Denshin / Blog / AI

Generative engine optimization: how to get cited by AI search

AI search answers questions without sending a click. Generative engine optimization is not a replacement for SEO, it is the same fundamentals plus passages that survive being lifted out of context. What makes content citable, which structured data matters, and how to measure it honestly.

Denshin Team · Product & Engineering · 27 August 2026 · 8 min read

A growing share of the questions your customers used to type into a search box now get answered by a model that reads the web for them and writes a paragraph back. Nobody clicks a blue link in that flow. If your page is not the thing the model reaches for and quotes, you were not in the conversation at all. That is the problem generative engine optimization is trying to solve, and most of the advice being sold about it is either recycled SEO or outright nonsense. This post is the practical version: what actually changes, what does not, and how to tell whether any of it is working.

What is generative engine optimization?

Generative engine optimization (GEO), sometimes called answer engine optimization (AEO), is the practice of structuring and writing your content so that AI search systems can find it, understand it, and cite it when they synthesise an answer. It is not a separate discipline from SEO. It is the same crawlable, fast, accurate, well structured, genuinely useful page, written so that individual passages survive being lifted out of context and pasted into somebody else's answer.

That last clause is the whole shift. Classic SEO optimises a page to win a position in a list. GEO optimises a passage to be worth quoting. The page still has to rank or at least be retrievable, because most AI answer systems are pulling from a search index underneath, but the unit that wins is smaller than the page.

What actually changes when answers are synthesised

Three things change in a way that matters for how you write and build.

The click is optional. A user who gets a complete answer may never visit. Your page can influence a purchase decision and generate zero sessions in analytics. Any measurement model that only counts sessions will tell you your best content is dying.
Context is stripped. A model may pull two paragraphs from the middle of your page. If those paragraphs only make sense after the four above them, they are unusable. Self contained writing beats clever narrative build up.
Attribution is a judgement call made by someone else's system. You do not control whether you are named, linked, or silently paraphrased. You can only make yourself the most convenient, most verifiable source available on that question.

What does not change: the model is reading pages that a crawler fetched. Everything that made a page findable and trustworthy for a search engine still applies, and if anything applies harder, because a synthesiser has to pick between a handful of sources rather than list ten.

The fundamentals still decide most of it

Before anything AI specific, get the boring layer right. In our experience shipping client sites, this is where nearly all of the real losses are, not in some novel AI tactic.

Crawlable. A sane robots.txt, a real sitemap, canonical URLs, no accidental noindex on a template, internal links that reach every page you care about.
Fast. Slow pages get fetched less thoroughly and rank worse, and a crawler on a budget is not going to wait for your hero animation.
Accurate. Wrong facts get you cited once and then get you a correction from a competitor. Dates, prices, and version numbers on a page you have not touched in two years are a liability.
Well structured. One h1, descriptive h2 and h3 headings, lists that are actually lists, tables that are actually tables.
Genuinely useful. There is no formatting trick that makes a thin page worth quoting. The passage has to contain information the model cannot confidently produce on its own.

We wrote most of this layer up already as part of the pre-launch checklist we run before shipping. If a site fails that checklist, GEO work is premature.

What makes a passage citable

Assume a retrieval system will chunk your page into pieces of a few hundred words and score each piece against a question. Write so that any given piece stands up alone.

Answer in the first two sentences under the heading. Not a preamble about how important the topic is. State the answer, then explain, then qualify.
Make headings match real questions. "What is generative engine optimization?" retrieves better than "The new landscape". Use the words people actually use, once, honestly.
Keep paragraphs self contained. Repeat the subject instead of leaning on "it" and "this" across a section boundary. Slight redundancy reads fine to a human and is what makes a chunk survive extraction.
Be specific and dated. "As of August 2026" beats "recently". A named method, a concrete number you can support, or a stated constraint gives a synthesiser something it cannot invent.
Show your basis. Link out to primary sources: vendor documentation, specifications, official announcements. Sourced claims are safer to quote, and systems that try to verify before citing will find the corroboration.
Use lists and small tables for enumerable facts. Steps, criteria, and comparisons extract cleanly. Prose that hides five criteria inside one long sentence does not.

