Being selected inside the answer replaced earning the visit: AI Overviews halve position-one clicks while AI citations churn and checkout moves into chat
Asked (summary):
SERP & AEO/GEO landscape, Q3 2026 — how answer surfaces suppress outbound clicks, how volatile AI visibility is, how commerce is moving inside answer engines, and what operating model follows for Q4.
The evidence base is 25 measured research findings on SERP clicks, citations, visibility and referrals published through 30 September 2026 across 13 independent hosts, plus 35 first-party Q3 announcements from OpenAI, Cloudflare and Perplexity (deduplicated here to 19 distinct moves). Some findings are experimental, some observational; single-site, sector-specific and preprint evidence is labelled as such throughout.
The quarter's core tension in one view
Left: what an AI Overview does to clicks on the same Google US results page. Right: how little of July's AI-cited domains survived to September — the surface you win is unstable.
without AI Overviewwith AI Overview — hover a pair for source and sample
share of July's cited domains still cited in September (higher = more stable). 7,949 prompts, 39,405 answer pairs, US English
A randomized US field experiment (preprint, 1,100 participants) found Google AI Mode cut click-through to external sites by 18.8 percentage points versus default search, while removing AI raised it 8.8 points — arxiv.org
In Semrush's manufacturing vertical, AI Overviews covered 57% of tracked search volume in July (up from 38% in January), yet AI assistants drove only 0.48% of site sessions, against nearly 80% from direct plus organic — semrush.com
Mention leaders and citation leaders diverge: of the top-15 manufacturing domains by AI mentions and by AI citations, only two (Vevor and Grainger) appear on both lists — semrush.com; separately, 25.4% of English AI citations led to pages with a disclosed commercial interest across 106,758 citations — ppc.land
A single-site B2B Search Console study (906,663 non-brand impressions over 90 days) measured 0.849% CTR at positions 1–2 — far below historical benchmark curves near 25% — and 96.5% of non-brand queries produced zero clicks — coworker.ai
Q3 platform moves: discovery, retrieval and control were all rebuilt
19 distinct first-party announcements, July–September 2026. OpenAI pushed commerce and ads into the answer; Perplexity replaced its retrieval stack; Cloudflare split publisher controls.
Operating model for Q4: three scorecards, not one
Scorecard 1
Classic SEO outcomes
Rank and impressions by query class
CTR by position, split SERPs with vs without an AI Overview
Non-brand clicks and zero-click share
Scorecard 2
Answer-engine outcomes
Mention share and citation share per engine
Source inclusion and sentiment in answers
Cited landing pages and monthly citation churn
Scorecard 3
Commercial outcomes
Qualified sessions and assisted conversions
Product-feed eligibility (merchant feeds, ACP)
In-answer checkout participation
Operate them with: prompt panels by intent and engine, re-run monthly to measure citation churn (retention ranged 14.8–64.6% across engines in Q3); content, entity and feed hygiene so products and pages are machine-legible; original evidence and data that answers must cite; and a crawler-policy review now that search, agent and training access are separately controllable.
Evidence table: all 25 research rows
Date
Study
Finding
Sample / scope
Evidence type
Source
Method: 25 measured research findings (SERP clicks, citations, AI visibility, referrals) published through 2026-09-30 across 13 hosts, no host supplying more than half, plus 35 first-party Q3 2026 announcement rows from OpenAI, Cloudflare and Perplexity deduplicated to 19 distinct moves; assembled 2026-10-06. Values are study-reported percentages, rates and counts under each study's own denominator; incompatible studies are shown side by side, never averaged. Observational results are associations, not causes; single-site, sector-specific and preprint rows are labelled. Cut for space: full metric definitions and comparison baselines (see linked sources) and 16 duplicate announcement extracts.