SEO 28 May 2026 13 min read 116 view

Deprecation of FAQ Schema in Google: What it Means for SEO, GEO, and AI Search

Updated: 28 May 2026

Language: 🇺🇦 🇺🇸 🇩🇪 🇪🇸

Vadim Kharovyuk

CEO & Founder of WebsCraft. 8 years in web development, focused on bringing AI into real products.

Deprecation of FAQ Schema in Google: What it Means for SEO, GEO, and AI Search

Announcement. On May 7, 2026, Google permanently disabled FAQ rich results for all sites without exception. This is the culmination of a process that began back in August 2023. But if you think it's just about the disappearance of accordions in search results, you're mistaken. Behind this technical decision lies a fundamental transformation of how Google and AI systems consume, process, and synthesize content. This article provides a detailed analysis: from the timeline of changes and the engineering of schema markup to the practice of GEO (Generative Engine Optimization) and how to optimize your site for next-generation retrieval search.

Official Google documentation on FAQPage structured data →

1. What Exactly Did Google Change — and Why Was It Expected

Timeline of Deprecation

The process of eliminating FAQ rich results occurred in two distinct stages. The first was in August 2023: Google announced that FAQ accordions in search results would henceforth be available only for "authoritative government and medical sites." According to John Mueller, Search Advocate at Google, this was a change in result display, not ranking. Within a week of the announcement, SEO tools across the industry recorded the same thing: the number of FAQ rich result impressions for commercial sites plummeted to zero. Pages that had carried FAQ snippets for years lost them overnight.

The second stage was on May 7, 2026: Google officially announced that FAQ rich results would no longer be displayed in Google Search for anyone. In June 2026, the corresponding report in Search Console and support in Rich Results Test will be disabled. In August 2026, support for FAQ rich results in the Search Console API will be phased out. The deprecation is complete.

What Disappeared from Search Results

For the absolute majority of websites — which is 99%+ of the open web — the SERP value of FAQ schema has become zero. Interactive accordions that expanded directly in search results disappeared, the visual size of snippets decreased, and console reports ceased. The March 2026 update further reduced FAQ rich result impressions by approximately half compared to the post-2023 level — even for sites that retained residual visibility.

Why It's Logical for Google

I think the reasons for this decision were obvious to anyone who has long observed the SEO market. FAQ markup gradually transformed from a way to structure useful answers for the user into a tool for capturing additional space in SERPs. Many teams added artificial FAQ blocks not because the page genuinely needed a "question-answer" format, but solely to increase the visual area of the snippet. As a result, this created more noise than value for search. In my opinion, Google solved the problem radically: instead of endless quality control of individual implementations, the company simply removed the display mechanism itself for most sites. This is clearly visible in the gradual phasing out of FAQ rich results in search.

But there is a deeper, strategic reason: Google is systematically moving towards generative synthesis of answers, rather than delegating them through extended snippets to third-party sites. AI Overviews now appear in 89% of search results for branded queries. Why show an accordion from someone else's site when Gemini can synthesize the answer itself?

Zero-Click: The New Normal

Statistics from 2025–2026 illustrate the scale of the transformation. According to Similarweb and SparkToro, 58.5% of searches in the US and 59.7% in Europe end without a single click. When an active AI Overview is present, the figure increases to 83% and higher. And in Google AI Mode, 93% of searches do not generate clicks at all. The search page is no longer a gateway; it has become the final destination.

2. The UI Layer and the Semantic Layer Are Different Systems

The Key Distinction That Is Lost in Most Publications

The most important detail in Google's official announcement — the one that is easily missed: Google explicitly stated that it will continue to use FAQ structured data to better understand pages, even after ceasing to display rich results. This phrase confirms what the SEO community has been debating since 2023: structured data and rich results are two different things.

Schema markup tells Google what a page is about in a machine-readable format. Rich results are a display function that used some of this data to show visual elements in SERPs. Google can stop showing a visual function without abandoning the data that informs its semantic model.

