Reading time: 11 minutes
Google does not require structured data for AI answers. Most studies with open methods also found no citation gain from schema markup for AI. Still, the markup helps traditional SEO and brand identification. So this guide shows what each company says and what to do next.
In short
- Google says AI search does not require structured data. It also says there is no special schema for AI Overviews or AI Mode.
- Among the companies we checked, only Microsoft says in its own documents that schema markup helps its AI. We found no documentation on the topic from OpenAI, Anthropic or Perplexity.
- In an Ahrefs test of 1,885 pages, adding schema brought no major citation gain on Google or ChatGPT.
- Keep the markup consistent with the visible text. Still, content, bot access and brand presence outside your site matter more.
Does Google require structured data for AI Overviews and AI Mode?
Google does not require structured data for AI Overviews or AI Mode. According to the company’s AI guide, there is no special schema.org markup for AI to add. However, the guide recommends keeping the markup for SEO, because it makes the page eligible for rich results.
According to Google’s documentation on AI features, a page can become a supporting link in AI Overviews and AI Mode if it is indexed. It must also be eligible to appear in Search with a snippet. The same page treats structured data as a good practice, so it should match the visible text.
Google’s AI optimization guide lists “Overfocusing on structured data” among the things you do not need to do. So Google does not tell you to mark up anything for AI. On the other hand, it does not say it ignores the markup either.
There is also a general limit. According to the general guidelines, Google does not guarantee that structured data will show up in results, even with correct markup. These pages cover AI features in Search. We found no equivalent page for the Gemini app.
Do ChatGPT, Copilot, Claude and Perplexity read structured data?
Among the companies we checked, only Microsoft says in its own documents that schema markup helps its AI systems understand content. On the other hand, we found no documentation from OpenAI, Anthropic or Perplexity about schema on regular pages. In its shopping help page, OpenAI mentions structured metadata when it picks products in ChatGPT.
What Microsoft says about structured data
According to Search Engine Land, Fabrice Canel of Bing confirmed in March 2025 that schema helps Microsoft’s language models understand content. He spoke at the SMX event in Munich. However, the confirmation came later, in the comments of a LinkedIn post.
In October 2025, the Microsoft Advertising blog went further. It says: “Schema is a type of code that helps search engines and AI systems understand your content.” So schema markup is a practice for AI search. In the end, though, the post admits that nothing guarantees selection.
The Bing Webmaster Guidelines apply to Bing, Copilot and grounding results. One item on structured data says: “Structured data may support clearer grounding but does not guarantee visibility or grounding traffic.” Also, the markup must reflect the visible content.
What OpenAI, Anthropic and Perplexity publish
We found no mention of structured data on OpenAI’s bots page, which explains OAI-SearchBot and GPTBot. The same goes for Anthropic’s help center, which describes ClaudeBot, Claude-User and Claude-SearchBot. Likewise, Perplexity’s crawlers page, about PerplexityBot and Perplexity-User, does not mention the topic.
There is one exception, tied to shopping. According to the ChatGPT shopping help page, OpenAI considers structured metadata from first-party and third-party providers, such as price and product description. But the passage does not say where this metadata comes from. It does not mention schema.org either.
What did 2025 and 2026 studies find about structured data and AI citations?
Most studies with published methods found no citation gain from schema. Only one study found an advantage, and only for product schema with filled attributes. Three of the four studies are observational, so they did not randomize pages. Also, all of them come from companies that sell SEO or AI tools or services.
The Ahrefs test, with control pages
In May 2026, Ahrefs compared 1,885 pages that added JSON-LD between August 2025 and March 2026 with 4,000 control pages. It measured citations in AI Overviews, AI Mode and ChatGPT, comparing the 30 days before and after the change. The changes were -4.6%, +2.4% and +2.2%, respectively. Only the first differed from zero in a statistically significant way. Ahrefs says it cannot pin that drop on schema, because both groups were already falling.
Ahrefs notes some limits. The pages already had more than 100 AI Overview citations, all schema types were pooled and only the HTML was analyzed. In the same work, pages cited by AI were almost three times as likely to have JSON-LD, out of 6 million URLs. For Ahrefs, that is correlation, because well-maintained sites have schema and also get cited. Ahrefs sells SEO and AI monitoring tools.
The searchVIU and SE Ranking tests
In October 2025, searchVIU ran a different test. It built a fictional product page with prices in different formats. Then it asked ChatGPT, Claude, Gemini, Perplexity and Google AI Mode for the prices. None of the five found the price that existed only in JSON-LD. On the other hand, ChatGPT and Gemini found visible prices, including in microdata and RDFa. It was a single test of direct page reading, but the company notes that schema may be used during indexing. searchVIU sells SEO software.
In November 2025, SE Ranking analyzed 129,000 domains and 216,524 pages across 20 niches. Pages with FAQ schema averaged 3.6 ChatGPT citations, against 4.2 for pages without it. This compares averages across pages. It is not an experiment, so it does not prove cause. SE Ranking sells a ChatGPT monitoring tool.
