87.4% of all AI referral traffic flows through a single engine — ChatGPT — according to Conductor's 2026 AEO/GEO benchmark report. The remaining 12.6% fragments across Perplexity, Gemini, Claude, and Google AI Overviews, each rewarding a structurally different content shape. Across roughly 1,000 documented citation queries aggregated from Ahrefs, RivalHound, Discovered Labs, Seer Interactive, ZipTie, SE Ranking, and the GenPicked Research Team's Fitness Wearables Bradley-Terry study, a single coherent pattern emerges: brand visibility diverges 9x across the five engines, and the agencies still optimizing for one averaged "AI visibility score" are reporting on a metric that describes no actual engine.
This report distills nine findings from that aggregate dataset. Each finding pairs a claim with sourced evidence and a concrete agency implication. The structure is deliberate. In a market where 72% of SEO-investing brands receive zero AI citations according to AuthorityTech's 2026 crisis report, the strategic question is no longer whether to invest, but where the marginal hour of effort pays off per engine.
The headline numbers
Eight benchmarks anchor the rest of the report. Every figure here is cited inline and used downstream in at least one finding.
| Metric | Value | Source |
|---|---|---|
| AI share of total website traffic | 1.08% | Conductor |
| ChatGPT share of AI referral traffic | 87.4% | Conductor |
| Gemini YoY referral growth | 388% | Conductor |
| Google AI Overviews trigger rate | 48% | Ahrefs |
| Brand-mention correlation with AI visibility | 0.664 | RivalHound |
| Backlink correlation with AI visibility | 0.218 | RivalHound |
| SEO-optimized brands invisible to LLMs | 72% | AuthorityTech |
| AI Overview impact on position-1 organic CTR | -58% | Ahrefs |
Nine findings from the cross-engine dataset
Citation concentration is severe and inverts the long-tail thesis of classic SEO.
The top 5 domains (Wikipedia, YouTube, Google, Reddit, Amazon) command 38% of all AI citations, the top 10 control 54%, and the top 20 capture 66%. Profound's monitoring across 680M citations confirms the same concentration pattern at scale. In tech verticals, Wikipedia and Reddit dominate; in B2B enterprise software, G2, Gartner, and LinkedIn dominate. The same five sources recur across nearly every category.
The strategic move is to win on terrain where the top 5 have under-invested: niche vertical publications, Reddit communities, YouTube tutorials, and the client's own E-E-A-T-rich author pages. Concentration is a constraint that also defines where smaller publishers still have room to compete.
ChatGPT and Perplexity reward structurally opposite content shapes.
47.9% of ChatGPT citations come from Wikipedia-style encyclopedic sources, while 46.7% of Perplexity citations come from Reddit. The same client query can return wildly different brand sets on each engine, with almost zero overlap in source provenance.
Build two citation calendars per client. An encyclopedic lane (Wikipedia, trade publications, research firms) drives ChatGPT and Claude. A community lane (subreddit participation, user reviews, forum discussion) drives Perplexity. One unified brief under-performs on both.
Brand mentions correlate 3x stronger than backlinks with AI visibility.
Brand mentions correlate 0.664 with AI visibility; backlinks correlate 0.218 — roughly a 3:1 advantage for unlinked mentions, the inverse of how most SEO retainers are priced.
Agencies allocating 80/20 toward link-building over mention-building are optimizing for the 2018 Google graph, not the citation graph LLMs sample. Rebalance toward roughly 50/50 for any AI-first client portfolio. Treat mention-building as a primary AEO line item, not a PR afterthought.
Front-loading bias means buried answers go uncited.
44.2% of LLM citations are pulled from the first 30% of an article's text. According to AirOps, only 15% of pages ChatGPT retrieves appear in final answers — the other 85% is discarded silently. The retrieval funnel is far narrower than the indexing layer suggests.
Lead every pillar page with the direct answer in the first 200 words. Save narrative and competitive framing for the back half. If the strongest claim sits below the fold, the LLM retrieves, scans, and silently drops the page.
Self-contained chunks of 50-150 words receive 2.3x more citations than long-form prose.
Content built as 50-150 word self-contained chunks earns 2.3x more citations than unstructured long-form. Pages with FAQPage schema are 3.2x more likely to appear in Google AI Overviews — but only when the schema wraps actual chunked content. Schema without chunking is markup the LLM cannot exploit.
Rewrite pillar pages as 5-7 answer-first sections, each ~120 words, each with its own question H3. Apply FAQPage schema after the structure exists. The lift compounds: chunked structure plus schema plus front-loaded answers stacks all three citation levers on the same page.
Distribution multiplies citations more than craft does.
Content distributed across multiple publications increases AI citations by up to 325% versus single-site publishing. Every additional placement becomes a citation source independently, and LLMs treat cross-publication agreement as a confidence signal.
Once a piece is built well, syndication is the highest-leverage hour. Repurpose every cornerstone asset across the client blog, two or three vertical trade publications, and one Reddit post written natively for that audience. The compounding is roughly linear in distinct hosts.
E-E-A-T predicts AI citations 4.5x better than domain authority.
Domain authority alone correlates 0.18 with AI citation probability, while E-E-A-T signals correlate 0.81 — a 4.5x difference, not marginal. Yet most legacy SEO audits still grade on referring domains and citation flows, not on author bios, primary sources, and last-reviewed dates.
Run an E-E-A-T audit on every page before it ships: named author, bio with credentials, cited primary sources, first-person experience markers, and a last-reviewed date. The DA-and-backlinks audit is no longer load-bearing for AI surfaces.
llms.txt is a vendor distraction with zero documented citation lift.
