Maneesh Jayachandran SEO | AEO | AI | Digital Growth

SEO & AEO

Schema, llms.txt and Chunking for AI Search: What Each Platform Has Said

By Maneesh Jayachandran

Last verified: October 10, 2026. Covers Google Search (AI Overviews and AI Mode), Bing and Microsoft Copilot, ChatGPT search, Perplexity and Claude.

For AI Overviews and AI Mode in Google Search, Google’s own documentation says you do not need special schema markup, an llms.txt file or content broken into small pieces. The normal requirements for appearing in Google Search apply, and structured data keeps its value for rich results.

Other platforms differ, so the answer depends on which one you mean. Bing is the only platform in this review that officially encourages formatting for AI: it says clear headings, tables and FAQ sections make content easier for its AI systems to reference, and it describes structured content as helping those systems interpret pages. OpenAI, Perplexity and Anthropic have published guidance about crawler access, but no public statement was found from any of them about schema or llms.txt.

On citations, the only controlled study of schema and AI citations found for this review, published by Ahrefs in May 2026, found no measurable increase in Google AI Overviews, AI Mode or ChatGPT citations for pages that were already heavily cited.

The evidence at a glance

Positions differ by platform, so the table gives each platform its own column rather than one verdict for “AI search”.

Tactic Google Search (AI Overviews, AI Mode) Bing / Copilot ChatGPT search Perplexity Claude (Anthropic)
Structured data / schema Not needed (official) Helps interpretation (platform’s own claim) No public statement found No public statement found No public statement found
llms.txt Not used by Search (official) No public statement found No public statement found No public statement found No public statement found
Chunking / AI-specific formatting Not needed (official) Recommended for structure (official) No public statement found Limited / indirect No public statement found

Three notes on reading this table:

  • Google Search is not the same as every Google product. Google Search says it does not use llms.txt. Lighthouse, Google’s separate developer auditing tool, checks the file in a category it marks as still under development. “Google Search vs Lighthouse” below explains the difference.
  • “No public statement found” means exactly that. No official guidance was found as of the date above. It is not evidence that a platform ignores schema, llms.txt or formatting.
  • “Limited / indirect” for Perplexity refers to an engineering post describing how Perplexity’s own system parses list- and table-heavy pages. It is not guidance for publishers.

How the evidence is labeled

Each claim in the tables below has one of four statuses:

  • Supported: a primary source, such as the platform’s own documentation, clearly supports the claim.
  • Contradicted: a primary source says the opposite, or materially conflicts with the claim.
  • Unproven: the claim is common or plausible, but no adequate primary evidence supports it.
  • Open: the evidence is incomplete, conflicting or too platform-specific for a firm verdict.

“Supported” applies only to the statement in that row. A row such as “Bing says structured content helps its AI systems” can be Supported because Bing says so; that does not mean the tactic has been shown to increase rankings or citations.

Sources are labeled by type, from strongest to weakest for this question: official documentation (including official source code), official blog posts, specifications and proposals, original studies with a disclosed method, and practitioner or vendor claims. Practitioner and vendor claims appear only as claims being evaluated, never as proof.

The same claim can have different statuses on different platforms. “Special schema is required” is contradicted for Google Search and simply unproven elsewhere.

No platform in this review says special or AI-specific schema is required to appear in its AI answers. Google says structured data is not required for its generative AI features but remains useful for rich results. Bing describes structured content as helping its AI systems interpret pages. No position was found from the other platforms.

“Schema markup” and “structured data” are used interchangeably here: code, usually based on the Schema.org vocabulary, that describes a page’s content in a machine-readable form. Advice about structured data often blurs three different roles, and the evidence differs for each:

  1. Rich-result eligibility. Structured data can make a page eligible for enhanced displays in classic search results, such as product or recipe details. Google documents this extensively.
  2. Helping a system understand a page. Google’s structured data documentation describes markup as helping it understand page content and entities. Bing makes a similar claim about its AI systems.
  3. Earning AI citations. This is the role behind most advice to add schema for AI search. Apart from a general Bing blog statement that structured content increases the chances of being cited, no platform documentation was found making this claim, and the controlled study covered below did not find the effect.
Claim Platform / product What the platform has said Status Basis
Special or AI-specific schema is required Google Search (AI Overviews, AI Mode) No special schema.org markup is needed Contradicted Official documentation · May 2026
Structured data is required for AI features Google Search (AI Overviews, AI Mode) Not required for its generative AI features Contradicted Official documentation · May 2026
Structured data still matters for rich results Google Search Recommended for rich-result eligibility, which it enables but does not guarantee Supported Official documentation
Structured content helps AI systems interpret pages Bing / Copilot Structured content helps its AI systems interpret and summarize content Supported (Bing’s own claim) Official blog · Nov 2025
Structured data guarantees grounding or visibility Bing / Copilot May support clearer grounding but does not guarantee visibility or grounding traffic Contradicted Official documentation · checked Oct 2026
Adding schema increases AI citations Google AI Overviews, AI Mode, ChatGPT None of these platforms says so; one controlled study found no measurable lift Unproven Original study · May 2026
Schema is used or rewarded by ChatGPT search ChatGPT search No public statement found Unproven Official documentation checked; nothing on this point
Schema is used or rewarded by Perplexity or Claude Perplexity; Claude No public statement found Unproven Official documentation checked; nothing on this point

