← All posts

Google may run two different searches — and you optimize for one

· Updated Nov 15, 2026 · 5 min read · By Mikhail Kuzmitskii

View as Markdown ↗

Spoiler, for anyone who skims: AI Overview and AI Mode pick their sources differently — by one hypothesis, from two different engines — so a top-10 page can be everywhere in one and absent from the other.

One histogram in the Ashmanov & Partners slides stopped me: source positions in AI Overview looked like any SERP chart — a fat top-3 and a long tail — while the AI Mode chart was almost flat. Same button, two shapes! What produces that — two different retrieval systems, or something simpler?

Two searches behind one button

The boldest architectural hypothesis about Google in 2026: AI Overview (AIO) and AI Mode (AIM) aren’t one “AI search” in two skins but two different systems on different retrieval engines. Its authors (Ashmanov & Partners, empirics across 600+ parameters) say AIO assembles its answer from the classic results, while AIM searches a vector semantic index, the way ChatGPT does. RESONEO’s October 2026 schemas point the other way: one system, magi, writes both answers, and AI Mode gets Search’s ranking signals; Google’s own guide describes both as RAG over its core ranking systems.

Google may run two different searches — and you optimize for one

Practically, “one template for everything” doesn’t work: optimizing only for the head query covers AIO and leaves a site nearly nonexistent in AIM.

The data: how they differ

ParameterAI Overview (AIO)AI Mode (AIM)
Answer lengthshort (1–2 paragraphs)~4× longer
Sources from the original query~60%~15%
+ from extra sub-queries10–15%~30%
Top-3 sources~40% of slotsflatter distribution
Commercial pages among sources45.6%29.2%
Source languagesas in the results (monolingual)multilingual
Ambiguous wordsone meaningall meanings
Correlation with organicstrongerweaker

The most telling piece is the histogram shape. In AIO the source-by-position distribution looks like classic search: dense on the top-3, exponential decay. In AIM it’s flat — what you’d expect from an embedding-index search, but also from a wide fan-out whose sources come from sub-query results rather than the head query. Each row alone could have another explanation — and fan-out could explain all of them together.

The AIO strategy: the classics still work

  • Top-10 still supplies ~60% of sources — rankings remain a lever.
  • Query fan-out gives pages outside the top-3 a shot — via sub-queries.
  • Freshness wins: recently updated pages systematically take the slots.
  • The commercial share is high (45.6%) — commercial pages are allowed into AIO.

The AIM strategy: rank for the sub-questions

  • Top-10 for the head query supplies only ~15% of sources — but AI Mode answers get Search’s ranking signals, so position on each sub-query still decides (RESONEO).
  • Atomic blocks and “vector similarity” are unconfirmed for Google: its guide calls chunking, rewriting for AI and special schema.org markup unnecessary, AI Mode included. What holds is a direct answer with the number up front, in visible text.
  • Multilingual: AIM cites foreign sources — your content can surface in other locales (and competitors arrive from other languages).
  • The answer is 4× longer → more citation slots, but each slot is won by a separate sub-intent (see where engines get their answers).

Uncomfortable? Yes, but it probably also means #7 is no longer a life sentence!

What to do

  1. Split the KPIs. Presence in AIO and in AIM are two different results; measure them separately, like rankings.
  2. For AIO — keep the classics: rankings, freshness, sub-query fan coverage.
  3. For AIM — cover the sub-intents: rank for the sub-questions and answer each in a self-contained section, without chopping the page into chunks.
  4. Check both surfaces. Our AI-visibility probe measures citation across 4 engines, and geo-metrics separately counts your AI-Overview exposure per keyword.

My subjective take: two columns in reports will soon look obvious, whatever the architecture turns out to be. One histogram convinced me to measure the surfaces apart; perhaps it’ll convince you.

Sources

  • GEO-HowTo 2026 — Kurdyukova–Volovich talk (Ashmanov & Partners) — the two-engine hypothesis, a 600+-parameter methodology: source histograms, original-query shares (~60% vs ~15%), commercial pages (45.6% vs 29.2%), language behavior.
  • GEO-HowTo 2026 — Alexander Runov’s talk (MGCom) — Google Search 2.0 (2026-05-19): “AI became the foundation of the search interface.”
  • US v. Google case materials (2025) — FastSearch as a separate retrieval system (embeddings, dot-product) — the architectural basis of the hypothesis.
  • RESONEO, “Understanding Google”, part 4 (October 2026) — schemas showing one system (magi) writing AI Overviews and AI Mode answers, with Search’s ranking signals.
  • Google Search Central, “Optimizing your website for generative AI features on Google Search” (May 2026) — both features as RAG over core ranking systems; no special schema.org markup needed.

FAQ

How is AI Overview different from AI Mode?

AI Overview is a short summary above the classic results, assembled mostly from them (~60% of sources from the original query). AI Mode is a separate conversational surface with an answer ~4× longer and only ~15% of sources from the original query — the rest come via sub-queries.

Is it true that AIO and AIM run on different engines?

It's a hypothesis by Ashmanov & Partners, empirically grounded on 600+ parameters: the shape of source histograms, language behavior, the share of commercial pages (45.6% in AIO vs 29.2% in AIM), and the handling of ambiguous words. Court records and leaks show no explicit split, Google's 2026 guide describes both as retrieving from the Search index with its core ranking systems, and RESONEO's October 2026 schemas show one system (magi) writing both answers. A wide fan-out can explain the differences without a separate vector engine.

How do you optimize for each?

For AIO classic SEO works: top-10 supplies ~60% of sources, plus freshness and query fan-out. For AIM position on the head query matters less, but position on the sub-queries does: per RESONEO's schemas, AI Mode answers receive Search's ranking signals. Cover the sub-intents with direct answers in visible text; Google says no special schema.org markup is needed.

Was this helpful?
Launch SEO/GEO with seo·matrix → or a request without Telegram →

Comments

  • …

Comments are moderated.