Schema markup for AI search: what actually matters
Schema markup helps AI search by disambiguating entities and feeding knowledge graphs, while the visible text on your rendered pages does most of the retrieval work. When an assistant like ChatGPT or Perplexity fetches a page to answer a question, the pipeline extracts the main content and splits it into passages, then ranks those passages against the query. Boilerplate and script tags are usually stripped during extraction, and JSON-LD blocks often go out with them. A competitor whose visible text states the answer plainly will beat you at that ranking stage even with no markup on the page, and immaculate JSON-LD will not close the gap. This is the most common misunderstanding in Generative Engine Optimization. Teams treat schema as a private channel to the models, a way to state facts the page never says out loud. No such channel exists. If a fact matters, it needs to appear in rendered, crawlable text, ideally in a heading or an early paragraph that survives content extraction. Structured data can confirm facts the visible copy already states, which is useful for knowledge graphs. A fact that exists only inside a JSON-LD block never reaches the passage ranker.
Where structured data genuinely helps
The strongest case for schema is entity disambiguation. Assistants regularly confuse companies with similarly named competitors or with unrelated brands in other industries. Our July 2026 study found that 27 of 30 brands were materially misdescribed by an AI assistant answering from memory, and 13 of 30 were still misdescribed when the assistant used live web search. Organization markup with a legalName, a logo, a short description, and sameAs links pointing to your Wikipedia article, Wikidata entry, LinkedIn page, and Crunchbase profile gives crawlers a machine-readable statement of which entity your domain belongs to. That mapping feeds Google's Knowledge Graph and Bing's equivalent, and those graphs are consulted when AI Overviews and Copilot ground their answers.
Product markup earns its keep for a similar reason. An Offer with a price, a currency, and an availability status gives shopping surfaces a canonical answer to questions your paragraphs might phrase loosely. Google's shopping graph and Bing both ingest it, and AI shopping answers on those surfaces resolve product facts from that structured feed rather than from prose. For SaaS companies, SoftwareApplication with an offers block does the same job. On classic Google surfaces, valid markup still produces rich results, and rich results still earn clicks while AI answers and traditional listings coexist.
FAQ structure helps too, with a caveat. Google restricted FAQ rich results to a narrow set of government and health sites in 2023, so FAQPage markup will rarely change how your listing looks. The underlying pattern still matters: a question phrased the way users ask it, followed by a direct two or three sentence answer, is close to the ideal retrieval passage. The markup makes each question and answer pairing explicit for any parser that reads it, and it costs almost nothing once the on-page structure exists.
Which types to implement first, and which to skip
A workable priority order for most companies: Organization sitewide, WebSite on the homepage, Product or SoftwareApplication on commercial pages, Article with author and datePublished on editorial content, and BreadcrumbList across sections. Use JSON-LD and keep every field consistent with the visible page. Validate with Google's Rich Results Test and the schema.org validator. One accurate, maintained Organization block is worth more than a sprawling implementation nobody audits.
Skip the long tail. Speakable has almost no consumer surface. HowTo lost its Google rich results in 2023. Marking up content that is not visible on the page violates Google's guidelines and does nothing for pipelines that read rendered text. Typing every div on the page adds maintenance cost with no consumer on the other end. A useful filter: if no engine has documented that it consumes a type, treat that type as optional.
llms.txt versus schema
llms.txt and schema solve different problems, and on current evidence schema is the safer investment. Schema has documented consumers: Google and Bing publish which types they read and what those types trigger. llms.txt, a proposed markdown file at the root of a domain listing the pages you most want AI systems to read, has no confirmed consumption by any major AI crawler as of mid 2026. It takes an hour to create and carries no downside, so publishing one is reasonable, but it is a speculative bet while Organization markup is a documented input. Put the serious hours into the markup and into rendered copy clear enough to survive passage extraction; the hour on llms.txt can come after.
Myths worth retiring
Four claims circulate in sales decks and deserve skepticism. First, that JSON-LD feeds model training directly: training pipelines work from extracted page text, and there is no evidence that any foundation model gives markup privileged weight. Second, that more types mean more AI visibility: invalid or exaggerated markup can earn a manual action from Google and buys nothing from language models. Third, that FAQ markup gets you quoted in AI answers: assistants quote the answer text itself, which works with or without markup wrapped around it. Fourth, that schema can rescue thin content: an engine that cannot find a clear description in your text will not reconstruct one from your JSON-LD.
Where schema fits in the workflow
Schema belongs at the end of the workflow, after the text it describes. Write the rendered copy first: a homepage opening that states plainly what the company does and who it is for, and product pages that answer the questions buyers actually type. Then add Organization markup with its sameAs links, plus the commercial types above, and validate everything. Finally, check what assistants actually say about your brand, because that is the metric this work exists to move. Discoverable runs that check as a free AI visibility snapshot with no login required, with Claude checks live today and other engines rolling out.
Keep exploring: Free AI visibility checkerWhat is GEO?The 9 best GEO tools (2026)AI Visibility Index: 40 brands measured