AEO field data · 2026
State of AI Search 2026
The short answer: ~93% of what people call "AI search" is still Google AI Overviews, and it runs off the normal Google index via Googlebot — not Google-Extended. ChatGPT is about 0.25% of volume but the fastest-growing referral source on the board. The assistants are the growth; Google is the rent. Share figures below are third-party clickstream estimates from Datos / SparkToro, not StackSwap data.
By Nick French, Founder of StackSwap · Published June 9, 2026
The 2026 AI-search landscape, by volume
Ranked by who actually has the search volume in 2026. The headline is the concentration: one surface is ~93% of it, and it is the one most "AEO tools" measure least well.
| Surface | Share of AI search | Trend | Data you get |
|---|---|---|---|
| Google AI Overviews | ~93% | Slowly declining (~2pp) | Limited — runs off the standard Google index |
| Bing / Copilot | ~3% | Growing | Basic analytics |
| ChatGPT | ~0.25% | Referrals up ~200% YoY | utm_source on referrals |
| Gemini | <0.5% | Growing | No referral data |
| Perplexity | <0.1% | Growing | Has analytics |
| Claude | <0.1% | Growing | No referral data |
Source: Datos / SparkToro clickstream analysis, 2026. Figures are directional estimates of AI-search-like volume, not first-party StackSwap measurements.
Why Google-Extended is not your AI Overviews switch
Google-Extended is the robots.txt toggle everyone debates. It only controls whether Google can train and ground Gemini on your content. It has no effect on whether you appear in AI Overviews. AI Overviews is built from the standard Google index — Googlebot. Indexed means eligible; block Googlebot or set noindex and you are out of ~93% of AI search. If an "AEO tool" flags Google-Extended as your AI Overviews setting, it is measuring the wrong toggle.
Most agents do not run JavaScript
GPTBot, ClaudeBot and PerplexityBot fetch raw HTML and do not execute your JavaScript. If your page renders client-side, the agent gets an empty shell and your copy is invisible — no matter how good it is. Server-render the content, or expose an agent-readable alternative. This is the single most common AEO failure, and most teams have never checked for it.
Branded and non-branded are two different jobs
"What is the best tool for X?" is a visibility game. "Is your product worth it for X?" is an accuracy game. They have different metrics and different owners, and scoring them as one number hides both. Separate the scoreboards before you try to move either.
The seven boring fixes that actually move AEO
In the order that moves the needle. None of it is exotic — it is plumbing, and every "AEO content framework" is downstream of getting these right.
- Unblock Googlebot. AI Overviews is ~93% of AI search and runs off the normal Google index — not Google-Extended. Everyone argues about the Google-Extended toggle; it only governs whether Google can train and ground Gemini on your content. Googlebot is the actual lever: block it (or set noindex) and you opt out of the 93%.
- Server-render the page. GPTBot, ClaudeBot and PerplexityBot read raw HTML and do not run your JavaScript. A client-rendered React app hands them a blank shell. Server-render the headings, body and schema so a crawler can actually read the words — this is the highest-leverage fix after crawler access.
- Find the one stray Disallow. A CDN default or a leftover line in robots.txt blocks the crawler at the door. No amount of schema saves a page a bot cannot fetch. Read robots.txt line by line and confirm a clean 200 response.
- Add two schema blocks: Organization and FAQPage. That is the difference between "according to StackSwap" and "according to a website." Organization gives an engine an entity to attribute facts to; FAQPage is the format engines lift most often. Get these two solid before touching the other ten types.
- Lead every section with the answer. Models lift passages, not pages. Put the answer in the first two sentences of each H2, then expand. The inverted pyramid is what gets a section pulled into a synthesized answer.
- Put a real last-updated date on it. On the web-search path, models reach for fresh sources. An undated page loses to a worse one that carries a machine-readable dateModified. Bump it when you actually update the page, and keep it honest.
- Ship an llms.txt and a clean sitemap. A one-file markdown map to your best pages, plus a sitemap declared in robots.txt. Boring, fast, and increasingly expected by the crawlers that assemble answers.
There is no Search Console for AI search
The data is fragmented, and Google gives you the least visibility into the slice that is 93% of the market. So you measure it yourself: prompt the chat surfaces, hit the APIs where they exist, watch what the agents fetch, and log who is cited by query, surface and week. The boring 93% (Google) pays the rent; the assistants are where you plant flags for next year. Optimize for the slope, not just the level.