearchBot from OpenAI, ClaudeBot and anthropic-ai, PerplexityBot and Google-Extended. That last one only governs training for Gemini, not ordinary Google search. Blocking it costs you no ranking.
An Organization block with address and contact gives a model the same detail again in unambiguous form. That is worth most where prose is ambiguous, such as opening hours and prices. Three types or more on one page score full marks.
An FAQ block helps because question and answer are already separated, which is exactly the shape a model quotes. It does not help when the questions are invented to place keywords. Take the five questions people actually ask you on the phone.
We calculate visibility and answerability separately and deliberately never add them. A site can be technically flawless and mute in substance, and the reverse exists too. A single score would blend both cases into one middle number nobody can act on.
We read text, not a browser. Loading time, contrast, keyboard operation and how the page behaves on a phone are outside this check. Those need a page that is actually rendered, and that is a different job with different tools.
Answer Engine Optimization. Type a question into ChatGPT, Perplexity or Google's AI overview and you get an answer instead of ten blue links. That answer carries either your name or somebody else's. AEO is the work of making sure it is yours.
SEO gets your page into a list someone picks from. AEO lets a model build an answer out of your copy. The first rewards findability, the second clarity. Much of the technical work is the same, what is asked of the writing is not.
Because the causes have nothing to do with each other, and neither do the remedies. Visibility is repaired in an hour of file work, answerability costs an afternoon of writing. One averaged number would hide which of the two you are looking at.
It reads up to six of your pages, preferring pricing, services and contact, and puts them to a language model under a strict instruction: use no prior knowledge. If the answer is not in the copy, it counts as unanswered, even where the model knows your company.
No, only the website. The difference matters enough to us that it sits as a rule in the instruction to the model. “The offering is not recognisable on the page” is a finding about your copy. “The offering is weak” would be presumption.
Where an answer is missing, something almost always stands in its place anyway. The tool quotes those spots verbatim so you can find them again on your own site. Usually they are turns of phrase like “tailored solutions” or “fair terms” that fill a line without saying anything.
Read it and adapt it. It is built from your own material and picks up your form of address, but a suggestion does not know your numbers. As a scaffold it saves you the start, as finished copy it will not do.
No. The address is used for the check and not stored afterwards. All that is counted is how often the check runs per day, so it does not get overloaded.
Every check reads up to six pages and then asks a language model. That costs money, and the tool is free. One check a day is enough to look at a website seriously, and it keeps the abuse out.
Three ways are open: fix it yourself using the SEO chapters of the documentation, take a template that already has the six answers laid out as sections, or have it done with the AI Website Service.
Each answers one question about your website, in under a minute, without an account.
Website, the guideline and seven skills sit ready in one ZIP file. Pick your Template and start today, for yourself or your clients.
Or we build it for you: see the AI Website Service
34 finished industry websites, each with built-in knowledge and seven skills. Structure, copy, structured data and llms.txt are set before you change a single thing.
One payment, everything included. We show this at most once.
We would like to understand how this site is used with Google Analytics, loaded via Google Tag Manager. Nothing loads and no cookie is set before your ok. You can change your choice anytime in the privacy policy.
Choose what may be active. You can change your selection anytime in the privacy policy.