By being the clearest source available when an answer engine assembles its reply: pages that answer first, structure that survives extraction, structured data that states facts plainly, an llms.txt index at the root, one consistent identity everywhere, and citations the estate earned rather than placed. This guide covers how answer engines read, what moves visibility, what does not, how to measure it, and where regulated estates need extra care. The channels are young. The principles are stable.
Every answer engine, whatever the brand on the front, runs a version of the same pipeline. Understand the three stages and most AEO decisions make themselves.
Bots fetch your pages, on a budget. Many read the HTML as served and never execute your JavaScript. What they cannot fetch cheaply, they skip, which is why server-rendered text, shallow stable URLs and a curated llms.txt index matter more than crawl-budget folklore ever did.
From each page the engine lifts candidate passages: headings, opening paragraphs, FAQ pairs, structured data. Extraction is brutal about context. A paragraph that only makes sense after the three above it does not survive; a paragraph that answers a question on its own does.
The engine decides which entities said what, and which source to cite. This is where consistency pays: a firm with one name, one description and one canonical URL per claim is easy to attribute. A firm with four variants of its own name splits its identity four ways.
Fig 1 · Crawl, extract, attribute Every answer engine runs a version of this pipeline over your pages and your llms.txt. The citation in the AI answer carries the one lime mark: it is the output every decision in this guide is aimed at.
Across the estates we run, these are the practices that change whether an estate appears in AI answers. None of them is a trick. All of them are also good for human readers, which is not a coincidence.
One question per page, asked in the h1 and answered fully in the first paragraph. If the opening survives extraction on its own, the page is quotable. This is the core of the Scrape-Bias Page Standard.
Real questions with plain-text answers, in the markup, not behind an accordion. FAQ pairs map almost one-to-one onto the prompts engines receive, which makes them the most directly reusable content on a page.
Valid JSON-LD on every page: Service or product schema for what you sell, FAQPage for what you answer, Organization for who you are. Schema states facts the engine would otherwise have to infer, and inference is where errors enter.
A plain-text index at the domain root: one line on what the organisation is, then the canonical pages. It gives a crawler on a token budget the curated version first. Ours is live at /llms.txt and took an afternoon.
Same firm name, same service names, same one-line descriptions: on the page, in the schema, in llms.txt, in directories and partner listings. Machines resolve identity by repetition. Every variant is a leak.
Coverage in trade press, analyst notes and standards bodies that engines already trust. This is the slow one, and the strong one: an estate cited by sources the engine weights is quoted with confidence. You earn it by publishing things worth citing.
Everything that gamed ranked search transfers badly to answer engines, because the unit of competition changed. You are no longer fighting for position nine instead of position eleven; you are either the material an answer is built from or you are absent.
Language models read for meaning, not term frequency. A page padded with query variants reads as low-quality prose and extracts as nothing in particular.
Content served to crawlers but not to people, white-on-white text, stuffed alt attributes. Engines compare what users see with what bots see, and the penalty for divergence is trust, the one currency you cannot buy back.
Fake reviews, self-placed citations, networks of sites quoting each other. Provenance analysis is exactly the kind of pattern matching these systems are good at. Manufactured authority reads as manufactured.
The engines change monthly and the tactic-of-the-week usually dies with the next model release. We are honest about the uncertainty: nobody fully knows how these systems weight sources today, let alone next year. The principles above are the part that has held.
Rank-tracking does not translate. Three measures do, and together they tell you whether the work is working.
Take the queries that matter commercially, the ones your buyers actually ask, and put them to the major assistants on a regular cycle. Record whether you appear, how you are described, and who appears instead. This is the headline number.
Traffic from assistant surfaces is small but unusually warm: the visitor arrives mid-conversation, already briefed. Track it separately from search, watch what it converts to, and expect the volume to understate the influence.
When engines cite sources, log which of your pages get cited and for what claims. Citations tell you which pages survive extraction, which is the fastest feedback loop AEO has. Pages that never get cited need rework against the standard.
For an FCA-regulated firm, a machine quoting your website is a distribution channel you do not control quoting promotional material you do. That changes three things. Precision, for once, is an advantage.
Every claim a machine might lift needs a clear, dated, attributable home on your own estate. If the canonical version of a claim lives in a PDF, a slide deck or a partner's site, the engine will find a worse version first.
Write every public sentence as if it will be quoted without its caveats, because it will be. Numbers you cannot evidence, superlatives you cannot defend and promises you cannot keep do not belong on extractable pages.
Pages built for extraction should pass through the same review as any financial promotion, with one extra test: read each paragraph in isolation and ask whether it is still fair, clear and not misleading when quoted alone.
Decide deliberately which AI crawlers may read what, per bot, in robots.txt, and record why. Blanket blocking removes you from answers your customers already read; blanket allowing should be a choice, not a default nobody reviewed.
AEO is not a separate project with its own agency and its own dashboard. It is a property of an estate that is built well. Here is where the work joins the rest.
Eight adoptable rules for building a page so machines quote it correctly. The standard behind this site.
A fixed-price four-week diagnostic. Machine visibility is scored inside its AI-readiness dimension.
Agentic migration plus the AI capability built into the live experience your customers touch.
Any system that replies with a synthesised answer rather than a list of links: ChatGPT, Perplexity, Google's AI Overviews, Copilot, and the assistants built on the same models. They crawl, extract and attribute, and the pages that survive that pipeline are the ones that get quoted.
Foundations move fast: answer-first rewrites, schema and llms.txt can ship in weeks, and engines that crawl frequently pick them up quickly. Earned citations and entity authority build over months. No honest practitioner will promise a date by which a specific assistant quotes you.
You can, and some publishers should. For a B2B estate the trade is usually poor: blocking removes you from answers your buyers already read, and the answer still gets written, sourced from whoever did not block. Decide deliberately, per crawler, in robots.txt.
Less than ranked links did, and that is the point: more of the journey happens inside the answer. What arrives instead is fewer, warmer visits from buyers who already know who you are, plus influence that never registers as a referral at all. Measure presence and assisted outcomes, not just clicks.
The same techniques apply, with a compliance layer on top. Every claim a machine might quote needs to be one you can stand behind in front of a regulator, on a page with clear provenance, reviewed the way you review any financial promotion. Precision, for once, is an advantage.
The DXP Value Survey includes an AI-readiness dimension: whether your estate is structured for machine readers and present in the answers your buyers already read. Four weeks, fixed price, benchmarked against UK FS&I estates.