A Practical Guide to AI Overviews

Tomohiro Iida · Published: June 26, 2026Last updated: September 7, 2026

Google's AI Overviews places an AI-generated summary above the regular search results, answering a user's question by drawing on multiple web pages and listing its sources. This guide is a deep dive focused specifically on AI Overviews for BtoB companies: what it is, how display varies, ordinary SEO groundwork, clear passage structure, structured data and first-party information, measurement, AI crawler handling, and the realistic limits of any of this. There is no way to guarantee that AI Overviews will display for a given query, and no vendor or technique can promise otherwise. We treat answer-first and one-point-per-paragraph writing as clarity practices; we have not observed evidence that those formats themselves increase citation rates.

What AI Overviews is

AI Overviews (labeled "AIによる概要" in Japan) is a Google feature that places an AI-generated summary above the search results, drawing on multiple web pages to answer a user's question and listing source links. It is the production version of what was previously an experimental feature called SGE (Search Generative Experience); people still search for "SGE optimization," but the current name is AI Overviews. Because it presents an answer directly rather than a list of links, users may get what they need without opening any individual page, which changes how visibility translates into traffic: being cited surfaces a company's name and link as the basis for the answer, while not being cited means depending on the regular results below it. Whether AI Overviews appears at all varies by query type, region, and time at Google's discretion; no method can guarantee it will show for a given search, and this guide focuses on the site-side foundations and observation methods that can be controlled without claiming a citation effect.

How display varies

• AI Overviews tends to appear for "informational" queries that can be summarized, such as explanations, how-tos, and comparisons, and less often for queries naming a specific company, highly time-sensitive searches, or contentious topics. The same keyword can show or not show depending on region, device, and time, so a single observation does not establish the real state.
• Sources overlap with top search rankings but do not always match them; a page can be cited without ranking at the top if a particular passage fits the answer, and a top-ranked page can go uncited. Before treating "we are not showing up" as a problem, first check whether AI Overviews appears at all for that query.
• Because display logic changes quickly, betting on optimizing for this one feature is less durable than building a site that communicates clearly to any AI search engine.

The foundation: ordinary SEO

• Indexing and crawling — pages must be crawled and indexed; misconfigured noindex tags, robots.txt blocks, incorrect canonical tags, frequent 5xx errors, or JavaScript-only rendering can keep a page out of consideration entirely, and if this is broken, nothing downstream matters.
• E-E-A-T (experience, expertise, authoritativeness, trust) — the author's real experience, cited sources, and clear operator/contact information help readers and search engines understand provenance and context; first-party information adds distinct material beyond a rehash of general knowledge, without guaranteeing citation.
• Matching search intent — a page needs to answer the actual intent behind a query, such as "what is" or "how to," rather than speaking in corporate-brochure language; structuring content as explanation, procedure, or comparison helps, since AI Overviews favors informational queries.
• Topic coverage and internal linking — pairing a comprehensive hub page with linked spoke articles signals depth on a topic; isolated pages with no internal links tend to be evaluated less favorably.

Writing clear, extractable passages

• State the conclusion first (answer-first) — put the answer at the top of the paragraph rather than building up through background first; a "conclusion, then reasoning, then detail" structure makes the main point easier to find and extract. This is a clarity claim, not a claim about citation probability.
• One point per paragraph — mixing multiple points together makes a paragraph harder to lift out and still make sense on its own; splitting by point so each paragraph is self-contained helps.
• Attach concrete evidence, figures, and conditions — a specific statement such as "under this condition, this metric improved by this amount" is more usable as answer material than a vague claim of effectiveness; cite sources where available, but never fabricate a figure that cannot be verified.
• Answer question-form headings directly — a heading phrased as a question, followed immediately by a concise answer, makes the correspondence clear and easier for a reader to scan; FAQ-style sections organize information in the same way. We have not observed that this format itself raises citation rates.
• Make text machine-readable — organize key points as lists, tables, and headings; content locked inside an image or a PDF-only document is less likely to be picked up.

Structured data, first-party information, and AI crawlers

Structured data (commonly JSON-LD) is a machine-readable annotation that helps convey a site's content accurately to search engines and AI. For a BtoB site, the core types are Organization, WebSite, BreadcrumbList, and Article or BlogPosting, with FAQPage used as a supplementary layer once the FAQ content is already clearly written in the body. It does not improve the content itself or guarantee citation or rich results, and structured data that does not match reality is counterproductive. First-party material such as a company's own procedures, case examples, judgment criteria, or figures adds information that a generic summary does not, and consistent name, address, and contact details (NAP) across the site reduce ambiguity about the entity. These are content-quality and identity practices; they do not guarantee citation in AI Overviews.

AI Overviews itself is a Google search feature: being included depends on ordinary indexing by Googlebot and on snippet controls such as nosnippet, and is separate from OpenAI's or Perplexity's crawlers, which serve ChatGPT and Perplexity rather than Google's AI Overviews. GPTBot (OpenAI, model training) and ClaudeBot (Anthropic, model training) are blocked or allowed as a matter of company policy on AI training use, while OAI-SearchBot (OpenAI, powers ChatGPT search) and PerplexityBot (Perplexity, powers Perplexity search) need to stay unblocked to appear in those services' AI search answers; training access and search-display access can be set independently. Google's "Google-Extended" is not a separate crawler but a robots.txt token controlling training/reference use by Gemini; blocking it does not affect Google Search crawling, indexing, ranking, or AI Overviews display.

Measuring results, and the limits of this work

Because display is not guaranteed and changes quickly, treat AI Overviews as one part of overall search and AI visibility rather than measuring it in isolation. Search Console (impressions, clicks, average position, CTR by query, indexing status), GA4 (sessions by channel and conversions by landing page), and whether AI services appear as a referral source are the practical metrics; Search Console cannot reliably isolate "clicks from AI Overviews" specifically, so the realistic approach is to track the broader trend over time rather than attribute results precisely. In short: display cannot be guaranteed, and any service promising otherwise should be treated with suspicion; betting on this one feature is less durable than building a foundation that serves any AI search engine; and AI Overviews is only one part of search traffic, so without a solid regular-search presence and a working inquiry funnel, appearing in AI Overviews alone will not produce results.