Schema markup helps AI search only at the margins. It can make a page easier for a machine to interpret, but no assistant cites a page because it carries JSON-LD, and Google says plainly that its AI features need no special markup. What decides whether ChatGPT, Claude, Gemini or Perplexity names your business is whether crawlers can reach your pages, whether those pages are indexed where the assistant searches, and whether you have enough pages that answer your buyers' actual questions.
That does not mean you should delete your structured data. It means you should put it in the right place on your list. For most small and mid-sized sites, that place comes after crawler access, indexing and content.
What does structured data actually do for AI answers?
Structured data is a machine-readable label on a page. It says: this is an article, this is its author, this is the organisation behind it, these are questions and their answers. Search engines use it to understand entities and to build rich results.
For AI features, Google's own documentation is the clearest statement. Its page on AI features and your website lists structured data that matches the visible content of the page as one recommended practice among several. The others are crawlability, page experience, and current Business Profile and Merchant Center data. Its guide to optimizing for generative AI features goes further and warns against overfocusing on structured data. The guide says structured data is not required for generative AI search, and no special schema.org markup gets a page into AI features.
OpenAI's documentation points the same way from a different angle. When the model uses web search through the API, the answer carries inline citations as url_citation annotations. Each one holds the URL, title and location of a cited source. The citation points to a page the search found. Nothing in that mechanism refers to markup.
Controlled testing has not found a shortcut either. CMSWire reports on an Ahrefs study of 1,885 pages that added JSON-LD schema. The experiment found no measurable lift in AI citations in AI Mode or ChatGPT, and a slight decline in Google AI Overviews.
In short, markup helps a machine read what is already on the page. It does not put a page into an answer.
What did we find when we compared markup with surface area?
We measure assistant answers every week, and we have gone through server logs from several sites we work with. One comparison stood out.
- The small site: complete markup, meaning server rendering, Article, FAQPage and Organization JSON-LD, and a sitemap, but only 75 URLs. It was still absent from answers on the exact topic of its own article.
- The large site: about 190,000 crawlable URLs. It was crawled constantly.
The crawl volume on a large site is hard to picture until you see it. We went through the full retained nginx logs of an edtech platform we work with, pickaclass.com, from 28 July to 11 August 2026:
- ClaudeBot made 19,272,125 requests
- Meta-ExternalAgent made 3,854,076
- GPTBot made 2,744,985
- Applebot made 2,655,555
- Amazonbot made 1,170,268
- Googlebot made 52,292
In the same fortnight, the platform's content pages received 1,918 fetches from live retrieval agents (ChatGPT-User, OAI-SearchBot, Perplexity-User). Each fetch is an assistant reading a page while it answers a real person. A client blog of a few dozen articles received 56 such fetches, spread across 24 articles.
We cannot pin the gap on a single factor. What the data does show is that the site with far more pages was crawled and fetched far more often, while complete markup did not make the small site appear in answers on its own topic.
Indexing is the other half. On 19 October 2026 we asked assistants for the exact titles of 18 articles on three client sites, each article older than three weeks. ChatGPT found 1 of 18 and Claude found 7 of 18. Google had indexed 29 of 30 articles on one of those sites. Being in Google's index did not mean an assistant could find the page.
Why doesn't markup decide who gets cited?
An assistant gets to a citation in four steps, and markup can matter in only one of them.
First, the assistant decides whether to search at all. On 23 September 2026 we tested 20 German business questions. The model answered 10 of them without searching and cited nobody. Markup cannot matter in an answer that never looks at the web.
Second, if it searches, it pulls candidates from a search index. If your page is not in that index, or a crawler was blocked, your page is not a candidate. For ChatGPT, that is why Bing indexing and IndexNow deserve more attention than most site owners give them.
Third, the model picks the sources whose text answers the question. In the September test, the model searched on the other 10 questions. On 9 of those 10 it cited primary sources: laws, ministries, the OECD and large accounting firms. It named firms only on 'who should I hire' or 'how to choose' questions. Wikipedia's entry on AI Overviews cites a June 2025 study that found Quora was the single most-cited source in Google's AI Overviews. Whatever drives that, it is a useful reminder that source selection in AI answers is not simply a matter of who has the most complete markup.
Fourth, answers vary. Two runs of the same question in our September test named two different sets of firms. A business can be named on one ask and missing on the next, so a single screenshot proves very little either way.
Which schema types are still worth having?
