SEO news: May 2026

Bing Webmaster Tools: reporting for AI search

While Google faces criticism for its lack of transparency about AI Overviews data, Microsoft revealed the future of its Bing Webmaster Tools at the SEO Week 2026 conference in New York.

Key metrics and features of the new reporting

• Citation Share: a brand new metric that shows you what percentage of the total citation volume in AI responses Bing assigns to your site compared to your competitors.
• Grounding Query Intent: Bing can now categorize user queries into 15 predefined AI intents. Specialists can finally see what type of deep contextual search (e.g., comparisons, fact-finding, creative solutions) the AI is using your site for.
• GEO-focused Recommendations: a module that will directly generate tips on how to optimize your site for Generative Engine Optimization (GEO ). Based on the data collected, the tool will suggest how to adjust the structure and data density of your content to make AI models cite you more often.

For the SEO community this is a major news, for the first time we will get our hands on an official analytics interface directly from the search engine. The gap in data transparency between Microsoft and Google is widened again with this step.

23 factors that determine whether AI quotes you

Getting a citation in generative search (such as Google AI Overviews, Perplexity or ChatGPT) is becoming the new equivalent of fighting for the top position on Google. Cyrus Shepard has compiled a ranking of factors based on an analysis of dozens of studies and confirmed that traditional SEO fundamentals are key to AI success, but are complemented by new requirements for text formatting.

The strongest factors for getting AI quotes

• Traditional Search Rank: Research (e.g. by Ahrefs) confirms that approximately 38% of citations in AI Overviews come from sites that are already in the top 10 organic results. Thus, traditional SEO is still the strongest stepping stone.
• Intent-Format Match: the AI prefers sites whose structure exactly matches the query type. For queries like "best..." (best) queries, the AI massively cites comparison tables and lists (listicles). For "how-to" queries, on the other hand, it requires clear step-by-step instructions.
• URL Accessibility and Indexing: the ability of AI bots to seamlessly crawl a page's code and read its content without technical barriers is absolutely essential.
• Structured data (Schema): although the correlation is slightly weaker than for text alone, the vast majority of studies have shown that properly deployed Schema markup helps AI to interpret entities correctly and increases the chance of inclusion in a response.
• The first 30% rule: analyses of the behaviour of LLM models show that 44.2% of all citations come from the first third of the article. AI models divide the text into smaller parts before processing and prefer to cite clear, declarative sentences placed right at the beginning of the text or paragraph.

Other attractions

• Ignoring llms.txt: The llms.txt file, which was supposed to serve as a new standard for communication with AI robots, received the lowest score in the importance rating (only 2.0 out of 10). For the citation algorithm itself, it is currently of almost no relevance.
• Language and local bias: AI search engines show a strong link to the location and language of the user. A French query from Paris will almost 100% generate only French citations from local domains.

AI citation optimization (GEO) does not reject traditional SEO, but builds on it. For AI to recommend your site, you need to maintain strong domain authority, write concise declarative answers right at the top of the page, and structure your content into tables and lists that AI can most easily read and process.

Analysis of 5 million query fanouts reveals what AI is looking for in the background

When a user enters a query into ChatGPT or Perplexity, the AI doesn't just look for that specific entry. In the background, it runs a series of hidden additional queries (called query fanouts) to expand context, search for reviews, compare brands, and verify data. This analysis shows that if your content doesn't answer these hidden queries, the chances of getting cited in AI drop dramatically.

Key Findings: How do AI models change user queries?

• Reciprocal Rank Fusion (RRF) algorithm: ChatGPT uses this algorithm to fuse results from multiple hidden sub-queries. A page that can answer multiple of these hidden sub-queries at once scores significantly higher and is more likely to be cited by the AI.
• The word "Best" as the absolute king: 24.3% of the time (almost one in four referral queries) ChatGPT secretly inserts the word "best" into the search, even if the user has not used it at all. The query "Which Samsung should I buy?" changes to "best Samsung Galaxy phones comparison 2026" in the background. This is why listicles score so massively in AI answers.
• "Reviews" in third place: AI search engines automatically search for third-party reviews. For businesses, this means that how AI talks about them is not determined by their own website, but by what people write about them on platforms like G2, Glassdoor or Sitejabber.
• Emphasis on timeliness: in 5.44% of cases, ChatGPT automatically adds the current year (2026) to hidden queries to force fresh and updated sources.

How does the behaviour of each AI search engine differ?

The study revealed huge differences in how comprehensively different tools approach background search:

• Perplexity (1.4 sub-query per prompt): does the least work. It just cleans the user query of filler words, simplifies it, and enters it into the search engine. It offers no new angle or deeper context.
• ChatGPT (2.1 sub-query per prompt): sticks to user intent, but actively adds brand-specific comparisons and keywords like "best", "top" or "reviews".
• Grok from xAI (6.8 sub-queries per prompt): works like a real researcher. One query (e.g. "best car camera") breaks down into 5 to 8 detailed searches. It purposely limits the search to the current year, and uses the site: operator to force results from its community-verified sources (massively searches Reddit, Wirecutter, ConsumerReports, or G2).

