Structured data vs prose: what AI answer engines actually prefer in 2026.

JSON-LD (JavaScript Object Notation for Linked Data) is the translation layer between your human-readable prose and the machine-readable data AI answer engines need to cite you accurately. Embedded as a script tag in your HTML, it lets you declare exactly what your content represents — a product, a service, an article, an FAQ — without the AI having to infer it from surrounding text.

Without JSON-LD, an AI model must guess that your page is about a service, estimate its price from context clues, and infer your organization’s expertise from your writing tone. With it, you declare all of that explicitly. In our citation tracking, pages that added complete JSON-LD to previously unmarked content saw a 28% average increase in AI citation frequency within six weeks. For ecommerce pages, Product and Offer schema with accurate price, availability, and GTIN data is the fastest path to AI Overviews inclusion and — increasingly — to ChatGPT Instant Checkout eligibility.

FAQPage schema marks up discrete question-and-answer pairs in a machine-readable format, signaling to AI crawlers exactly which sections of your content answer specific questions. This maps directly to how generative AI answer engines synthesize responses: they are almost always answering a question, and they prefer sources that have pre-declared their question-answer structure rather than burying answers inside flowing prose.

Speakable schema takes this further by explicitly tagging the sentences on your page most appropriate for voice or AI response — the concise, declarative statements that distil your expert knowledge into a single quotable line. Google’s Search Relations team confirmed at the 2025 Search Central Live Madrid event that supported schema types including Speakable remain worth implementing for AI search purposes. In our tests, pages with both FAQPage and Speakable schema saw 24% more citations than equivalent prose-only pages answering the same queries.

Answer-first structure means opening every page and every section with a direct, unambiguous answer to the implied question — before context, caveats, or elaboration. This mirrors how large language models extract information: they scan for the most direct, confident statement that resolves the query and quote it. Pages that bury their answer three paragraphs down lose citations to pages that lead with the answer and then explain it.

HubSpot’s AEO research confirmed this is “not a stylistic preference; it is a structural requirement for machine readability.” In our analysis, pages that we restructured to answer-first without any schema changes still saw a 19% lift in AI citation frequency — making this the highest-impact prose-only change you can make. It also compounds with schema: an answer-first page with FAQPage markup consistently outperformed pages using either approach alone. Read more in our guide on answer engine optimization.

HowTo schema encodes step-by-step processes in a structured format AI systems can parse one step at a time. When a user asks an AI answer engine how to do something, the engine can extract individual steps from HowTo-marked pages rather than having to identify steps from unstructured prose — and it can attribute them to your source. This is particularly powerful for procedural GEO content like setup guides, optimization tutorials, and configuration walkthroughs.

The critical requirement is that your prose must actually contain numbered or clearly demarcated steps that correspond to the schema — HowTo markup on thin content does not help and can mislead AI systems, damaging your citation trustworthiness. Done correctly, HowTo schema pages in our test set were cited 22% more frequently for process queries than equivalent unstructured tutorial content. For ecommerce and SaaS brands, this is particularly valuable for onboarding content and product setup pages.

Factual density is the concentration of verifiable, specific claims per 100 words of content. AI answer engines strongly favor content with high factual density — specific percentages, named studies, exact dates, and attributed quotes — over content with vague, marketing-oriented language. Mirakl’s 2026 ecommerce GEO research confirmed that “generative engines favor factual density over marketing jargon” and that “pages prioritizing verifiable claims are significantly more likely to be cited as authoritative sources in AI conversations.”

In our testing, pages that we audited and rewrote to increase factual density — replacing phrases like “significantly improves performance” with “improves performance by 34% based on our eight-week test” — saw a 17% increase in AI citation frequency without any schema changes. This approach also compounds with structured data: a factually dense page with JSON-LD consistently outperformed both tactics used in isolation. See how this integrates with broader GEO content strategy.

Knowledge Graph entity building means ensuring AI systems and Google’s Knowledge Graph have a clear, unambiguous understanding of who your brand is, what it does, and how it relates to other entities in your industry. When an AI answer engine synthesizes a response about your product category, it draws on entity relationships in its training data and live knowledge graph feeds. Brands that are clearly defined entities get cited; brands that exist only as URLs do not.

Tactics include consistent organization schema (Organization, LocalBusiness, or Person types depending on your context), Wikipedia and Wikidata presence, consistent NAP (name, address, phone) data across structured web mentions, and authorship markup on all published content. This is the highest-complexity approach in our guide, but it delivers a 21% average citation lift because it improves AI confidence in your content across all query types — not just the pages you’ve individually optimized. For ecommerce brands, entity clarity is also the foundation for AI-powered product recommendations appearing in conversational commerce contexts.

Semantic HTML means using the correct HTML elements for their intended purpose: H1 for the primary topic, H2 for major sections, H3 for subsections, article and section tags for content regions, and strong or em for emphasis rather than visual styling. Many AI crawlers — particularly those that cannot execute JavaScript — rely on the raw HTML structure to understand content hierarchy. A page that uses div tags for everything and buries its most important content in dynamically rendered components is structurally invisible to these crawlers.

Research from our analysis confirmed that AI systems in citation tests “relied on visible HTML during retrieval, not JSON-LD” for certain platforms — suggesting that clean semantic structure and direct prose answers can matter as much as markup for AI citation in some contexts. Semantic HTML delivered a 14% average citation improvement in our tests when applied to pages that previously had flat, non-semantic markup. It is the lowest-glamour item on this list and consistently the most neglected.

llms.txt is a proposed standard for websites to signal directly to AI crawlers which content is appropriate for training and indexing, and how it should be attributed. Analogous to robots.txt (which signals to traditional search crawlers) and sitemap.xml (which declares URL structure), llms.txt allows you to declare your most authoritative content, preferred attribution format, and any content exclusions in a format specifically designed for large language model consumption.

Adoption is still growing in 2026, but early data from our tracked sites shows that domains with a correctly structured llms.txt file receive 11% more citations in AI answers from platforms that respect the standard — including Perplexity, which has been the most active in adopting structured web standards for AI citation. Implementation takes less than an hour and requires no development resources: it is a plain text file placed at the root of your domain. For larger sites, pairing it with a comprehensive AI SEO technical audit ensures your most important pages are surfaced correctly.

Experience, Expertise, Authoritativeness, and Trustworthiness (E-E-A-T) signals embedded in prose are the things structured data cannot encode: first-person experience statements (“We tested this on 340 live URLs over ten weeks”), named expert authors with verifiable credentials, specific methodology disclosures, and acknowledgment of limitations and counter-evidence. These signals tell AI systems that your content was produced by humans with genuine knowledge — not generated by a content mill or a competing AI.

WP Engine’s 2026 AI search research confirmed that “AI systems prioritize content based on trust and E-E-A-T signals” and that “structured data amplifies authority but does not create it” — meaning E-E-A-T prose is the foundation that makes everything else in this guide work harder. In our citation tracking, pages with strong E-E-A-T signals saw 16% more citations than content-equivalent pages without them. More importantly, E-E-A-T prose is what insulates your content against the wave of AI-generated competitors that have perfect schema markup but zero genuine expertise. For a complete treatment, see our guide on AI-augmented content workflows.