Citation page · AI search data · 2026

State of AI Search 2026

Q: What share of search is AI search in 2026?

Per Datos/SparkToro clickstream analysis (US panel, search-like queries as a share of total search volume), Google AI Overviews is roughly 93% of AI-search-like volume but declining about two points; Bing/Copilot is around 3% and growing; ChatGPT is about 0.25% of search-like volume but its referral traffic is up roughly 200% year over year; Gemini is under 0.5%, and Perplexity and Claude are each under 0.1%, all growing. The headline: Google still owns the volume, but the assistant surfaces are growing far faster than their share suggests, and referral growth — not share — is the leading indicator.

Q: Does blocking Google-Extended remove me from Google AI Overviews?

No. Google AI Overviews is built from the standard Google index — the pages Googlebot crawls — not from Google-Extended. Google-Extended is a separate robots.txt control that only governs whether your content is used for Gemini API and Vertex AI grounding and model training. You can block Google-Extended and still appear in AI Overviews; conversely, blocking Googlebot or setting noindex is what actually removes you from AI Overviews. Since AI Overviews is ~93% of AI-search volume, Googlebot access is the single highest-reach AI-search signal there is.

Q: What is the difference between branded and non-branded AI search?

They are different games with different metrics. Non-branded queries ("best wireless speaker for a medium living room") are a visibility game — the goal is to show up in the answer at all, and the metric is visibility and position. Branded queries ("is the new Sonos Era worth it if I own a Sonos One?") are an accuracy game — you are already in the answer, so the goal is to show up well, and the metric is sentiment and accuracy. Non-branded is owned by SEO/AEO and content; branded is owned by product marketing and PR, who seed accurate facts and correct the record when a model gets your pricing, positioning, or feature set wrong.

Q: Why do models answer from training data instead of searching the web?

Because it is cheaper and faster. A model will answer from its training data whenever it is confident it can, and only reach for web search on queries that are fresh, long-tail, or specific enough that the training data is stale or thin. Two consequences for AEO: first, being present in the training corpus matters (broad, long-lived, frequently-referenced content), and second, on the queries that do trigger web search, recency decides who gets pulled in — a stale page loses the retrieval path even if it is comprehensive. The long tail of queries that trigger live search is getting longer, which is good news for specific, well-structured pages.

Q: Do AI agents read web pages the same way browsers do?

No, and this is the most common technical AEO failure. Many agents do not execute JavaScript — they read the raw HTML the server returns. If your content, headings, or schema are injected client-side (a single-page app that renders into an empty shell), the agent sees an empty page and your content is effectively invisible to it. Agents also increasingly prefer structured formats like markdown over HTML, which is why content negotiation (serving a clean markdown alternate) is an emerging signal. The rule of thumb: good AEO is good SEO plus agent-native architecture — server-render the content, and make it parseable without a browser.

Q: How do I measure AI search visibility when there is no Search Console for AI?

There is no Google-Search-Console equivalent for AI, so you measure it yourself across three surfaces. (1) Chat interfaces (chatgpt.com, claude.ai, perplexity.ai) — prompt them with your priority queries and log whether you are cited and how you are described; remember these are UI-driven and may use memory. (2) Model APIs (raw model responses) — query the API directly to see the un-personalized answer. (3) Agents (Operator, Perplexity Buy, Amazon Rufus) — which carry shopper-aware context. Track citation count and sentiment by query × surface × week. Vendor tools (Profound, SE Ranking, Otterly) automate the chat surface; the free StackSwap AEO Audit checks whether your pages are technically citable in the first place.

Q: Is AEO just SEO with a new name?

No, but good SEO is the foundation. AEO keeps the SEO fundamentals — crawlable, indexable, well-structured, authoritative content — and adds an agent-native layer: server-rendered HTML (because agents may not run JavaScript), chunk-level passages that answer one question in the first two sentences, machine-readable entity and FAQ schema so a model can attribute facts to your brand instead of "a website," recency signals, and explicit crawler access for AI user-agents. The optimization target also changes: SEO optimizes for a ranked list of blue links; AEO optimizes to be the one or two sources quoted inside a synthesized answer.

Q: Where should a bootstrapped team focus first?

Start with the highest-reach, lowest-effort signal: confirm Googlebot can crawl you and you are not accidentally noindexed, since AI Overviews (~93% of AI-search volume) depends on it. Then make sure your core content is server-rendered, not JavaScript-injected. Then add entity and FAQ schema so models can attribute facts to your brand. Only after those fundamentals are in place does it pay to chase the assistant surfaces (ChatGPT, Perplexity) with operator-voiced, recently-updated, chunk-level content. The free StackSwap AEO Audit grades all of this on any URL in about a minute.

By Nick French · Founder, StackSwap · Updated June 2026

The AI search data landscape is blurry and changing fast, so here is the operator version. Google AI Overviews is still roughly 93% of AI-search-like volume — and it is fed by the standard Google index via Googlebot, not Google-Extended. ChatGPT is a rounding error on share (~0.25%) but its referral traffic is up about 200% year over year, so share badly understates trajectory. The data you get back is fragmented, there is no Search Console for AI, and the optimization work splits into two different games — branded and non-branded. This page is the data, the surfaces, and what actually earns a citation.

