Schema, AEO, GEO & on-page SEO in 2026: how AI reads and cites a page

First, the mental model: how an AI engine uses your page

Before tactics, get the pipeline straight, because every recommendation below maps to a stage of it. When a modern AI engine answers a question, your page passes through four gates. It must be crawled (the bot can fetch it), parsed (the content and structure are understood), retrieved (it is selected as relevant to a sub-question), and finally cited (the model quotes or attributes it in the answer). Fail any gate and the later ones never happen.

Traditional SEO mostly optimized the first two gates and a ranking. AI search adds a wrinkle: the engine rarely runs one query. It decomposes the question into several sub-questions — the fan-out, and retrieves sources for each. A page that answers the main question and a few of its branches is far more likely to be pulled in. One 2026 analysis found pages ranking for the main query plus at least one fan-out query were 161% more likely to be cited, accounting for more than half of all citations measured.

It helps to see the four disciplines as layers on this pipeline rather than competing choices. On-page SEO governs the crawl and a share of the parse. Schema deepens the parse by adding machine-readable meaning. AEO shapes content so the retrieve-and-extract step finds a clean answer. GEO works the trust and authority signals that decide the final citation. You do not pick one; you stack them, and a weakness in any layer caps everything above it. A brilliantly written, perfectly structured page that a crawler cannot reach earns nothing, and a fast, crawlable page with thin, hedged content gives the model no reason to quote it. The rest of this guide walks each layer in the order an engine encounters it, with the concrete things to ship at each stage.

On-page SEO in 2026: still the foundation

Nothing gets cited that cannot first be crawled and understood, so on-page SEO has not gone away — it has become table stakes. The fundamentals that matter most for AI accessibility:

• Crawler access. Confirm your robots.txt allows the AI agents you want, GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended, and publish an llms.txt summarizing your site for models.
• Rendering. Content hidden behind client-side JavaScript is a frequent, invisible cause of non-citation. Server-render or statically generate anything you want quoted.
• Heading hierarchy. A clean H1 > H2 > H3 outline tells a model what the page covers and how sections relate. Messy or skipped levels degrade parsing.
• Speed and Core Web Vitals. They influence crawl efficiency and remain a ranking input, which still feeds the sources AI draws from.
• Internal linking. Links between related pages build the topical cluster and entity relationships engines use to gauge authority.

Think of on-page SEO as making the page findable and legible. It is necessary and no longer sufficient, which is where the next three layers come in. The mistake teams make is treating it as finished work; in practice it is the layer that silently fails most often, because a single blocked agent or a JavaScript-only render can erase everything built on top of it.

Structured data: making your page legible to machines

Schema markup (structured data, usually JSON-LD) does not force a citation, but it gives engines explicit, machine-readable context: what this page is, what entities it describes, and how they relate. In 2026 the types that earn their place on a B2B site are concrete:

• Organization — identity, foundingDate, sameAs links, and knowsAbout to assert topical expertise. This anchors your brand as an entity.
• Article / TechArticle — for content pieces, with author, datePublished and about to connect the piece to entities.
• FAQPage — pairs a question with a self-contained answer, a format engines extract cleanly.
• BreadcrumbList — encodes site structure and the page's place in it.
• Product / Service / Offer — for commercial pages, including transparent pricing where you have it.

The point of schema is not the rich result it might trigger in classic search; it is disambiguation. When a model can tie your page to a known entity with clear attributes, it can attribute and reuse your content with more confidence. Entity density matters here too: pages with many recognized, well-linked entities are selected more often than thin ones.

Recent events made that distinction official. On May 7, 2026, Google deprecated FAQ rich results entirely, after its March 2026 core update had already narrowed rich-result display for FAQ, Review and How-To markup on non-primary content. Read carefully, neither change devalues structured data: FAQPage remains valid schema, AI crawlers and Bing still parse it, and the same March update saw sites with clean entity schema improve their citation rates in Google's AI answers. What ended is the era of schema as a SERP-display trick. What remains, and grew, is schema as entity verification — with Organization properties like knowsAbout, which declares the topics your company genuinely covers, emerging as one of the highest-impact additions for AI source selection.

Disambiguation is worth dwelling on because it is where many brands quietly lose. If a model cannot tell which company "AC" refers to, or which of three similarly-named products you are, it hedges — and a hedge usually means it cites someone clearer instead. Consistent naming, accurate sameAs links to authoritative profiles, and agreement between your schema and the wider web all reduce that ambiguity. The goal is to make you the obvious, low-risk entity to name.

