From SGE to AI Overviews: What Changed

ClickRadius Institute · Published May 26, 2026

In three years, Google moved from the opt-in SGE experiment to production AI Overviews and a separate conversational AI Mode surface. Those names refer to different products and release stages, not one universally deployed interface. This account separates what Google announced from what a dated, scoped observation would still need to measure.

Act I — SGE: the experiment (May 2023 – May 2024)

Google previewed the Search Generative Experience at I/O in May 2023: an opt-in Search Labs feature that generated an AI summary above the classic results, with source links attached. SGE was deliberately hedged — users had to enroll, coverage was narrow, answers were slow and verbose, and Google labeled the whole thing experimental.

But the experiment settled three questions that defined everything after:

• Format: the answer sits above the results, making the generated block — not position #1 — the top of the page;
• Sourcing: the answer cites a small set of sources, chosen by the system, not strictly the top-ranked pages;
• Intent: Google was willing to answer the query itself rather than merely referee links — a reversal of a 25-year-old contract.

The SGE year is when the discipline now called GEO got its academic foundation: the Princeton-led paper “GEO: Generative Engine Optimization” (released late 2023, presented at KDD 2024) tested content interventions and found that statistics, attributed quotations, and credible source citations changed the selected source's relative visibility/impression — up to 40% in the authors’ reporting — while keyword stuffing did little. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.

Act II — AI Overviews: the rollout (May 2024 – early 2026)

At I/O 2024, SGE graduated: renamed AI Overviews and shipped to U.S. users by default, no Labs enrollment required. Expansion to more than 100 countries followed within months. The differences from SGE were substantive, not cosmetic:

• Selectivity. Overviews stopped trying to answer everything. They appeared where generation added value — informational, multi-step, and question-shaped queries — and stayed away from many navigational and transactional ones. Early rough edges (including some widely publicized wrong answers in the first weeks) pushed Google toward more conservative triggering and heavier grounding.
• Speed and compression. Answers got shorter, faster, and more confident, with citations presented in evolving formats — link cards, inline attributions, sidebars.
• Scale. AI Overview coverage varies by measurement source, query cohort and class, collection date, locale, device, account state, and feature definition. Meanwhile Google began testing AI Mode as a conversational Search surface in Labs during 2025.

The click economics changed measurably in this act. Zero-click behavior varies by feature, measurement source, query cohort, collection date, locale, and device. This page does not assert a universal current rate. Ranking #1 under an Overview that resolved the query became the emblematic frustration of the period: technically winning a contest whose prize had moved.

Act III — AI Mode and a redesigned Search interface (May 2026)

At Google I/O on May 19, 2026, Google described a redesigned Search experience and made Gemini 3.5 Flash the default model inside AI Mode. It did not establish AI Mode as the global default or formally demote ordinary results; normal SEO foundations remain relevant. Google’s leadership described the release in these terms:

This is the biggest upgrade to our Search box in over 25 years.— Elizabeth Reid, VP of Search, Google (Google I/O 2026)

It’s our biggest upgrade to Search ever.— Sundar Pichai, CEO, Google (Google I/O 2026)

AI Overview coverage still varies by measurement source, query cohort, collection date, locale, device, account state, and feature definition; this page does not assert a universal current percentage. Google also announced that Information Agents would launch first for AI Pro and Ultra subscribers during summer 2026. Announcement, account eligibility, observed availability, and an actual agent result remain separate evidence states.

The practical measurement consequence is narrower: ordinary rankings, AI Overview appearances, AI Mode answers, citations, clicks, and conversions are different events. One cannot be substituted for another, and an announcement does not establish which result surface every user receives.