In practice, this means structuring texts to include clear recommendations, comparisons with competitors, statistics, user reviews and clear answers at the beginning of sections, because these are exactly the parameters that the background search engine algorithms require.

Where does the AI get its information from? The three layers of data that define the visibility of tags

When an AI (ChatGPT, Gemini or Perplexity) answers your query, its knowledge doesn't come from one magical place. The entire system relies on three distinct technological layers of data. Each of them has its own specifics, expiration dates and risks of errors (hallucinations).

The three layers from which the AI draws

• Training data (The Foundation): this is a gigantic static dataset (public websites, Wikipedia, books, licensed databases) on which the model was trained before its launch. This is where the AI makes semantic associations about your brand (e.g. that the brand "Patagonia" belongs to the concept of "ecology"). Problem: Once the training is over, the data is frozen. The AI doesn't know what happened yesterday, and if information is missing, it tends to hallucinate.
• RAG or Real-time Anchoring: Retrieval-Augmented Generation technology works like an open-book test. When you ask a query, the AI first looks up the actual documents through traditional search engines (Google, Bing) and generates an answer based on those documents. This drastically reduces the risk of hallucinations and the AI retrieves the actual data.
• MCP Interface and API (The Execution Layer): the most advanced layer for AI agents. Models use Model Context Protocol (MCP) and direct APIs to connect to live specialized databases (e.g., Ahrefs' AI agent connects directly to real SEO keyword and backlink data in this way). AI in this mode does not read text from the web, but works with clean structured data.

What are the implications for marketing and brands?

If you want the AI to find and recommend you, you need to feed all three layers:

• Build off-site mentions (PR and discussions): for AI to fixate on you in the training data, you need to exist outside your website - in the media, on Reddit, Wikipedia or professional forums. A website on your own domain is invisible to the basic training model.
• Cover the semantic neighborhood (Query fan-out): AI often decomposes queries into sub-queries. If you're selling a design tool, you need to cover related topics as well (e.g., "agile vs. waterfall"), because those are the ones AI will pull in through the RAG layer to explain the broader context to the user.
• Pure technical accessibility: for the RAG layer you need to have a perfectly traversable website (pure HTML, speed) so that bots can download your content in a flash and process it as a verified source (ground truth). In contrast, the llms.txt format is not yet realistically respected by any major player in 2026.

Traditional search engine optimization is not dying, but transforming. By making massive use of Google and Bing indexes in the RAG layer, high organic rankings remain the best way to get into the "open book" from which AI generates its answers and citations in real time.

Google I/O 2026: Search completely changes its face after 25 years, the era of AI agents is upon us

At this year's Google I/O conference, the biggest search box upgrade in a quarter century resonated. The search engine is finally leaving the role of a mere link broker and transforming into a complex agent platform that not only searches for the user, but also performs tasks. This entire ecosystem is powered by the newly introduced Gemini 3.5 generation.

The main technological pillars of the new search

• Transition to Gemini 3.5 Flash: This new model becomes the default engine for AI Mode in search. Compared to previous versions, it delivers 4x faster answer generation and an extreme shift in tool orchestration. It can process huge amounts of data in parallel with minimal latency and cost.
• Always-on Agents: users get the ability to create and run autonomous search agents (led by the Gemini Spark assistant). These micro-services run continuously in the background: they crawl the web themselves, monitor changes, structure data, compare offers and perform complex multi-step tasks without the user having to manually click on links.
• Deep multimodal integration: the search field now fully supports real-time combined inputs. The user can shoot video, speak and type a text prompt at the same time - the search engine analyzes the visual and audio context together and responds immediately to the situation.

Personal Intelligence and Privacy

The new search securely connects to the user's personal apps such as Gmail, Google Photos and Google Calendar. Google is building on Personal Intelligence, which gives the user absolute control over what data they make available to agents, with all sensitive data subject to the highest security standards (Frontier safeguards).

Google I/O 2026 confirmed that we are experiencing a change in traditional SEO focused only on keywords and positions. In the agent era, people will no longer crawl sites to find the answer - software agents will do it for them.

Sources

https://searchengineland.com/bing-webmaster-tools-teases-new-ai-reporting-updates-475659
https://signal.zyppy.com/p/ai-citation-ranking-factors
https://peec.ai/blog/patterns-we-see-in-chatgpt-query-fanouts
https://ahrefs.com/blog/how-does-ai-get-its-information/
https://blog.google/products-and-platforms/products/search/search-io-2026/