Short answer: AI assistants cite pages that rank in search, load fast as plain HTML, and answer a question clearly in the first sentence or two under a heading. To get cited, let the right crawlers in, put the answer first, be specific, keep your company facts the same everywhere, and measure who gets cited for the questions your buyers ask.

More and more people ask ChatGPT, Perplexity or Gemini before they ever open a search results page. Google answers many searches with an AI Overview at the top. When that happens, the assistant either cites you or cites someone else.

Making your content the source those assistants quote has a few names: Answer Engine Optimisation (AEO), Generative Engine Optimisation (GEO), AI search optimisation. The names matter less than the mechanics, which are simple and quite concrete. This guide covers how assistants choose sources, how to write for them, the technical basics, and how to measure results honestly.

How do AI assistants choose what to cite?

AI assistants that search the web work in four steps. They turn your question into several searches, pull the top results, read parts of a few pages, and write an answer with links to the passages they used. Each step throws most pages away, so a page has to survive all four to be cited.

How an AI assistant retrieves and cites a source

One question, four stages. Only a few sources survive to the answer.

1. 1

   Rewrite

   Your question becomes several machine queries, each aimed at a different part of the answer.

   5searches fired

2. 2

   Retrieve

   Each query hits a real search index. If you do not rank here, nothing later can save you.

   50results ranked

3. 3

   Read

   A handful of pages are fetched and split into passages. Slow pages and pages that only render with JavaScript drop out.

   9pages actually read

4. 4

   Compose

   The answer is written from the passages that fit. Those passages get the citation link.

   3links cited

The numbers show the shape of one example run, not a measurement.

Three places a page loses the citation

• It never ranked

  Stage 2 is ordinary search. No ranking, no reading, no citation. SEO is the ticket in.

• The answer was buried

  Stage 3 extracts passages, not pages. An answer in paragraph nine is invisible.

• The source was vague

  Stage 4 prefers a source it can name. No author, no organisation and no date mean lower confidence.

Step 1: your question becomes several searches

Ask “is headless WordPress worth it for a marketing site?” and the assistant does not search for that exact sentence. It splits it into smaller questions, such as cost, speed, downsides and timeline, and searches for each. Google describes the same idea for AI Overviews in its guide to AI features.

So the question you think you are targeting is rarely the one that finds you. A page that fully answers one narrow question gets pulled into many bigger answers.

Step 2: retrieval is ordinary search

Those smaller searches go to a real search index. If your page is not near the top, the assistant never sees it. Good SEO did not go away. It became the ticket in.

Step 3: the assistant reads passages, not pages

A fetched page is split into chunks, and each chunk is scored against the question. You compete paragraph by paragraph. A long guide with the answer buried in the middle can lose to a short page that answers in its first sentence.

Pages also drop out for boring reasons. Content that only appears after heavy JavaScript runs is often missed, and very slow pages get skipped.

Step 4: the citation goes to the passage

When the assistant writes its answer, it links the source of each claim. Two things help: the passage can be quoted on its own, and the source is easy to name. “According to Studio IDD, a design and development studio” is easy for a model to write. “According to a website” is not.

How should you write so AI can quote you?

Write the answer first. If someone read only the first two sentences under each heading, they should already have the answer. Context, details and opinions come after.

The same section, reordered

Assistants extract passages under a time and token budget. The order decides what they see.

Buried answer

Our approach to migrations

The reading budget often runs out around here

About ten weeks for a content site with a handful of templates and no e-commerce.

Result: the assistant cites a competitor.

Answer first

How long does a headless WordPress migration take?

About ten weeks for a content site with a handful of templates and no e-commerce.

The budget runs out here too, but the answer is already above it

Result: a self-contained passage worth quoting.

Same page, same facts, same word count. The only change is which sentence goes first.

• Use the questions people really ask as headings. “How long does a website migration take?” beats “Our approach”. Your sales calls and support emails are the best source of real questions.
• Answer completely in one or two sentences. Then explain. This is the opposite of how most of us learned to write, and it is the change that does the work.
• Be specific. “It depends” cannot be quoted. Name the tools, the steps, the trade-offs and the things that change the answer.
• One idea per section. A paragraph that mixes price, timeline and team cannot be lifted cleanly, so it does not get lifted at all.
• Show where facts come from. Link to the original source for any number. Assistants and readers both trust sourced facts more.

Do not overdo it. A page that is only heading, answer, heading, answer reads like a spec sheet. Let the first two sentences serve the machine, and let the rest serve the human reader, with real opinions and examples. Google says the same thing: write for people first (source).

Which AI crawlers should you allow?

Allow the crawlers that power AI search and live answers. You can block the ones that only collect training data without losing visibility. A common mistake is blocking the wrong ones, for example when a plugin adds a default block list.

