How Google AI Overviews surface sources

First, what happened to SGE?

Google introduced Search Generative Experience (SGE) in 2023 as a Search Labs experiment, as its announcement explains. The current feature is called AI Overviews; older screenshots may show a past interface or behavior. The useful publisher question is how Google describes its present Search AI features finding supporting information, not how to optimize for an old SGE layout.

Google describes AI Overviews as part of Search, not a separate index to which publishers submit pages. Its AI features documentation says Overviews appear when systems determine they add value beyond classic Search, so they do not appear for every query.

The basic path: retrieve, then generate

Google describes the approach as retrieval augmented generation (RAG), or grounding. Search ranking systems retrieve relevant, up to date pages from the Search index; models review information from those pages and generate a response with links intended to help people explore supporting material. This is not simply a copy of one page or a summary of only the first conventional result.

A page may contribute a fact or explanation without being reproduced word for word, and a response may draw on multiple sources. Google describes the broad architecture, not which passage supplied every sentence in a particular answer. For the underlying concept, see Rankwox’s RAG explainer.

Why one question can trigger related searches

A broad question can contain several information needs. Google says AI Overviews and AI Mode may use query fan out: issuing multiple related searches across subtopics and data sources. Its example starts with a question about a weed filled lawn; related searches might cover herbicides, non chemical removal, and prevention. This illustrates the technique, not a template for every query.

For “Which payroll system suits a 40 person restaurant group?”, related needs might include scheduling integrations, tax handling, and multi location reporting. A page that genuinely answers one need may be relevant without repeating the whole question. Publishers cannot see or predict the exact searches generated, and mass producing thin pages for imagined variations is not a sound response. Google favors useful, distinctive material for people.

What the links do and what they do not prove

Links in an Overview give readers routes to pages Google considers relevant and a way to investigate further. They show that a page appeared as a supporting link for a particular search at a particular time; they do not reveal the complete selection process, prove every sentence came from that page, or establish its conventional organic position.

Results can change with the query, sources, models, and product presentation. Google’s AI Overviews help page warns that generated responses can contain mistakes and advises checking important information against supporting pages and other results. Record query, date, market, answer, and links as a snapshot not a durable placement commitment.

Eligibility is a gate, not a citation strategy

Google says a page must be indexed and eligible to appear in Search with a snippet to qualify as a supporting link in AI Overviews or AI Mode. There are no additional technical requirements. Make sure Googlebot can access the page, essential information is available as text, and the content is useful and reliable. Structured data should match visible page content; special AI schema does not guarantee inclusion.

A site level control also applies. Google’s Search generative AI control documentation says owners can manage inclusion in Search Console; inclusion is the default, but child properties may inherit a parent setting. Check the property and inheritance that cover the site. Eligibility still does not guarantee crawling, indexing, serving, or citation.

Publisher work that improves the underlying page

The defensible work is strong Search practice for the page’s real audience. Keep it crawlable, internally discoverable, readable on mobile, and clear about its subject. Put essential information in accessible text, use useful headings, and make claims specific enough to verify.

Add value a generic answer cannot easily supply: a documented method, original analysis, relevant experience, or clear trade offs. A payroll software comparison, for example, should explain which operating assumptions shape its recommendations instead of repeating a feature list. Organize information coherently, but do not mistake short chunks or answer first formatting for a citation trick. See Rankwox’s guides to content chunking and answer first content.

A practical agency audit for an important page

A repeatable audit separates fixable access problems from unobservable selection decisions. First, record the exact URL, query, audience, locale, device, and date; similar queries can produce different results. Next, check Search Console and URL Inspection for indexing and the page Google received. Confirm that essential copy is crawlable and not hidden by robots directives, authentication, or rendering failures.

Then review the Search generative AI setting for the relevant property, including whether it inherits a parent choice. Read the page as a customer: does it answer the business question, explain its scope, and offer value beyond a generic overview? Improve weaknesses for readers rather than imitating a guessed model format. If an Overview appears, note its links and context, recheck over time, and report observations not guaranteed rankings or causal proof. This process makes “why were we not cited?”

actionable without pretending to know Google’s undisclosed selection mechanics.

Measure visibility without inventing attribution

Google says AI feature appearances are included in the Search Console Performance report under the Web search type. This helps evaluate overall Search performance, but it does not mean every impression or click is labeled with a particular Overview citation. Pair Search Console trends with analytics outcomes and annotate significant site changes; avoid claiming causation from one movement.

Separate technical eligibility, observed appearances, and business outcomes such as qualified visits or conversions. They are related, not interchangeable. Rankwox’s How AEO works hub links this topic to related guides, including vector embeddings and SEO.

The honest answer about source selection

Google has described the broad pattern Search retrieval, generated synthesis, and supporting links and says related searches may expand the material considered. It has not published a fixed formula for predicting which eligible URL will be linked for each query. No word count, markup, heading pattern, or “AI score” guarantees citation. Aim to build a useful page that Search can access, then measure carefully and qualify conclusions.