The UI Layer Has Lost Value — The Semantic Layer Remains Relevant

The old playbook from 2018–2022, which treated FAQ schema as a CTR multiplication tool, no longer works. But this doesn't mean Schema.org has lost its meaning altogether. On the contrary — Schema.org has evolved from a display signal to a learning signal: LLM systems don't "parse" structured data in the classic sense — they absorb it, embed it into their internal knowledge graph, and reuse it to generate answers.

What Remains Relevant for Large Projects

Enterprise projects, e-commerce, and knowledge sites continue to benefit from structured data. Among the schema types that still generate rich results are: Product, Review and AggregateRating, Article, Recipe, Video, Organization, LocalBusiness, BreadcrumbList. These types will survive any changes — because they carry real content signals for Google and AI systems, rather than just capturing SERP real estate.

3. Googlebot ≠ AI Agent: Modern Search Has Two Different Content Consumption Models

The Classic Search Bot

Googlebot operates within a deterministic paradigm: indexing, canonicalization, building a link graph, parsing structured data, and rendering the DOM structure. It looks for clear signals—meta tags, canonical URLs, schema hints—and processes them according to formal rules. Schema.org was initially created precisely for such deterministic parser models.

AI Agents and LLM Systems

AI agents—Perplexity, ChatGPT Search, Google AI Overviews, Copilot—function fundamentally differently. Their work is based on retrieval, semantic extraction, chunk relevance scoring, answer synthesis, and probabilistic interpretation of content. They don't just look for formal signals in HTML or schema.org; they attempt to interpret the document's meaning at a semantic level. This is why, for modern AI systems, text structure, section logic, and context are often more important than the presence of FAQ JSON-LD. I've discussed this in more detail in the article about Perplexity's RAG architecture and retrieval-based AI search.

The key difference: where Googlebot reads a schema attribute and checks its compliance with the allowed vocabulary, an LLM reads the entire document and builds a semantic model of its content. Unlike Google's parsers, which check what's allowed, LLMs try to understand what is.

Consequences for Content Architecture

AI search is increasingly oriented not towards formal schema hints but towards the logical structure of the document and the quality of textual context. This means that well-structured HTML with a clear semantic hierarchy can yield a greater retrieval effect than overloaded JSON-LD with minimal textual content.

4. Under the Hood of AI Search: How LLMs Actually Process HTML

Evolution of Parsing: From Regex to Semantic Retrieval

Early web scraping systems used regex and CSS selectors—simple, deterministic tools. Modern AI retrieval pipelines are fundamentally different and involve several sequential processing layers.

How AI Sees a Web Page

The first step is HTML sanitation and DOM simplification: the system removes navigation noise, ads, scripts, tracking elements, footers, and menus. Only the main content remains. Then, HTML is converted into a text structure—often a Markdown-like representation with a flattened DOM tree. The document is simplified before being fed into the model's context window.

What happens to JSON-LD in this process? Structured data is not fed directly into the model; it's processed separately. In some retrieval pipelines, JSON-LD might be ignored during document cleaning. AI systems are increasingly receiving information directly from textual content rather than from auxiliary markup.

Chunking and Embeddings: The Heart of a Retrieval System

After cleaning, the document is broken down into semantic blocks—chunks. The choice of chunking strategy directly impacts retrieval quality: segmenting documents into smaller, semantically concentrated blocks ensures that the retrieved data fits within the LLM's context window while minimizing the inclusion of distracting or irrelevant information.

Each chunk receives its own vector embedding—a numerical representation of the text's semantic meaning. This allows the AI system to work not with individual words but with the context and meaning of a fragment. I've already explained in detail how embeddings help models "understand" the meaning of text in the material "Embeddings in Simple Terms: How AI Understands Meaning, Not Just Words."