The study that found an exception
In February 2026, Kurt Fischman of Machine-like published a preprint with 730 citations from ChatGPT (GPT-4o with browsing) and Gemini (1.5 Pro). It covered 75 commercial searches and 1,006 pages. The first model showed a negative link between schema and citation. The author blames a method artifact. After he corrected it, schema presence did not predict citation (odds ratio of 0.678; p = 0.296). Google organic position was the main predictor. Each lower position came with about 24% lower odds of citation.
The exception was Product or Review schema with filled fields, such as price, rating and specifications. These pages had a citation rate of 61.7%, against 41.6% for pages with generic schema, such as Article or Organization. The models tested are from earlier generations, and the author runs an AI agent company. So the result still needs replication.
Does the Semrush +21.60% measure structured data for AI?
No. In January 2026, Semrush listed “structured data elements” with a +21.60% association with AI citations. But Semrush itself says it looked only at visible text. It did not evaluate metadata, HTML structure or schema markup. So that number says nothing about schema markup for AI citations. Semrush also sells SEO tools.
Which types of structured data make the most sense for brands?
For most brands, three types cover the essentials: Organization on the home page, Product with Offer on product pages and Article on content pages. This structured data helps Google and other search engines understand the company and its offers. However, the evidence of an effect on AI citations is weak.
- Organization: Google recommends the markup on the home page or on a page that describes the company, such as About, as the guide to SEO and AI explains. Include name, logo, contacts and official profiles in sameAs. No study we found measured the effect of sameAs. In the preprint, Wikidata links appeared on fewer than 4% of pages with schema and could not be evaluated.
- Product and Offer: this type has the most support. OpenAI mentions product metadata in shopping, and the preprint saw an advantage with filled attributes.
- Article and BreadcrumbList: they serve traditional SEO, and plugins such as Yoast already generate both.
- LocalBusiness: Google’s AI guide says Business Profile and Merchant Center can help local businesses and products be visible in AI answers.
- FAQPage: the FAQ rich result left Google Search in May 2026, as the guide on how to create an FAQ explains. Also, the SE Ranking data show no gain in ChatGPT.
What should you do with structured data before investing in AI?
Keep the markup that already has a role in SEO and make it correct. Do not create schema markup “for AI”. First, check what matters more: the text in the initial HTML, bot access and the brand’s presence on other sources. A short checklist solves most cases.
- Generate the JSON-LD on the server: bots that read only HTML miss markup inserted by JavaScript, as the guide on JavaScript SEO and AI crawlers explains.
- Mark up only what the page shows: name, price, author and dates must match the text, so the markup does not contradict the page.
- Validate the structured data: use Google’s Rich Results Test and fix the errors, because a template error repeats on every page.
- Measure with a control: Ahrefs suggests 5 to 10 test pages and 5 to 10 control pages with similar citation levels, compared after 30 days.
- Fix the bigger causes first: blocks, weak content and missing outside sources matter more, according to the guide on why AI does not cite your brand.
Does structured data that differs from the visible text hurt?
Yes, in two ways. On Google, markup that describes something the page does not show breaks the guidelines and can trigger a manual action. In AI bots, on the other hand, the price that existed only in JSON-LD never reached the answer in the searchVIU test.
Google’s guidelines say not to put content in structured data that readers cannot see. If the JSON-LD describes a performer, for example, the HTML must describe the same performer. A manual action removes eligibility for rich results, but according to Google it does not affect search ranking. Bing says markup that is irrelevant, inaccurate or misleading may be ignored.
The practical risk shows up in price and stock. If the schema says one thing and the screen says another, the markup contradicts the page and may be ignored. We found no test showing which of the two values AI repeats. So generate the markup from the same data that feeds the page, and do not edit it by hand.
Frequently asked questions
Only if the service fixes the markup that already has a role in SEO, such as Organization and Product. According to Google, there is no special structured data type for AI, and the Ahrefs test found no citation gain. So ask the agency for a test with control pages before paying more.
In part. SEO plugins such as Yoast already generate basic schema, such as Article and BreadcrumbList. But check that the home page has a complete Organization and that product pages have Product and Offer. The guide on WordPress GEO plugins shows what each one does.
We found no proof of that. Google says Organization can help tell the company apart in results, but we found no study that measures this effect in AI answers. Meanwhile, keep the company’s name and details identical on the site and in official profiles.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- Google Search Central, Optimizing your website for generative AI features on Google Search (2026).
- Google Search Central, AI features and your website (2025).
- Microsoft Advertising, Optimizing Your Content for Inclusion in AI Search Answers (2025).
- Search Engine Land, Microsoft Bing/Copilot use schema for its LLMs (2025).
- Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved (2026).
- searchVIU, test of prices in JSON-LD, microdata and RDFa with five AI systems (2025).
- Machine-like, Does Schema Markup Predict AI Citation? (2026).
- SE Ranking, study on ChatGPT citation factors (2025).