What Google says

Google published a guide to optimizing for generative AI features in Google Search on May 15, 2026. It states that there is no special schema.org structured data you need to add for AI Overviews or AI Mode, and that structured data is not required for its generative AI features. It still recommends structured data as part of SEO, because it supports eligibility for rich results.

That is consistent with Google’s longer-standing structured data documentation. Google’s introduction to structured data describes markup as helping Google understand page content, and says not to mark up information that is not visible to users. Its general structured data guidelines say markup enables a search feature but does not guarantee it will appear.

Taken together, Google gives structured data a documented role in rich results and in understanding pages, and no documented role as an AI-feature requirement.

What Bing says

Bing says more about structure and AI than any other platform in this review. In its February 10, 2026 announcement of AI Performance reporting in Bing Webmaster Tools, Bing advises that clear headings, tables and FAQ sections help surface key information and make content easier for AI systems to reference accurately.

A November 2025 Bing Webmaster Blog post goes further. It says structured content, including schema-marked product pages, FAQs and comparison tables, helps AI systems interpret and summarize content, and presents this as improving the chances of being cited.

Bing’s Webmaster Guidelines, which cover how content is surfaced in Bing and Copilot, are more cautious. As of October 2026, they say structured data may support clearer grounding (basing an AI answer on retrieved web content) but does not guarantee visibility or grounding traffic, and that markup must match visible content. They also advise a logical H1 to H6 heading hierarchy, stating facts and definitions explicitly, and keeping each page accurate and focused on its topic.

Two limits apply to all of this. These are Bing’s own descriptions of its systems, not independent measurements. And they apply to Bing and Copilot only: they do not establish how any other AI assistant treats markup, including assistants that use Bing infrastructure.

What OpenAI, Perplexity and Anthropic have said

None of the three has published guidance, as of this review, telling site owners to use schema markup to appear in its AI answers.

  • OpenAI. Its publishers and developers FAQ focuses on allowing OAI-SearchBot, its search crawler, if you want to be eligible for ChatGPT search. For shopping, OpenAI documents its own product feed specification rather than schema.org markup on product pages.
  • Perplexity. Its crawler documentation covers PerplexityBot access. A September 2025 engineering post on its search API mentions that sites with a lot of list or table content suit more formulaic extraction rules. That describes Perplexity’s own parsing system; it is not guidance about schema markup.
  • Anthropic. Its crawler documentation explains its bots and how to control them. No publisher guidance on structured data was found.

This silence does not show that these platforms ignore markup. It means the question is open, and claims about how they treat schema are not backed by the platforms themselves.

A controlled study of schema and AI citations

Ahrefs published a controlled study of schema and AI citations on May 11, 2026. It tracked 1,885 pages that added JSON-LD structured data (markup added to a page as a block of script code) between August 2025 and March 2026. It compared them with 4,000 matched control pages and measured citations in Google AI Overviews, AI Mode and ChatGPT over the 30 days before and after the change. Using a difference-in-differences comparison, which measures the change in the treated pages against the change in the controls, it found:

  • AI Overviews: a 4.6% decline, which was statistically significant but which Ahrefs did not confidently attribute to schema
  • AI Mode: a 2.4% increase, not distinguishable from zero
  • ChatGPT: a 2.2% increase, not distinguishable from zero

The study did not establish a positive citation effect. Its limits matter for how far that result travels:

  • Every page in the sample was already heavily cited, with more than 100 AI Overview citations before the change. It says nothing about pages AI systems do not yet surface.
  • The observation windows were 30 days.
  • It measured JSON-LD in the page HTML only.
  • It is an industry study by a company that sells SEO software, not platform documentation.

Figures showing that a high share of AI-cited pages carry structured data are a different kind of evidence. They are correlations: they show that cited pages often have markup, not that the markup caused the citations.

FAQ and HowTo rich results: what changed

Google has retired both of these rich-result types:

  • August 8, 2023: Google limited FAQ rich results to well-known, authoritative government and health websites, and limited HowTo rich results to desktop.
  • September 13, 2023: HowTo rich results stopped appearing on desktop as well, ending the feature.
  • May 7, 2026: FAQ rich results stopped appearing in Google Search. Google’s documentation changelog records the deprecation notice and the removal of the FAQ documentation in June 2026.