Structured data and consistent entity information

Structured data does not make a model like you. It removes ambiguity about what your page is and who published it, which is exactly the thing a retrieval and citation pipeline has to resolve. Add JSON-LD for the types that describe your content honestly:

Article or BlogPosting with headline, datePublished, dateModified, and a real author.
Organization on the site, with your legal name, logo, and sameAs links to the profiles you actually control.
BreadcrumbList so the hierarchy of the site is explicit.
FAQPage only where the page genuinely contains questions and answers a user asked for. Do not bolt a fake FAQ onto every page.

Then make the entity data consistent everywhere: same company name, same spelling, same address, same founding details on your site, your business listings, and your social profiles. Contradictory information across the web is the single most common reason an AI answer gets your company's basic details wrong.

Make sure the crawler can see the text

This is the failure we see most often on modern front ends, including ones built on the stack we use ourselves. If your content is fetched by JavaScript after load, some crawlers will render it, some will index a shell, and some will fetch the HTML once and give up. You do not get to know which.

The fix is not exotic: server render or pre-render the pages that carry your content. A marketing site or blog should ship real HTML with the text in it. Keep the client side app for the parts that need interactivity, like an admin panel behind a login, where crawlers were never going to reach anyway. We covered how we split those two concerns in S3 and CloudFront for an admin heavy SPA. A quick check: fetch your own URL with curl and read the output. If the article body is not in there, no amount of GEO advice helps you. The complementary machine readable layer, robots directives and AI crawler user agents included, is covered in llms.txt and AI crawlers.

How to measure GEO honestly

You cannot get impression level data from most AI answer surfaces, and pretending otherwise is how agencies build dashboards that mean nothing. Here is what is actually observable as of writing in August 2026, and what each signal is worth.

Signal	What it tells you	Limits
Referral traffic from AI assistant domains	Somebody was cited and a human clicked through	Undercounts badly. Most answers never produce a click
Manual spot checks: ask the assistants your key questions	Whether you appear at all on the questions you care about	Non deterministic, personalised, and a small sample
Brand mention tracking	Whether your name is entering the conversation	Mentions are not links and not always favourable
Server logs and crawler hits	Which AI fetchers are actually retrieving your pages	Fetch is not citation. User agents change
Branded search volume and direct traffic	Downstream effect of being named in answers	Correlational at best. Many other causes

Pick two or three, record a baseline before you change anything, and re-check on a fixed schedule rather than reacting to a single bad answer. Treat every one of these as directional. The honest framing for a stakeholder is: we can see whether we are present and trending, we cannot see a rank.

What to be sceptical of

Nobody can guarantee citation in an AI answer. There is no submission form, no ranking API, and no relationship you can buy that reliably puts you in a synthesised paragraph. If a vendor promises guaranteed placement in AI answers, they are selling either luck they will take credit for or tactics that will hurt you.

Specifically, do not go near the things that were bad practice before and are still bad practice now: cloaking different content to crawlers than to users, buying links, spinning out hundreds of thin generated pages to blanket a keyword set, or stuffing fake structured data. Search vendors have decades of experience detecting exactly this, and an AI answer layer sitting on top of a search index inherits those defences. The downside is a manual action on your domain, which is a far worse outcome than not being cited. The same scepticism you would apply to any studio pitching you a guarantee applies here.

What to do next

A realistic first pass, in order:

Curl your top ten pages and confirm the body text is in the raw HTML.
Rewrite the opening two sentences under each h2 on those pages so they answer the heading directly.
Fix the headings themselves to match questions a real person would ask.
Add Article, Organization, and BreadcrumbList JSON-LD, and validate it.
Reconcile your company details across your site and every listing you control.
Set a baseline for two measurement signals and diary the date.
Re-check in a quarter. Change one thing at a time so you can attribute the difference.

None of this is a trick. It is the same work that made content good for readers, applied with the knowledge that the next reader might be a machine summarising you for somebody else. If you want a second pair of eyes on whether your site is actually retrievable, or you are rebuilding and want the rendering and structured data right from the start, talk to us.

Tags: Generative Engine Optimization, AI Search, SEO, Structured Data, Content Strategy

All posts · Work with Denshin