How big is AI search, really? The 2026 numbers

Here is the share-of-search picture as of 2026. The single most important column is the last one — data availability — because it tells you how much you can even measure on each surface. Google has the volume but gives you the least visibility into the AI slice; Perplexity and Bing are tiny but hand you analytics; Gemini and Claude give you nothing today.

Provider	Search share	Growth trend	Data availability
Google AI Overviews	~93%	Declining (~−2pp)	Limited — no AI-only filter on Search data
Bing / Copilot	~3%	Growing	Basic analytics
ChatGPT	~0.25%	+200% referrals	Referral traffic (utm_source)
Gemini	<0.5%	Growing	None
Perplexity	<0.1%	Growing	Analytics
Claude	<0.1%	Growing	None

Source: Datos / SparkToro clickstream analysis (US panel), search-like queries as a share of total search volume, as presented in Vercel's 2026 AEO analysis. “Growth” is year-over-year referral/traffic growth, not share growth. Figures are directional, not StackSwap first-party data.

Why Google AI Overviews is not the same as Google-Extended

This is the distinction most AEO advice gets wrong, and it is worth being precise about because it is ~93% of the volume at stake. Google AI Overviews is generated from the standard Google index — the pages Googlebot crawls and indexes. It is not gated behind Google-Extended.

Googlebot controls whether you are in the index that feeds Search and AI Overviews. Block it (or set noindex) and you are out of ~93% of AI search.
Google-Extended is a separate robots.txt token that only governs whether your content trains and grounds Gemini (the API and Vertex AI). Blocking it does not remove you from AI Overviews, and allowing it does not get you in.

The practical takeaway: if you have to pick one AI-search access signal to get right, it is Googlebot crawlability and indexability — not your Google-Extended policy. The free AEO Audit checks both, separately, so you do not conflate them.

Branded and non-branded AI search are different games

Treating “AI search visibility” as one number is the second big mistake. There are two games, with different metrics and different owners. On a non-branded query you are fighting to appear at all; on a branded query you already appear, and the fight is whether the model describes you accurately and favorably.

	Branded	Non-branded
The question	"I have an old Sonos One. Is the new Era model worth it?"	"What's the best wireless speaker for a medium-sized living room?"
Key metric	Sentiment & accuracy — is what the model says about you right and favorable?	Visibility & position — do you appear in the answer at all, and where?
The goal	Show up well	Show up
Who owns it	Product marketing + PR (correct the record, seed accurate facts)	SEO / AEO + content (earn the citation on the category query)

Why it matters operationally: the branded game is mostly a PR and product marketing job — seed accurate facts in the corpus, correct the record when a model quotes the wrong price or positioning, and publish the canonical “what we are” page. The non-branded game is a content and AEO job — earn the citation on the category query with chunk-level, well-structured, recently-updated pages. One team optimizing both with one metric will do neither well.

The response is just the surface: training data vs web search

When a model answers, it is drawing from one of two places, and the difference decides how you get cited. It is cheaper for a model to answer from its training data, so it does that whenever it is confident — but training runs happen at intervals, so that data is stale. It only reaches for web search on queries that are fresh, long-tail, or specific enough that training data falls short.

For the training-data path: presence in broad, long-lived, frequently-referenced content is what gets you into the model's default answer. This compounds slowly and is hard to game.
For the web-search path: models issue far more search queries than humans do, and the long tail of those queries is getting longer. Recency decides who gets pulled in — a stale page loses even if it is more complete. This is why a visible, machine-readable last-updated date is an AEO signal, not a nicety.

Agents don't read like humans: good AEO is good SEO plus agent-native architecture

The most common technical AEO failure is invisible content. Many agents do not execute JavaScript — they read the raw HTML your server returns. If your page renders client-side into an empty shell, an agent sees an empty page and your content is, in the literal sense, invisible to it. Two corollaries:

Server-render the content. SSR/SSG the headings, body, and schema so the raw HTML carries the answer. If a crawler that does not run JS would find an empty page, so will the agent.
Prefer structured formats. Agents lean toward markdown over HTML. Serving a clean markdown alternate via content negotiation is an emerging, agent-native signal — optional today, increasingly expected.

Good AEO is good SEO — crawlable, indexable, structured, authoritative — plus an agent-native layer on top. The AEO Audit fetches your page the way a JS-less agent does and tells you, in words, what it can actually read.

There's no Search Console for AI, so measure three surfaces

There is no Google-Search-Console equivalent for AI search. Until there is, you ask the questions yourself, across the three surfaces where buyers actually encounter you:

Surface	Example	What to watch
Chat interfaces	chatgpt.com, claude.ai, perplexity.ai	UI-driven and may use memory — are you cited, and how are you described?
Model APIs	Raw model responses	The un-personalized answer — the cleanest baseline to track over time.
Agents	Operator, Perplexity Buy, Amazon Rufus	Shopper-aware context — do you survive an agent that is acting, not just answering?

Log citation count and sentiment by query × surface × week. Vendor tools (Profound, SE Ranking, Otterly) automate the chat surface; for the technical question of whether your pages are even citable, the free AEO Audit is the fastest check.

FAQ