FAQPage and HowTo remain types in the Schema.org vocabulary; what changed is Google’s rich-result display. The change is not evidence about how other AI systems treat FAQ markup, and it says nothing against visible question-and-answer content written for readers. Advice to add FAQ schema to win Google rich results is out of date. Claims that FAQ markup drives AI citations remain unproven.

Do you need an llms.txt file?

Not for Google Search: Google says Search does not use llms.txt, and that having one neither helps nor harms Search visibility. For Bing and Copilot, ChatGPT search, Perplexity and Claude, no public statement was found saying their AI systems read publishers’ llms.txt files. llms.txt is a proposed Markdown file that gives AI agents a summary of a site and links to its key pages; it is not a ratified standard.

Claim Platform / product What the platform has said Status Basis
llms.txt is a web standard Not platform-specific The proposal describes itself as a proposal to standardize Contradicted Proposal · 2024, revised 2026
Google uses llms.txt for AI Overviews or AI Mode Google Search Search does not use llms.txt; publishing one neither helps nor harms Search visibility Contradicted Official documentation · June 2026
A Google product checks llms.txt Lighthouse (not Google Search) Audits the file’s format, in a category marked as under development Supported Official source code · Lighthouse 13.5.0, Sept 2026
AI companies read publishers’ llms.txt files ChatGPT search; Perplexity; Claude; Bing / Copilot No public statement found Unproven Official documentation checked; nothing on this point
Wide adoption shows AI systems use it Not platform-specific Publication is not consumption; log studies observed little fetching by AI search crawlers Unproven Proposal; vendor log studies · 2026

What llms.txt is, and what it isn’t

llms.txt was proposed by Jeremy Howard of Answer.AI on September 3, 2024, and revised in August 2026. The proposal describes itself as “a proposal to standardise on using an /llms.txt file to provide information to help agents use a website.” The file is a Markdown document at a site’s root, or at a path within it, with a short summary and links to the pages an AI agent is most likely to need.

Its stated purpose is narrower than “AI search visibility”. The proposal says llms.txt information is “used on demand, when an agent needs information about a topic while assisting a user.” It does not describe the file as a ranking signal, and the file is not a ratified standard.

Google Search vs Lighthouse

Google’s generative AI guide says that Google Search does not use llms.txt files, including for its generative AI features, and that creating one for other services neither helps nor harms visibility in Google Search. Google added that clarification to the guide on June 15, 2026.

Lighthouse, Google’s open-source tool for auditing web pages, is a separate product. Since version 13.5.0, released in September 2026, it has included an llms.txt audit in an “Agentic Browsing” category that its source code describes as “still under development and subject to change.”

The audit checks whether the file is well formed: for example, whether it is a Markdown file with an H1 heading and links. It is a quality check for sites that want to be usable by AI agents, and it tells you nothing about whether Google Search reads the file.

Publishing a file vs AI systems reading one

Several AI companies publish llms.txt files for their own developer documentation, including OpenAI, Anthropic and Google’s Gemini API, as the proposal itself notes. Perplexity’s documentation site does the same. That is publication, for their own documentation. No public statement was found that their search or answer products read other sites’ files.

Server-log studies published by SEO and security vendors point the same way. A 2026 Ahrefs analysis of llms.txt requests across sites in its analytics data reported that 97% of the files it observed received no requests in May 2026. Most fetches came from bots that are not AI tools, and the search crawlers of ChatGPT, Perplexity and Claude fetched the files only rarely. A much smaller study that checked requests against crawlers’ published IP ranges found no verified AI-crawler fetches over four days of logs.

These are vendor studies, not platform statements, and a fetch does not show that a file was used. They support “little observed consumption”, not “never read”. Counts of how many sites publish llms.txt measure adoption by site owners, not use by AI systems.

“Chunking” means three different things

You do not need to break content into small pieces for AI Overviews or AI Mode in Google Search; Google says so directly. Bing recommends clear structure such as headings and tables. These positions do not conflict, because “chunking” is used for three different things:

  1. Writing bite-sized content for AI: a publisher deliberately rewriting or splitting content into small, AI-targeted pieces.
  2. Clear document structure: ordinary sections, headings, lists and tables that make a page easier to read and navigate.
  3. Retrieval chunking: an AI system splitting documents into segments internally so it can retrieve relevant passages. This is part of retrieval-augmented generation (RAG), where a system looks up relevant text before generating an answer.
Claim Platform / product What the platform has said Status Basis
Content should be split into tiny pieces for AI Google Search (AI Overviews, AI Mode) Not needed for its AI features Contradicted Official documentation · May 2026
Clear headings, tables and FAQ sections help AI reference content Bing / Copilot Recommended Supported (Bing’s own claim) Official blog · Feb 2026
AI systems prefer bite-sized content ChatGPT search; Perplexity; Claude No public publisher guidance found Unproven Official documentation checked; nothing on this point
Retrieval systems split documents internally OpenAI API; Bing Documented as system design Supported Official documentation (OpenAI API); official blog (Bing) · May 2026

Writing bite-sized content for AI

This is the meaning Google’s guidance addresses. Its generative AI guide says there is no requirement to break content into tiny pieces for AI to understand it better. No guidance was found from any platform in this review recommending that publishers rewrite content into AI-targeted fragments.