During a query, the system finds the semantically closest chunks through nearest-neighbor search and passes them into the model's context for answer synthesis. This is how Retrieval-Augmented Generation (RAG) works—an architectural approach that underlies visibility in most modern AI search systems.

Why Semantic HTML is Becoming More Important

From the perspective of a chunking system, the connection between an `

` heading and its corresponding `
` paragraphs is critical for correctly delineating semantic blocks. The document's logical hierarchy, minimization of noise, and efficient use of the context window enhance retrieval quality for AI agents.

This also raises the issue of token efficiency: overloaded markup code complicates extraction and reduces the signal-to-noise ratio for LLMs. Traditional crawlers follow links and parse HTML; LLM engines do the same plus entity extraction. JSON-LD offers an easy out-of-band signal they can absorb without natural-language parsing. But if this signal is noisy or irrelevant, it becomes an obstacle.

5. A Practical Approach to Code: Keep FAQ Schema or Abandon It?

Scenario A—Existing Projects: The No-op Strategy

For sites where FAQ schema is already implemented—Google has explicitly stated: there's no need to rush to remove FAQPage structured data. Unused valid schema does not incur any penalties from the search engine. The markup does not create direct penalization. If it describes the actual Q&A content of the page, keep it. It's not worth spending engineering resources on mass deletion.

There's even an argument for keeping it: Perplexity, ChatGPT Search, Gemini, and Google AI Overviews parse FAQ schema as a primary signal when extracting Q&A answers. Pages with clean FAQ schema are disproportionately cited in AI answers compared to pages with the same content in plain format. This claim remains debatable (a separate Search/Atlas study from December 2024 found no correlation between schema coverage and citation frequency), but it demonstrates that the data is ambiguous.

Scenario B—New Projects and Releases

The approach here is different. There's no point in building complex FAQ-generator pipelines if the ultimate goal is only SERP display, which no longer exists. Reducing backend/CMS complexity and reorienting engineering resources towards content architecture and retrieval optimization is a rational decision.

What Remains a Priority

The following types of schema retain full relevance and continue to generate rich results or provide semantic value for AI systems:

Organization—key for entity recognition in AI systems;
Article—a fundamental signal for content pages;
Product + Offer—critical for e-commerce, AI agents cite specific prices;
BreadcrumbList—semantic context for thematic clustering;
canonical metadata—basic technical SEO hygiene.

What is becoming less of a priority: schema spam, FAQ-driven CTR engineering, excessive content duplication in JSON-LD without corresponding textual content.

6. Transitioning from SEO to GEO: How to Optimize Content for AI Search

GEO as a New Layer of Optimization

Generative Engine Optimization (GEO) is the practice of structuring content and managing online presence to increase visibility in the responses generated by AI systems: ChatGPT, Perplexity, Google AI Overviews, Claude, Copilot. GEO influences how LLM systems extract, summarize, and present information in response to user queries.

The scale of the transition: AI-referred sessions have grown by 527% year-over-year in the first five months of 2025, according to the Previsible 2025 AI Traffic Report. 43% of marketers are actively implementing GEO strategies, compared to almost zero in 2025. GEO is no longer a niche discipline.

The Principle of Logical Atomicity

The key requirement for retrieval systems: one section = one complete thought. AI agents extract chunks and synthesize them into a response. If a content block is not self-contained, it will either not be retrieved or will lose context during extraction.

Rule of thumb: the answer to the intent is in the first sentences of the section. AI systems with real-time retrieval assess page relevance primarily by its initial content. The first 200 words of any article must directly and fully answer the main query, not lead up to it.

Document Semantic Structure

Instead of schema-overengineering, focus on the logical hierarchy of the document: proper use of <article>, <section>, consistent H2–H4 hierarchy, clear semantic boundaries between blocks. Generative engines parse meaning, not keywords. Content optimized for GEO is structured around clearly defined entities, statements, and relationships.