A concrete example helps. An Organization block that asserts identity and expertise carries a name, a foundingDate, a set of sameAs links to your profiles, and a knowsAbout array naming the topics you have authority on. That last field does quiet work — it connects your brand entity to the topic entities a model reasons about when it decides who to cite for a given question. Pair it with an Article block whose author resolves to a real person or organization, and you have given the engine a clean chain from page to author to topic.

One caution: schema is a claim, not a guarantee. Engines cross-check structured data against the visible page and against what other sources say about you. Marking up a claim you do not support in the body, or that no third party corroborates, does not help and can hurt. Treat schema as a way to make true things legible, not a lever to assert things you have not earned.

Common mistakes that quietly block citation

Most pages that fail to get cited are not missing a clever tactic; they are tripping one of a few recurring problems. In audits these come up again and again:

• The answer is buried. A page that opens with three paragraphs of throat-clearing before the substance loses the front-loading advantage entirely, the model often never reaches the good part.
• Content depends on JavaScript. If the quotable text only appears after a client-side fetch, many crawlers see an empty shell. This is one of the most common invisible causes of non-citation.
• One keyword, one angle. A narrow page targeting a single phrase cannot satisfy the fan-out. Comprehensive pages that answer the question and its branches win because they are retrievable for several sub-queries at once.
• Hedged, promotional copy. Writing that is both vague and self-promoting fails twice: hedging reduces citation odds, and promotional tone correlates negatively with being quoted.
• No author, no entity. Anonymous content with no identifiable author and no entity markup gives the model nothing to trust. E-E-A-T is not a slogan here; it is a retrieval input.
• Stale with no signal. Engines favor fresh, clearly dated content for many queries. A page with no visible update date and aging facts gets passed over for a current one.

AEO: making your content extractable

Answer engine optimization is the discipline of structuring content so an engine can lift a clean, correct answer with minimal effort. It is the bridge between classic SEO and full GEO, and it is mostly about format and placement.

• Front-load the answer. Roughly 44% of ChatGPT citations come from the first third of a page. Put a direct, complete answer in the opening, then expand. Mirror the user's question in your H1 or first H2 and answer it in one or two sentences.
• Write self-contained blocks. Engines extract passages, not whole articles. Each section should make sense lifted out of context, without depending on the paragraph before it.
• Use definitive language. Cited text is nearly twice as likely to use clear, declarative phrasing rather than hedging. Make claims and support them; cut the qualifiers that add nothing.
• Format for extraction. FAQ blocks, short direct-answer paragraphs, and clean lists give the model obvious units to quote.

AEO is where SEO habits start bending toward how machines actually read. A page written purely to rank reads differently from one written to be quoted, and in 2026 you want the second. A useful test before you publish: take any section out of the page and read it cold. If it answers a real question on its own, an engine can lift it. If it only makes sense after the three paragraphs above it, it will not travel.

GEO: making your brand citable

Generative engine optimization targets the final gate: being named and recommended inside the generated answer. This is where the signals diverge most sharply from classic SEO, because the model weighs trust and authority, where the ranking algorithm weighs relevance.

• Brand mentions over backlinks. In 2026 research, brand mentions correlate with AI citation at roughly 0.664, versus about 0.218 for backlinks. Contextual mentions of your brand across the web, even unlinked, move citation more than link building does.
• Earned media. Models trust third-party sources over self-description. Coverage in publications, communities and comparison platforms the engines read is where citation authority accrues.
• Entity authority and E-E-A-T. The vast majority of AI citations come from sources with strong experience, expertise, authoritativeness and trust signals. Named authors with real track records, accurate entity data and consistent descriptions all feed this.
• Fan-out coverage. Cover the question and its branches comprehensively so you are retrievable for several sub-queries at once, beyond the head term alone.
• Avoid promotional tone. Overtly promotional copy correlates negatively with citation. Editorial, factual writing is both more trustworthy and more citable.

• Contextual brand mentions (even unlinked) 0.66
• Traditional backlinks 0.22

Correlation with citation (0 = none, 1 = perfect). Source: 2025–2026 studies correlating ranking and off-page signals with AI-Overview and chatbot citations.

The uncomfortable truth of GEO is that much of it happens off your own site. You can perfect your pages and still lose if the wider web has little to say about you. That is why citation engineering pairs owned content with earned mentions: the owned work makes you quotable, and the earned work makes you trusted. Skip the second half and you have an extractable page that the model still has no independent reason to believe.

The engines are not interchangeable

"AI search" is a convenient shorthand for systems that behave differently under the hood, and optimizing well means respecting those differences. They share the fundamentals above, but each sources answers in its own way.