The common types are worth keeping, as long as they describe only what is visible on the page. Here is what each one does and does not do.
| Type | What it tells a machine | Why keep it | What it will not do |
|---|---|---|---|
| Organization | Your name, logo, URL and profiles | Ties your pages to one entity and separates your brand from similar names | Make an assistant recommend you |
| Article | Headline, author, published and modified dates | Makes authorship and dates explicit, so readers and machines can see how current a page is | Get a page into an index it was never crawled into |
| FAQPage | Question-and-answer pairs shown on the page | Mirrors the question-and-answer format assistants answer in | Replace the visible answer, which Google wants the markup to match |
| LocalBusiness | Address, opening hours, service area | Keeps your details consistent with your Business Profile | Beat directories and competitors whose pages already answer customers' questions |
| Product or SoftwareApplication | Name, price, platform, ratings | Gives clean facts for comparison questions | Create comparisons your pages do not cover |
The LocalBusiness row comes from something we measured. On 20 October 2026 we put 20 patient questions about an orthopaedic practice in Bonn to the four assistants. They named the practice in 12 of 80 answers: Claude 5 of 20, Gemini 4, ChatGPT 2 and Perplexity 1.
- The directory jameda.de was cited in 12 answers.
- Competing practices' own websites were the most cited sources.
- 'Who to choose' questions named practices.
- Questions about costs, risks and treatments named none.
The sources that won were pages with substance about the questions patients asked.
Two rules keep markup useful:
- Mark up only what a visitor can see. Google's guidance asks for exactly that.
- Do not paste FAQPage onto pages that have no visible questions. Mismatched markup works against you.
What should you fix before touching markup?
Work through this order. Each step is a precondition for the next.
- Crawler access. Our own mistake shows how easily this goes wrong. On our own site, we wrote a robots.txt rule to hide a single redirect page, /b. It was written as a prefix, and robots.txt rules match prefixes. So from 11 August to 21 October 2026, the rule also closed /blog, /business and every article in our business blog to all crawlers.
Bing Webmaster Tools had listed our business page as 'discovered but not crawled' since 11 August 2026. Google had none of the business pages indexed. We changed the rule to /b$ on 21 October 2026. If you are unsure about your own file, see how to allow GPTBot and other AI crawlers in robots.txt.
-
Rendering. If your text only appears after JavaScript runs, some crawlers may see an empty page. Check whether ChatGPT reads JavaScript websites before you invest in anything else.
-
Indexing in Google and Bing. Submit sitemaps, confirm your pages are actually indexed, and use IndexNow where it is supported.
-
Pages that answer buyers' questions. This is the surface that the comparison above points to. Expect few clicks. Across about two weeks of logs from three of our course websites, roughly 1.2 million requests in the two busiest logs, Copilot sent 1 human visitor. ChatGPT, Perplexity, Gemini, Claude, Meta AI and the others sent 0. Assistants mostly answer in place, so the goal is being named in the answer, not being clicked.
-
Markup. Add it once the steps above are done, and keep it in line with what the page visibly says.
How do you know whether any of this worked?
Measure the answers, not the markup validator. Here is the method we use:
- Pick the questions your buyers ask.
- Put each question to each assistant three times.
- Count a business as named for a question if any of the three asks names it.
- Compare weeks only on the same questions with the same number of asks.
This is cheap. With API prices of October 2026, a report of 20 questions, 4 assistants and 3 asks costs about $5.40.
Our own baseline shows why measuring matters. On 21 October 2026 we put 20 English and 20 German questions about getting a business into AI answers to ChatGPT, Claude, Gemini and Perplexity, three times each. ItGrows was named in 0 of 160 answers, and none of our pages was cited. That was the same day we fixed the robots.txt rule. You can read more in our own visibility measurement.
ItGrows runs this measurement weekly for clients, alongside publishing articles on their own domains and fixing crawler access and indexing. Nobody can guarantee that an assistant will name a business. To see where you stand today, the free check shows in about 30 seconds whether ChatGPT and Perplexity name you for five of your buyers' questions.
Frequently Asked Questions
Does ChatGPT read JSON-LD schema markup?
OpenAI's web search documentation describes citations as the URL, title and location of pages found through search, and it says nothing about schema. A controlled study reported by CMSWire found no measurable citation lift in ChatGPT after pages added JSON-LD. Treat markup as having no direct effect on whether ChatGPT cites you.
Is FAQPage schema worth adding for AI search?
It is worth adding to pages that already show questions and answers to visitors, because it describes that content accurately. It will not make an assistant cite the page. The visible question-and-answer text does the real work.
Can schema markup hurt AI visibility?
The Ahrefs study reported by CMSWire found a slight decline in Google AI Overviews citations after pages added schema, though that does not prove the markup caused it. Markup that does not match visible content goes against Google's guidance, so keep it accurate and do not overdo it.
What should a small business do first if it has no schema at all?
First confirm that AI crawlers are not blocked, that your pages render without JavaScript, and that they are indexed in Google and Bing. Then publish pages that answer the questions your buyers ask. Add Organization and Article markup afterwards, keeping it consistent with what each page shows.