Linguistic Precision for Embedding Models

Embedding models work better with established terminology, clear entity definitions, and minimal ambiguity. Key principles of effective GEO: structure content with direct answers in the first 40–60 words, maintain fact density with statistics every 150–200 words, cite authoritative sources, and implement correct schema markup.

Content for a Retrieval-First Web

The winner will not be the site that manipulates SERP UI better, but the one whose content is built on the principles of:

Chunkable sections — blocks that can be extracted while retaining full meaning;
Explicit definitions — terms are defined directly in the text, not through links;
Contextual self-sufficiency of blocks — each chunk is understood without necessarily reading the rest of the page;
Readable-by-agents structure — minimal decorator noise, maximum information density.

GEO is not a replacement for SEO, it's an additional layer. Brands that succeed in GEO in 2026 are typically the same ones with a strong traditional SEO foundation. The optimization principles overlap significantly, but GEO adds specific requirements for content structure, citation-friendliness, and information saturation that SEO alone does not cover.

What Website Owners Should Do in 2026

In my opinion, the main mistake after the deprecation of FAQ rich results is a panic reaction. Some teams started mass-deleting FAQ schema as if the markup itself suddenly became "harmful." I don't think this is the right approach. The problem isn't the existence of schema.org, but that the market has used it for years primarily as a UI hack to increase snippet real estate.

I don't see the point in mass-deleting FAQ JSON-LD from old projects if it already exists and is supported without additional costs. The presence of schema itself does not create a penalty. But at the same time, I wouldn't build an SEO strategy around FAQ blocks and CTR engineering through structured data anymore.

In my view, in 2026, the focus is shifting in a different direction: from manipulating SERP UI to optimizing content for retrieval and AI interpretation. This means that the key is not the amount of schema markup, but the quality of the page's semantic structure.

I would pay attention to a few things:

logical document structure;
consistent H2/H3 hierarchy;
clear semantic boundaries between sections;
minimization of boilerplate and noise content;
contextual self-sufficiency of text blocks.

Modern AI systems work through retrieval pipelines, chunking, and embeddings. Therefore, content increasingly needs to be chunk-friendly — meaning it can be easily broken down into separate logical fragments that can be extracted, analyzed, and used to generate a response.

I also think that entity clarity is an underestimated topic. For AI agents, it's not just about which words are in the text, but how unambiguously the document explains entities, terms, and the relationships between them. The less ambiguity, the easier it is for the retrieval system to correctly interpret the content.

Separately, I would treat H2/H3 not just as design elements or SEO structure, but as retrieval boundaries. In many AI pipelines, headings help the system define the boundaries of semantic chunks and understand which block answers a specific intent.

And most importantly, in 2026, the winner will not be the site with the most aggressive schema optimization, but the site whose content is easier to:

parse;
chunk;
embed;
interpret;
cite by AI systems.

7. Conclusion: FAQ Schema Isn't "Dead" — The Search Model Itself Has Changed

I believe the deprecation of FAQ rich results is just a symptom of a much deeper transformation in search. The real change is that Google and the entire modern AI search stack are gradually moving from a system of rich snippets to a system of generative answer synthesis. The logic that defined SEO optimization from 2018–2022 — fighting for SERP real estate, FAQ accordions, maximizing snippet expansion — is no longer central.

In my opinion, modern AI systems rely less and less on auxiliary display signals and increasingly derive meaning from the structure, context, and semantic integrity of the document. This doesn't mean schema.org has lost its value. Rather, its function has changed: from a display trigger to a semantic fingerprint, from a tool for increasing CTR to one of the additional learning signals for retrieval and LLM systems.

The new reality for the web requires a re-evaluation of priorities:

Parse — content must be free of noise;
Chunk — blocks must be semantically self-sufficient;
Embed — terminology must be precise and unambiguous;
Cite for AI systems — the structure must be such that an LLM can lift a block and insert it into an answer without loss of meaning.

Categories