How AI answers get written — and how to be in them
Answer engines do not rank pages, they retrieve passages and quote the ones they can attach a claim to. This is the reference for how that works: what decides whether you are quoted, how to write so you can be, how to measure it without fooling yourself, and what to let AI crawlers read.
39 entries, each written twice — once for someone who has never met the term, once for someone who has to implement it. Last revised 2026-09-07. Also available as one markdown file, and every entry answers Accept: text/markdown.
How answer engines actually work
The mechanics that decide what gets quoted. Everything else on this page follows from these.
- Generative Engine Optimisation (GEO)Making your brand the one an AI assistant names and cites when someone asks it a question in your category — the answer-engine equivalent of ranking.
- Grounding & RAG (retrieval-augmented generation)The technique that makes an assistant search the live web, read what it finds, and write its answer from those pages — which is why citations exist at all.
- Chunking & passage retrievalRetrieval systems do not read your page — they read paragraph-sized slices of it, and each slice has to make sense alone.
- Embeddings & semantic searchText turned into coordinates, so a machine can find passages that mean the same thing even when they share no words with the question.
- Query fan-outOne question typed by a person becomes several different searches run by the machine — so you compete for questions nobody ever typed.
- Google AI Overviews & AI ModeThe AI-written summary Google places above the normal results, with a handful of linked sources — often answering the question so completely that nobody scrolls.
- How the engines differChatGPT, Claude, Perplexity, Gemini and Copilot retrieve differently, cite differently and change at different speeds — so one engine's answer is not a proxy for the rest.
Being the source that gets quoted
What makes one passage citable and another invisible — the half you control directly.
- AI citations — why one source gets quotedThe links an assistant shows beside its answer, chosen at answer time for how well a passage supports a specific claim — not for how good the site is overall.
- Entities & the knowledge graphBeing a thing a machine recognises — a company with a name, an address and confirmed profiles elsewhere — rather than a string of characters that happens to appear on a page.
- Where AI gets its answers about your categoryThe handful of sites an engine keeps quoting when someone asks about your market — usually not the vendors' own sites, and usually the fastest route to being named.
- Structured data (schema.org / JSON-LD)A machine-readable description of what your page is about, which helps search engines show rich results and helps AI understand your content.
- Canonical URLA tag that tells search engines which address is the ‘official’ one when the same page can be reached by several URLs.
- Meta descriptionThe short summary of a page that search engines often show under its title in results — your chance to earn the click.
Measuring AI visibility
How to get a number that survives scrutiny, and how to tell the honest tools from the confident ones.
- Share of voice in AI answersThe proportion of AI answers about your category that name your brand — the closest thing answer engines have to a ranking position.
- Designing the prompt set you trackChoosing the questions you measure — the decision that determines whether AI-visibility tracking tells you anything, and the one most often got wrong.
- Sampling and variance in AI answersThe same question asked twice can name different companies — so AI visibility is a statistic to be sampled, never a position to be looked up.
- Choosing the competitor setThe brands you measure yourself against — the denominator that makes share of voice mean anything, and a list the engines will happily correct for you.
- Which AI-visibility numbers are worth reportingFour numbers that survive scrutiny — presence rate, share of voice, citation share and sentiment — and the vanity metrics that do not.
- Proving a change workedEstablishing a baseline before you change anything, and knowing how long to wait — because in a noisy channel, an unmeasured improvement is indistinguishable from a lucky week.
- AI referral traffic & attributionVisits that arrive from an AI assistant — a small, high-intent trickle that most analytics setups quietly file under "direct", and a serious undercount of the influence behind it.
- Zero-click answersThe normal outcome of being cited: the reader gets what they needed inside the answer and never visits you — which is worth something, and worth measuring differently.
- Language and country in AI answersAnswers change with the language they were asked in and the country they were asked from — so measuring in one and selling in another tells you about somebody else's market.
Crawlers, permissions and policy
Who may read your site, how to prove who they really are, and how to say so in a file.
- AI crawlers & robots.txtThe bots that AI companies use to read the web — and the robots.txt file where you decide which ones may read your site.
- llms.txtA simple file at your site root that gives AI tools a curated map of your most important pages.
- Content Signals (search, ai-input, ai-train)A line in robots.txt that separates three different permissions people used to grant by accident all at once: indexing, answering, and training.
- Verifying AI crawlersProving a visitor claiming to be GPTBot really is GPTBot — because the user-agent string is a claim anyone can copy, and blocking on it wrongly costs you the crawlers you wanted.
- Blocking crawlers (Cloudflare / WAF challenges)When your bot protection (like Cloudflare) shows automated visitors a challenge page, it can also block the search and AI crawlers you actually want.
What goes wrong
The failure modes: invented facts about your brand, tactics that backfire, and text written to manipulate the reader's assistant.
- Hallucinations about your brandWhen an assistant states something confidently wrong about you — a price you do not charge, a service you do not offer, a location you have never had.
- What does not work in AI visibilityThe tactics being sold for getting cited by AI that either do nothing or actively damage you — worth knowing before you buy them.
- Prompt injection (and why it is a web-publishing problem)Text on a web page written to give orders to the AI that reads it — an attack on assistants that browse, and a temptation for site owners that backfires.
The rest of the stack
E-mail authentication, DNS, security headers and page speed. None of it wins an AI answer on its own, and all of it can lose you one: a page that will not load, or a domain whose mail is spoofable, undermines the trust everything above depends on.
SEO
- Structured data (schema.org / JSON-LD)A machine-readable description of what your page is about, which helps search engines show rich results and helps AI understand your content.
- Canonical URLA tag that tells search engines which address is the ‘official’ one when the same page can be reached by several URLs.
- Meta descriptionThe short summary of a page that search engines often show under its title in results — your chance to earn the click.
Security
- HSTS (HTTP Strict Transport Security)A setting that tells browsers to only ever open your site over the secure HTTPS connection — never plain, interceptable HTTP.
- HTTP security headersSmall instructions your server sends with every page that tell the browser to switch on built-in protections against common attacks.
- SSL / TLS certificates & HTTPSThe technology behind the padlock — it encrypts the connection between your visitor and your site so nobody in between can read or alter it.
- SPF (Sender Policy Framework)A DNS record that lists which mail servers are allowed to send email using your domain, so receivers can spot forgeries.
- DKIM (DomainKeys Identified Mail)A cryptographic signature added to your outgoing email that proves the message really came from your domain and wasn't tampered with.
- DMARCA policy that sits on top of SPF and DKIM, telling receivers what to do with mail that fails authentication — and emailing you reports of who's sending as you.
Performance
- Core Web Vitals (LCP, INP, CLS)Google's three user-experience measurements — how fast the main content appears (LCP), how quickly the page reacts to taps and clicks (INP), and how much it jumps around while loading (CLS).
- TTFB, caching and compressionHow long the server takes to start answering (TTFB), and the three levers that make every page faster for everyone: caching, a CDN and compression.
Reading is one half. Check whether your own page can be quoted — it names the sentences, takes seconds, and calls no model. Then ask the engines what they say about you.