• Google AI Overviews lean heavily on Google's existing index and ranking systems, so strong classic SEO carries more weight here than elsewhere. Fresh, well-structured pages that already rank have an edge, and the fan-out behaviour is pronounced.
• ChatGPT blends training knowledge with live search for current queries, and it drives the largest share of AI referral traffic. Brand presence in its training-era sources and in the pages its search tool retrieves both matter, which rewards durable, widely-referenced content.
• Perplexity is citation-native: it shows its sources prominently and favors pages that read as clean, authoritative references. Clear structure and verifiable claims tend to perform well.
• Claude and Gemini increasingly retrieve from the live web and reward the same extractable, entity-rich, well-attributed content, with their own ranking quirks for which sources they trust.

The implication is operational: do not optimize for "AI" in the abstract. Track each engine separately, because a page that is cited in Perplexity can be invisible in AI Overviews, and the fix for one is not always the fix for the other.

Why ranking and citation have split

The reason all four disciplines now matter, rather than SEO alone, is a measurable decoupling. In mid-2025, most AI Overview citations came from the top ten organic results. By early 2026 that share had fallen sharply, with some studies putting it as low as one in six. A page can hold position one and still be absent from the AI answer above it, because the citation engine weighs extractability, entity authority and trust differently than the ranking algorithm weighs relevance and links.

Practically, this means you can no longer treat "we rank well" as "we are visible." The two are now separate scoreboards, and you have to play both. The teams that win the citation scoreboard are usually the ones that already mastered the ranking one and then learned the new signals — not the ones starting from either half alone.

How to measure it

None of this is worth doing if you cannot tell whether it worked, and AI visibility is measurable once you instrument it. The core metrics, in plain terms:

• Citation rate. Across a fixed set of buyer-relevant prompts, how often does each engine name you? Track it weekly so you can see movement and attribute it to specific work.
• Share of answer. On those same prompts, how often do you appear versus each competitor? Citation is relative; this tells you whether you are gaining on the brands that currently own the category.
• Accuracy and sentiment. When you are cited, does the model describe you correctly, right category, current features, fair framing? A confident, wrong description reaches buyers as fact.
• Source diversity. How many distinct trusted sources reference you? Breadth is the leading indicator of durable citation, because it survives model refreshes that a single source would not.

Run these across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, because each retrieves and weighs sources differently. A page can be strong in one and absent in another, and you only see that by checking all five, and the fix for a gap in one engine is not always the fix for another.

Two practical cautions on measurement. Sample the same prompts on a fixed cadence rather than improvising new ones each time, or you will mistake your own wording changes for movement in the engines. And capture the full response, not just a yes or no on whether you appeared, because the wording of how you are described is often where the real problem hides and the easiest win lives.

Internal links and topical authority

One on-page factor gets overlooked because it feels like old SEO advice: internal linking. It still matters, and arguably matters more for citation than for ranking. When you connect related pages into a tight cluster, you do two things at once. You make the relationships between your entities explicit, which helps an engine understand what you are an authority on. And you give a model more than one retrievable surface for the same topic, so a question that fans out into sub-questions can land on several of your pages instead of one.

The practical pattern is a hub and spoke. A central page covers a topic broadly and links out to focused pages that each go deep on a sub-topic; those pages link back to the hub and across to each other where it genuinely helps the reader. Done well, this reads as a coherent body of work rather than a pile of disconnected posts, and that coherence is itself a trust signal. Thematic consistency across a domain is one of the things engines weigh: a site that covers one area thoroughly looks more authoritative on it than a site that publishes a little about everything.

Keep the anchor text descriptive and the links relevant. A link labeled with the actual topic of the destination page tells the model something; a link labeled "click here" tells it nothing. The aim is not to flood the page with links. It is to map the real relationships between your ideas so a machine can follow them.

A small habit pays off here: when you publish a new page, spend five minutes adding links to it from the older pages it genuinely relates to, not only links out from it. New content tends to arrive orphaned, and an orphaned page is hard for both crawlers and models to place in the context of what you already cover. Wiring it back into the cluster on day one is the cheapest authority you will ever build.

The 2026 ship list

If you do nothing else, do these, in this order — each gate before the next. The sequence is not arbitrary: there is no point engineering citations for a page a crawler cannot read, and no point polishing schema on content that has no direct answer to extract. Work bottom-up.

1. crawl Allow AI agents in robots.txt; publish llms.txt; server-render quotable content.
2. parse Clean H1>H2>H3 outline; valid JSON-LD (Organization, Article, FAQPage, BreadcrumbList); accurate entities.
3. extract Front-load a direct answer; self-contained sections; definitive language; FAQ blocks.
4. cite Build brand mentions and earned media; strengthen author E-E-A-T; cover the fan-out; keep tone editorial.
5. measure Track citation rate and share of answer across ChatGPT, Claude, Gemini, Perplexity and AI Overviews weekly.

FAQ