What actually changed, era by era

What stayed the same — the through-line for businesses

Across all three acts, the selection logic has rewarded the same fundamentals, which is the most strategically useful fact in this whole story:

1. Retrievability. Every era pulled candidates from Google’s index. Blocked crawlers, rendering failures, and thin indexation have been disqualifying since day one.
2. Public readability. Test whether important content is present in the public render and organized for the intended reader. Do not infer a cross-version extraction preference.
3. Evidence quality. Statistics, quotations, and citations should be accurate, attributable, and scoped. The KDD 2024 benchmark does not establish that later Google grounding became stricter or selected a page for those features.
4. Identity records. Audit first-party identity facts, directories, and third-party sources independently. Their consistency does not reveal an AI Mode trust score or deeper-reasoning weight.

Do not infer commercial stakes from an unversioned 48% figure, a claimed global default, an announced Information Agents rollout, or an unsupported population-wide zero-mention rate. Measure the actual Google surface and business population in question, and keep ordinary monthly AI Presence separate from unobserved agent delivery.

How each era rewrote the practitioner’s job

The three stages changed what practitioners could observe. Tracing that evolution shows why a current record must identify the exact Google surface rather than treating SGE, AI Overviews, and AI Mode as interchangeable.

The SGE era created the observer’s job

During the experiment, instrumentation meant recording prompt sets, returned answers, cited sources, and conventional rankings separately. A retained baseline makes later product versions comparable; a recollection does not.

The Overviews era created the dual-scoreboard job

From May 2024 onward, a ranking and an Overview citation became two measurements that should be recorded side by side. The KDD 2024 GEO benchmark reported source-visibility changes for several content interventions in its own experimental setting; it did not prove that those interventions cause citation selection in later Google products. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.

The AI Mode era creates the entity manager’s job

For a conversational surface, teams should keep first-party facts, off-site records, and observed answers as separate evidence planes. Consistency can be audited, but it does not reveal a private trust rule or prove selection.

The lesson across all three stages is to version the evidence. A preview, announcement, eligible account, observed answer, citation, and downstream visit each prove different things.

The tactical checklist for the AI Mode era

• Audit robots.txt and rendering for Google’s crawlers — including Google-Extended policy decisions made years ago and forgotten.
• Restructure priority pages answer-first, one canonical answer per question, with Article/FAQPage/Organization JSON-LD.
• Enrich those pages with the evidence triad: real statistics, attributed quotes, credible external sources.
• Reconcile your entity across the web — name, category, location, facts — until every source a triangulating engine checks agrees.
• Record citations with a fixed prompt set and schedule. ClickRadius reports a separate monthly AI Presence sample across ChatGPT, Gemini, Perplexity, Claude, and Grok and estimates five readiness-evidence families at the submitted URL; neither result stands in for a direct AI Mode observation.

Frequently asked questions

Is SGE the same thing as AI Overviews?

SGE (Search Generative Experience) was the experimental, opt-in version Google ran in Search Labs starting in May 2023. Google launched AI Overviews to U.S. users in May 2024 and later expanded availability. Current coverage, triggering, and citation presentation still require a dated observation for the relevant locale, account, device, and query.

What is the difference between AI Overviews and AI Mode?

AI Overviews can appear within the conventional results experience, while AI Mode is a conversational search surface. Google announcements and observed availability must be versioned by locale, account, and date. Google's May 19, 2026 announcement described a redesigned Search experience and made Gemini 3.5 Flash the default model inside AI Mode. It did not establish AI Mode as the global default or formally demote ordinary results; normal SEO foundations remain relevant.

How do I measure whether my business is cited?

Record the exact Google surface, query, locale, account state where relevant, time, returned answer, and cited URL. Accessible pages, attributable evidence, valid structured data, and accurate off-site records can be audited separately, but none guarantees selection. ClickRadius's five-engine monthly AI Presence panel must not be presented as direct AI Mode coverage.

AI Readiness and observed citation are different questions. Get a free AI Readiness Score for five setup families observed at the submitted URL, or see ClickRadius pricing for separate monthly AI Presence samples across five supported engines. The score cannot determine whether Google will cite the site.