Answer Engine Optimization: how to get cited by ChatGPT
Search engine optimization gets you onto a list. Answer engine optimization gets you into an answer. That is not a nuance, it is a different place on the screen, and it is quietly redistributing who gets found.
| Question | Weight | What a model looks for | Most common failure |
|---|---|---|---|
| What do you offer? | 20 % | A noun someone could order | „We support businesses“ |
| Who is it for? | 20 % | A named, narrowed group | „For anyone who wants to grow“ |
| What makes you different? | 20 % | A claim that would be false for a rival | „High quality, fair prices“ |
| What does it cost? | 15 % | A from-price, a range or a worked example | „On request“ |
| How do people reach you? | 15 % | Email and phone as text on the page | A form and nothing else |
| Where are you based? | 10 % | City and service area in body copy | Only in the imprint |
People preparing a purchase increasingly type their question into ChatGPT, Claude or Perplexity instead of a search box. What comes back is not ten blue links but a paragraph with two or three named sources. That screen has no positions four through ten. Either you are in it, or you are absent.
That is the whole job of AEO: not ranking better, but being quotable.
Why a model cites anything at all
An assistant with web search does two things in sequence. First it searches, usually through a conventional search engine, and pulls a handful of pages. Then it reads those pages and builds an answer in which individual claims are attributed.
For the second step it needs paragraphs that stand on their own. A sentence beginning „as described above“ drops out, because the model cannot ship the rest of your page along with the answer. So does anything that would fit any provider. When three websites say the same thing, the model names the one it knows something else about.
Which leads to an uncomfortable rule: what gets cited is what is checkable and specific. A number, a method, a restriction, a place. Everything a copywriter added for rhythm is, for this purpose, dead weight.
The six questions it hangs on
From hundreds of audited pages we distilled six questions every prospect asks before buying, and which a model therefore has to be able to answer. They are in the table above with the weight we give them. Three deserve a closer look.
What do you offer
The first question, and the one most pages fail. „We support businesses on their journey“ is not an answer, because it fits every business. A model cannot build an association from it and would rather name someone it is sure about.
The test is simple: does your opening paragraph contain a noun somebody could order? „We build oak staircases for period houses in Bristol“ travels further than any sentence containing the word solution.
What makes you different
The most expensive gap. Your difference is the only reason a model names you instead of the others. And it has to be phrased so that it would be false for a competitor.
„We do not take jobs under twenty square metres“ is a difference. „High quality“ is not, because nobody claims the opposite. The test takes ten seconds: could your closest rival copy the sentence word for word without lying? Then it is not a difference.
What does it cost
Most service websites carry no price. That costs twice. The prospect leaves because they cannot judge the order of magnitude, and the model cannot answer a pricing question with you in it. Asked „what does X cost in Manchester“, it answers with the two providers who published a number.
Three forms commit you to nothing and are entirely sufficient: a from-price, a range, or a worked example. „A kitchen with us usually lands between £12,000 and £20,000“ is an answer.
The technical half: can the model reach the text at all
Everything above is wasted if the assistant cannot read the content. Six measurements settle that in about fifteen minutes.
AI crawlers in robots.txt. Check for GPTBot and OAI-SearchBot from OpenAI, ClaudeBot and anthropic-ai, PerplexityBot, and Google-Extended. Blocks are usually accidental, inherited from a template or set by a plugin. Worth knowing: Google-Extended only governs whether your content feeds Gemini’s training, not your ranking in ordinary Google search. Blocking it costs you no ranking, but it costs visibility in answers.
Content in the delivered HTML. The text has to be in the HTML the server sends. If the browser assembles it from JavaScript, a share of models see an empty page. Some crawlers execute JavaScript, many do not, and none guarantee it. Under roughly sixty words of raw text, a page is an empty frame.
Structured data. An Organization block with address and contact restates the same facts in unambiguous form. That matters most where body copy is ambiguous, such as opening hours and prices. A FAQPage block helps because question and answer are already separated, which is exactly the shape a model quotes in. It does not help when the questions were invented to hold keywords.
llms.txt. A markdown file in your root that says in a few lines what the site is about and which pages matter. It replaces no copy, but it helps with orientation. How to build one is in the llms.txt guide.
The hedge words
Reading other people’s websites, one group of words keeps surfacing that sounds polite and destroys information: bespoke, tailored, holistic, on request, suited to your needs.
These are not mistakes. They simply sit exactly where a number, a name or a restriction could have been. Where one appears, information ends for a model, and usually for the reader too.
The rewrite is unglamorous. „Bespoke advice for your needs“ becomes „A 45-minute first call, free, followed by a written proposal within three working days“. Same space, two checkable statements more.
What AEO is not
It is not a trick, and it is not a schema markup contest. Marking up claims that do not appear in the copy creates a contradiction, and models punish contradictions harder than gaps.
It is also not word count. A 400-word page that answers six questions gets cited more often than a 3,000-word page that talks around them.
And it is not a one-off. Every new service, every new location, every price change is a statement that should be on the page before somebody asks.
The order that works
- Measure. First find out whether a model can reach your text at all. The AEO Check does both halves in about twenty seconds and keeps them apart on purpose: visibility and answerability are never merged into one number, because the remedies have nothing to do with each other.
- Close the three heaviest gaps. Offer, audience, difference. Together they are sixty percent of the weight.
- Add price and contact. A range and an email address as text, not as an image.
- Set structured data, but only for what the copy actually says.
- Publish an llms.txt in your root directory.
- Measure again and ask the assistants yourself.
What this has to do with a template
In a website that is built up front, all of this can be decided in advance. Our AI Website Templates carry the six answers in their copy structure, ship the structured data and include an llms.txt. The text is written without hedge words, because the knowledge sits in the template and the built-in skills hold the line while you rewrite.
That is no substitute for your own claims. Only you know your difference. But the place for it is already there, marked up correctly, exactly where a model looks for it.
Frequently asked questions
Answer Engine Optimization. It is the work of making a language model choose your website as a source for an answer and name you in it. The neighbouring terms GEO (Generative Engine Optimization) and LLMO describe the same field with a different emphasis.
No, it sits on top. Assistants with web search almost always reach for search results before they answer. If you do not rank, you do not get cited. AEO decides whether a ranking turns into a mention.
Ask the assistants yourself, in several phrasings, and record whether and how you are named. Then read your server logs for GPTBot, ClaudeBot, PerplexityBot and OAI-SearchBot. None of this traffic appears in Search Console.
It helps with orientation, not with citation. The file tells a model which pages matter and what the site is about. What gets quoted is still what stands on the page. Text first, file second.
Days to weeks. Assistants with web search read your page at the moment of the question, so a change lands almost immediately. Anything that comes from the training snapshot only shifts with the next snapshot, and nobody controls that.
You do not have to name a fixed price. A from-price, a range or a worked example is enough and commits you to nothing. „On request“, by contrast, reads to a model as a missing value, and it answers accordingly, which a prospect reads as expensive.
Keep reading
- llms.txt: structure, example and what it actually does What an llms.txt is, how it is structured, where it belongs and what it does. With a full example, a template and an honest assessment of its limits.
- AI Website Templates, What They Are and Why They Win What separates an AI website template from an ordinary one, why machine-readable context changes the economics, and when a normal template still wins.
- How to Build a Website with Claude Code (Complete 2026 Guide) A practical guide to building a real website with Claude Code, from setup to launch. What the tool does well and where it needs guardrails.
- Financial advisor website templates: 7 finished sites compared Seven finished website templates for financial advice and wealth management compared, with page count, focus and which firm size each one fits.