Crawler	Company	What it does	If you block it
OAI-SearchBot	OpenAI	Builds the index for ChatGPT search	You can disappear from ChatGPT search answers
ChatGPT-User	OpenAI	Fetches a page when a user’s question needs it	ChatGPT cannot open your page live
GPTBot	OpenAI	Collects data for model training	Your content is not used for training. Search is not affected.
Claude-SearchBot	Anthropic	Indexes pages for Claude’s search results	Less visibility in Claude’s search answers
Claude-User	Anthropic	Fetches a page when a user asks Claude about it	Claude cannot open your page live
ClaudeBot	Anthropic	Collects data for model training	Your content is not used for training
PerplexityBot	Perplexity	Builds Perplexity’s search index	You can disappear from Perplexity answers
Google-Extended	Google	Controls use for Gemini training and grounding	No effect on Google Search or ranking
Googlebot	Google	Normal Google Search crawling, also used for AI Overviews	You disappear from Google entirely

Sources: OpenAI, Anthropic, Perplexity, Google. Bot names change, so check these pages once or twice a year.

There is no separate bot for Google’s AI Overviews. They use the normal Google index, and you control them with the same snippet settings as regular search results.

What technical basics matter for AI search?

Server rendered HTML and speed

Assistants fetch pages with a time limit and usually without running much JavaScript. If your content is built in the browser, it may simply not exist for them. Server rendered or pre-built HTML gets read in full. This is one of the reasons to move off a heavy theme, which we cover in our guide to when headless WordPress is worth it.

Structured data

Schema.org markup states facts that assistants would otherwise have to guess. The ones worth your time:

• Organization: who you are, where you are and your official profiles. This makes you easy to name.
• Article with a real author and dates. Anonymous, undated pages are trusted less.
• FAQPage on question and answer sections. Google stopped showing FAQ rich results for most sites in 2023, but the markup still tells machines clearly which passage answers which question.
• Service for what you sell, and BreadcrumbList so a page is not an orphan.

llms.txt

An llms.txt file is a short text file at the root of your site that describes it and links to your key pages. It is a proposal, not a standard, and support by the big assistants is unclear. It takes twenty minutes to write. Do it, but do not pay anyone for a project around it.

The same facts everywhere

Your company name, address, services and key facts should match on your site, LinkedIn, Google Business Profile and the directories in your field. Assistants compare sources. Conflicting facts make them less sure about you.

How do you measure AI citations?

Rank tracking does not tell you whether you are cited, and AI visibility tools each measure in their own way, so their numbers rarely match. A simple manual method is slow but transparent.

1. Write down 30 to 50 questions your buyers really ask, in their own words. Use support emails, sales calls and your site search. Then freeze the list, so months can be compared.
2. Once a month, ask each question in ChatGPT, Perplexity, Gemini and Google, and note who gets cited: you, a competitor or nobody.
3. Track your share of citations against named competitors. It is the number that stays meaningful when the assistants change.
4. Watch referral traffic from chatgpt.com, perplexity.ai and gemini.google.com in your analytics. It will be small, but these visitors arrive already knowing what you do.

What a citation baseline looks like

Forty buyer questions, four assistants, once a month. A spreadsheet is enough to start.

Question asked	ChatGPT	Perplexity	Gemini	AI Overviews
best headless CMS for a marketing site	competitor	you	competitor	competitor
how much does a WordPress migration cost	you	you	competitor	you
product design studios in Bulgaria	competitor	competitor	nobody	competitor
how long does it take to build an MVP	nobody	you	competitor	competitor
is Elementor bad for Core Web Vitals	you	competitor	competitor	nobody
…and 35 more questions on the same sheet. Here 6 of 20 checks cite you: 30%.

• You are cited
• A competitor is cited
• Nobody in your market

Share of citations

• You30%
• Competitor A41%
• Competitor B17%
• Competitor C12%

Same 40 questions, checked monthly.

Track the share, not the rank. A five-point move from one month to the next is a real result; one lucky answer is not.

Remember what you cannot see: people who read an answer that mentioned you and never clicked. Treat citation share as the early signal and searches for your brand name as the later one.

A four-week plan

1. Week 1, access and baseline. Check robots.txt against the table above. Make sure your key pages show their content in “view source”, not only in the browser. Write your question list and record who gets cited today.
2. Week 2, rewrite five pages. Pick your five most important pages. Turn headings into questions. Put a complete, specific answer in the first two sentences under each. Do not change URLs.
3. Week 3, structured data and facts. Add Organization, Article and FAQPage markup. Make your company facts match everywhere. Publish llms.txt. Fix any page that is slow to respond.
4. Week 4, measure and write. Run the questions again and expect little change so early. Then pick three questions where nobody is cited and write the best answer on the internet for each.

What AI search optimisation does not fix

• It does not make a business different. If your answer is the same as everyone else’s, the assistant writes one sentence and cites whoever it read first.
• It does not create demand. If nobody asks about your field, there is nothing to be cited for.
• It does not replace expertise. Pages that get cited again and again have a method, an example or an opinion that only comes from doing the work.

Keep reading

• Headless WordPress: when it is worth it and when it is not: fast, server rendered pages are one reason assistants read some sites and skip others.
• The six-week MVP: how AI changes the design process: what AI really speeds up in building products, and what it does not.

A fast site that machines can read starts with how it is built. See our web development work, or our AI knowledge assistants, which make your own documents answerable with sources. Questions? Get in touch.