AI search optimization

Optimising for a ranked list and optimising to be quoted are different jobs that share most of their infrastructure. This guide is the second job: what to actually do, in what order, and how each step differs from the SEO equivalent you have probably already done.

It assumes the argument for any of it is settled. The GEO guide makes that argument and explains the mechanism. This one is the work.

How it differs from ranking

Four differences drive everything below.

The unit is a passage, not a page. Ranking is a judgement about a document. Being quoted is a judgement about a paragraph. A page can rank first on the strength of its whole and contain nothing an engine can lift.

Context is stripped. Your navigation, your layout, the three paragraphs above the one being read — none of it comes along. What survives is text that means something alone.

Identity is a separate problem. A ranked result carries your domain in the URL. An answer carries your name only if the engine is confident which company that name refers to.

Access is per-engine and absolute. There is no partial credit. A crawler that is disallowed reads nothing, and the site is not ranked lower in that engine — it is absent from it.

Week one: the things that gate everything else

Fix crawler access

Read your robots.txt agent by agent and separate three questions that look like one: which bots decide whether an assistant can quote you, which feed a search index, and which only collect training data. Blocking the third costs nothing measurable; blocking the first removes you from that engine entirely.

The specific line to check is Google-Extended, which is not a crawler at all — it is a switch governing whether Gemini and Google’s AI answers may use your content, and it is the one most often disallowed by accident during a general clean-up. The crawler checker reports every agent separately and names the line responsible for each block.

Confirm your text arrives in the HTML

Fetch one of your own pages without JavaScript and count the words. If the body is assembled in the browser, most answer-engine crawlers see an empty page, and nothing else in this guide can help.

This is the most consequential finding in the whole discipline and the least discussed, because it is invisible from a browser. Everything looks fine. It is fine, for people.

Week two: identity

Pick one name. The legal name or the trading name — one — and use it in every declared field: your schema.org Organization, your og:site_name, your title suffix. Three correct names written by three people at three times is the ordinary state of a company website and, to a machine matching strings, three companies.

Then add sameAs: LinkedIn, Crunchbase, GitHub, a Wikipedia entry if you have one. It is a five-minute change and it is the only item on this list that recruits sources you do not control — each link is a thread tying a mention of your name elsewhere back to you.

Name what you sell, too. A Product, SoftwareApplication or Service node on the page that sells it turns “what does Acme do” from an inference into a reading. The entity analyzer shows which of these a page currently declares and which it leaves to be guessed.

Month one: the pages themselves

Write passages that survive being lifted

Take your ten most important pages and read each paragraph as if it had arrived alone on a stranger’s screen. Most will fail, and they fail in a small number of ways:

• They open with a pronoun. “This means that…”, “It also…”, “They therefore…” — each refers to something the reader cannot see. Start with the subject.
• They are too long. Past roughly a hundred words a passage is truncated rather than quoted.
• They are abstract. A paragraph with no number, no name and no date is grammatical and quotes as nothing.

Then break the page up. Subheadings, and a question with its answer directly beneath it. Put anything enumerable in a list or a table rather than a sentence with commas — those are the formats that come out of a page whole. The citation checker scores one page on seven of these signals and shows you the passages that currently qualify.

Answer the question in the first paragraph

The single highest-leverage editorial change, and the one that runs against how most marketing pages are written. A page that spends four paragraphs establishing why the question matters before answering it has buried the quotable passage under the part nobody quotes.

State the answer, then justify it. The justification is what makes the page worth reading; the answer is what makes it worth citing.

Make claims specific

Replace the strongest vague claim on each page with the number behind it. “Significantly faster” carries nothing into an answer; “four minutes instead of nineteen” is a sentence an engine can use and a reader can check. Where there is no number behind a claim, you have found something more useful than a copy edit.

Month one, in parallel: markup

Start by checking what you already have parses. A malformed JSON-LD block is present in your source, absent from every consumer, and completely invisible in a browser — one trailing comma or one smart quote and the whole declaration is skipped in silence. That is work you have already paid for and are not receiving. The schema checker keeps the blocks that fail and reports the parser’s own error.

Then, in this order:

1. Organization with sameAs on the homepage, referenced as the publisher elsewhere.
2. Article on anything dated, with an author, datePublished and an honest dateModified. Both dates: without the second, a page you revised last week reads as untouched.
3. Product and Offer on anything sold. Asked what something costs, an engine otherwise has to decide which number on a pricing table is the price.
4. BreadcrumbList site-wide — usually a template change rather than a content one, and cheap.
5. FAQPage where you have real questions with real answers on the page. Its role in Google’s rich results has narrowed, but a question paired with its answer is still the cleanest shape an engine can lift.

Never mark up something the page does not show. Markup that does not match visible content is the one category of mistake here with a real downside, because it is the one an engine treats as deliberate.

Quarter one: off the site

Everything above is within your control and has a ceiling. Past it, the work is getting mentioned accurately in the places a model already reads — comparison pages, review sites, industry write-ups, forum threads where your category comes up.

This is ordinary marketing with a new reason attached. It is not about a link passing authority; it is about a claim being corroborated somewhere the model looks. Which means accuracy matters more than volume: one correct listing on a source an engine trusts does more than ten directory entries with your old positioning.

Two things worth doing deliberately. Correct outdated descriptions of you wherever you find them — they persist, and a model reading three sources will average them. And link out from your own pages to the sources behind your claims: a page that cites nothing asks to be taken on faith, which is the position an engine is least willing to grant a source it does not recognise.

What not to bother with

• Keyword density, in any form. Nothing about generative retrieval rewards it, and the pages that read as keyword-stuffed are the pages that get paraphrased.
• Publishing volume for its own sake. Thin pages produced at scale give an engine more surface and nothing more to quote.
• Chasing a specific engine’s quirk. These systems change without notice. Everything in this guide is a property of being legible, which is stable; a trick that works on one engine this quarter is not.
• llms.txt, first. Worth twenty minutes eventually. Worth nothing before the six items above it, and worth precisely nothing at all if your robots.txt blocks the engine that would read it.

Then measure it

All of the above is readiness. Whether it worked is a separate question, and answering it badly is easy — the measurement guide covers prompt sets, the grounded-versus-recalled distinction, and how to read a number of this kind without over-reading it.

The free GEO audit scores a site across all six pillars in one pass, if you would rather start from a list than from this one.