Generative engine optimisation: getting cited by AI answers

A growing share of queries never reach a website. The user asks an assistant, gets a synthesised answer, and follows a citation only if they want to verify something. Being ranked first is no longer the same outcome as being the source.

What these systems actually do

Whether it is Google AI Overviews, ChatGPT search, Perplexity or Copilot, the pipeline is broadly the same: retrieve candidate documents, extract relevant passages, synthesise an answer, and attribute the passages used. The unit of citation is the passage. A page that ranks well but buries its answer in the ninth paragraph is a poor extraction candidate even when it is the better document.

Let the crawlers in

The most consequential and most overlooked factor. If your robots.txt disallows GPTBot, ClaudeBot, PerplexityBot or Google-Extended, you are not eligible for citation in those products at all. Plenty of sites blocked them reflexively in 2023 and have not revisited the decision since.

There is a genuine trade-off — allowing crawling means allowing training use in some cases — but it should be a decision, not an inherited default.

Answer first, elaborate second

Every section should open with a direct, self-contained answer in roughly 40–60 words, then expand. Written this way, a passage lifted out of the page still makes sense — which is exactly what extraction does to it.

This is the same discipline that wins featured snippets, applied to every section rather than only the introduction.

Headings as questions

People ask assistants in natural language. Headings phrased the way the question is actually asked give the retrieval layer a much stronger match than a noun-phrase heading does. "How long does an AWS migration take?" outperforms "Migration timelines" consistently.

Cite your own sources

Content that attributes factual claims to named, linkable sources is measurably more likely to be cited itself. It is also simply better content: a claim with a source attached is one a reader can check.

Authorship is not decoration

Answer engines and search engines both weight who wrote something. An article under a named author with a profile page, stated credentials, a resolvable external identity such as LinkedIn, and Person structured data linking them to the organisation is a stronger candidate than the same text published anonymously.

This is E-E-A-T, and its influence has grown rather than diminished as generated content has become cheap.

Structured data still matters

Organization, Person, Service, Article and FAQPage schema, linked through a single JSON-LD graph with @id references, lets an engine resolve who you are as an entity rather than as a string. Entity resolution is what makes a brand mention count as evidence about you.

llms.txt

An emerging convention: a plain-text file at the site root that maps your most valuable content for AI consumption. Support is not universal yet, and it is not a ranking factor. It costs almost nothing to publish and gives a crawler a curated route into your best material.

Measuring it

Referral traffic from assistant domains is visible in analytics but understates the picture badly, because most AI answers produce no click at all. Track brand mentions inside AI responses directly — run your key queries against the major assistants on a schedule and record whether you appear. It is manual, and it is currently the only honest measurement available.

What has not changed

Crawlable, fast, well-structured pages with genuine expertise behind them. GEO is not a replacement for SEO; it is the same foundation with more attention paid to passage-level clarity and to who is credibly saying it.