SE Ranking's study of 300,000 domains found zero correlation between llms.txt presence and AI citations. The file appears on only 10.13% of domains surveyed. Neither Google nor OpenAI list it as a primary citation lever in their official documentation.
Add it defensively if you wish — it carries no risk — but do not let a vendor sell it as the program. The 80/20 work is chunked content, E-E-A-T signals, brand mentions, and per-engine distribution.
Cited brands recover disproportionately when AI Overviews appear.
When Google AI Overviews appear, position-1 organic CTR drops 58% from the historical 7.3% baseline. But Seer Interactive's analysis of 53 brands across 5.47M queries shows cited brands earn 35% more organic clicks per impression and 91% more paid clicks in those same overviews. Fast Company's coverage of the same Seer dataset highlights the same divergence.
Citation appearance is now as material to revenue as ranking position. The brands losing the 58% of clicks are missing from the overview. Cited brands recover; non-cited brands stay exposed.
The per-engine cross-tab
Across the aggregate dataset, each engine rewards a distinct primary signal. The table below maps the dominant lever for each, with the supporting citation. Use it as the master key when splitting a content brief across engines.
| Engine | Dominant signal | Documented benchmark | Primary agency lever |
|---|---|---|---|
| ChatGPT | Encyclopedic authority | 47.9% Wikipedia-source share (Discovered Labs) | Tier-1 publication brand mentions; entity-rich author pages |
| Perplexity | Community voice | 46.7% Reddit-source share (Discovered Labs) | Subreddit participation; user-review surface area |
| Google AI Overviews | E-E-A-T plus schema | 0.81 E-E-A-T correlation (ZipTie); 3.2x FAQ schema lift (Frase) | FAQPage schema on chunked content; author attribution |
| Gemini | Multi-modal, video-weighted | YouTube share now 9.51%, up 34% in six months (Ahrefs) | YouTube tutorials; transcript-rich pages |
| Claude | Clean authorship, entity density | 97.3% brand-mention rate per answer (Profound) | Cited-source content; named-author bylines |
Where traditional SEO investments stop transferring
72% of SEO-investing brands receive zero AI citations. The cause is structural, not tactical. Traditional SEO optimizes for Google's link-and-authority graph. LLMs sample a different graph — one built from entity density, experience-based content, self-contained question-answer chunks, and front-loaded answers. Optimizing for one graph often directly suppresses performance on the other.
Only 38% of pages cited in AI Overviews rank in Google's top 10 for the same query; 31% rank beyond position 100. A client can be invisible on Google and cited on ChatGPT, or rank #1 on Google and be missing from Perplexity, Gemini, and Claude simultaneously. The two systems are no longer correlated tightly enough for one to proxy the other in a monthly report.
The revenue translation
Seer Interactive tracked 53 brands across 5.47M queries representing 2.43B impressions. When AI Overviews appear, paid CTR crashes 68% (from 19.7% to 6.34%) — but cited brands earn 91% more paid clicks than non-cited brands in the same overviews. The spread is extreme. The channel is either protective or actively punishing, with very little middle ground for brands that ignore it.
On the lead side, HubSpot's internal AEO program reported a 1,850% increase in qualified leads sourced from AI. Conductor's 2026 survey of 250+ enterprise executives shows 56% of CMOs made significant AEO investment last year and 94% plan to increase next year. AI-sourced visitors who do click through spend 68% more time on the website than organic search visitors — a quality premium most agency dashboards still are not surfacing.
Five operational shifts for the agency report
Shift 1 — From rank tracking to citation tracking. Monthly reports should show citations gained, queries newly cited, engines where citation appeared, and the organic and paid traffic delta. The rank-tracking dashboard omits the variable that explains the revenue.
Shift 2 — Per-engine content calendars. One for the encyclopedic / brand-mention lane (ChatGPT and Claude), one for the community lane (Perplexity), one for the schema and E-E-A-T lane (Google AI Overviews), and a video layer (Gemini). Folding them into one editorial plan dilutes every signal.
Shift 3 — Brand mention as a tracked KPI. Treat mentions in tier-1 trade publications, Reddit threads, and YouTube transcripts as a monthly target with explicit quotas — for example: 3 tier-1, 10 tier-2, 50 community-tier mentions per quarter. Reporting them is the closest agency proxy to the 0.664 correlation that actually moves AI visibility.
Shift 4 — E-E-A-T audit before content ships. Author attribution, source citations, data points, first-person experience markers, last-reviewed date. The pre-publish checklist is now load-bearing. E-E-A-T predicts AI visibility 4.5x better than DA, which means the audit cannot remain an optional step for the editor.
Shift 5 — Citation lift as the retainer narrative. If the dashboard shows 2x citation growth month-over-month and a corresponding lift in cited-brand CTR, the retainer defends itself on data. Conductor's 2026 survey shows late-mover agencies face compounding competitive pressure as {n("94%")} of CMOs plan to increase AEO investment. The window for late entrants is narrowing each quarter.
Tooling landscape
Profound raised $96M at a $1B valuation in February with more than 10% of the Fortune 500 monitoring through the platform, positioned at the enterprise tier. Conductor and Ahrefs publish regular benchmarks integrated into their existing SEO platforms. ZipTie focuses on E-E-A-T audit and schema recommendations for the tactical, SMB-friendly buyer. Am I Cited tracks passage-level citation patterns at granular content depth.
GenPicked covers the agency workflow end-to-end: daily citation tracking across the five engines, per-engine content optimization, competitive intelligence, and white-labeled monthly reports built for retainer defense. The layer that wins is the one closest to the per-engine evidence the report has to show every month.