Clear structure: sections, headings and tables

This is closer to what Bing recommends: clear headings, tables and FAQ sections that make information easier for its AI systems to reference. Read in context, Google’s statement is about rewriting content into small pieces for AI, not about ordinary headings and sections.

Bing’s guidelines say well-structured content helps retrieval and citation accuracy. No platform data was found, however, measuring whether clear structure changes how often a specific AI system cites a page.

How retrieval systems split documents

Retrieval chunking happens inside AI systems, and the documentation that describes it is written for the people who build those systems. OpenAI’s API documentation for its file search feature, for example, splits files into chunks of 800 tokens with 400 tokens of overlap by default. That is a developer feature, not a description of ChatGPT search.

Bing’s May 2026 post on the role of its index treats keeping a page’s meaning intact, when the page is split and transformed to ground AI answers, as a challenge for the index itself. The post is about Bing’s own system, not guidance for publishers.

Some practitioners, such as iPullRank, argue that Google’s advice conflicts with how retrieval-augmented generation works. Separating the three meanings narrows that disagreement: retrieval chunking is real, but the system does it, and no platform guidance was found telling publishers how to pre-chunk content for it. Whether publisher-side structure changes retrieval results in a particular engine remains open.

Is optimizing for AI search different from SEO?

For Google Search, no. Google’s generative AI guide says its AI features are rooted in its core Search ranking and quality systems, so optimizing for them is still SEO. Google’s AI features documentation says a page is eligible to appear as a supporting link in AI Overviews or AI Mode if it is indexed and eligible to be shown in Google Search with a snippet, with no additional technical requirements.

Google also documents a Search Console setting, Search generative AI, that includes sites in these features by default and lets a site owner exclude a site from them. It had rolled out to all sites by August 31, 2026.

For the other platforms, the published guidance does not answer the question either way. Bing gives content-level advice for AI, covered above. OpenAI and Perplexity document their search crawlers and how to allow them, and Anthropic documents its crawlers and controls. Whether any of these platforms rewards anything beyond ordinary, accessible, well-made pages is an open question.

What official sources say to prioritize instead

Google Search, per its generative AI guide, AI features documentation, helpful content guidance and May 2025 post on AI experiences in Search:

  • Make sure pages can be crawled and indexed, and are eligible to show with a snippet.
  • Publish original, helpful, people-first content, rather than material that could be easily produced by a generative AI model or repeats what is already widely available.
  • Provide a good page experience.
  • Keep structured data accurate, consistent with visible content, and used for the rich results it supports.
  • To limit how content appears, use the existing preview controls, such as nosnippet, data-nosnippet, max-snippet and noindex, or the Search generative AI setting in Search Console to exclude a site from these features.

Bing and Copilot, per its Webmaster Guidelines, AI Performance announcement and index post:

  • Use clear headings, tables and FAQ sections so key information is easy to reference.
  • State facts and definitions explicitly, and keep each page accurate and focused on its topic.
  • Keep facts current. Bing’s index post says a stale fact produces a misleading answer when it is used to ground an AI response.

ChatGPT search and Perplexity: their published publisher guidance is about access. To be eligible, make sure their search crawlers, OAI-SearchBot and PerplexityBot, are not blocked.

Claude: Anthropic’s crawler documentation explains its bots and how site owners can control them. No guidance on what to prioritize for visibility in Claude was found.

What is still unknown

  • Whether ChatGPT search, Perplexity or Claude give any weight to schema.org markup when choosing sources to cite.
  • Whether any AI search or answer product reads publishers’ llms.txt files at query time.
  • Whether adding schema helps pages that AI systems do not already cite; the controlled study above only covered heavily cited pages.
  • Whether clear page structure measurably changes citation rates in a specific AI system.
  • How Lighthouse’s agentic browsing checks will develop, given that the category is marked as under development.

Sources and verification

The primary sources used in this review, all checked on October 10, 2026:

Platform guidance on this topic changed several times in 2026. This article will be reviewed regularly and whenever one of these platforms publishes new guidance. For the wider areas this review sits within, see SEO & AEO.

Back to Insights