Skip to content
Primores AI wiki

What 1 Billion Data Points Say About AI Search — Ahrefs' 2026 Studies, Calibrated

What 1 Billion Data Points Say About AI Search (Ahrefs 2026, Calibrated)

TL;DR: Across 14 studies and 1B+ data points in 2026, Ahrefs measured how AI search actually works. Six findings matter: (1) AI surfaces are separate discovery layers — Google AI Mode and AI Overviews reach the same conclusion 86% of the time but share only 13.7% of citations, and only 38% of AI Overview citations come from Google’s top 10; (2) the traffic collapse is accelerating — position-1 CTR drops 58% with an AI Overview, up from 34.5% eight months earlier; (3) the strongest visibility correlate is YouTube mentions (0.737), far above Domain Rating (~0.3) or backlinks (weak), with content volume near-zero (0.19); (4) schema markup has no measurable effect on AI citations (the one quasi-experimental result here); (5) “Best X” listicles are 43.8% of ChatGPT’s cited page types; (6) AI answers are volatile on the surface (change every 2.15 days) but stable in meaning (0.95 semantic similarity). Calibration: these are single-vendor, mostly correlational studies — strong large-N measurement, not peer-reviewed. Treat as Tier 2 evidence; the schema null is the lone near-causal finding. See seo/zero-click-strategy § calibration for the cross-vendor + institutional anchoring.

Why this page exists — and how to trust it

In the first half of 2026 Ahrefs published a connected run of studies on AI search, each built on its Brand Radar index, Google Search Console aggregates, or Web Analytics panel. Together they’re the largest public measurement of AI-search behavior available. They’re also all from one vendor that sells SEO tools — so this page exists to synthesize the findings and grade them, not to repeat them uncritically.

Evidence tier (load-bearing — read before citing any number below):

This cluster
SourceSingle vendor (Ahrefs), self-published
ScaleReal, disclosed, large — millions of SERPs/URLs/prompts per study
MethodMostly observational / correlational (Brand Radar, GSC, Web Analytics)
ExceptionThe schema study is quasi-experimental (difference-in-differences, 1,885 treated + 4,000 matched controls) — the one near-causal result
Peer reviewNone
TierTier 2 — single-vendor large-N measurement. Above “vendor-blog estimate” (real n + disclosed method); below peer-reviewed/institutional.

The correlations study says it plainly: “correlation isn’t causation… that doesn’t mean improving these metrics will automatically boost your AI visibility.” Carry that caveat into every correlation below. Where an institutional primary source corroborates the direction (e.g. the Pew Research clickstream study on the CTR collapse), that’s noted — and lives in full at seo/zero-click-strategy § “How solid are these numbers?”.

The six findings

1. AI surfaces are separate discovery layers — ranking ≠ being cited

The most strategically important cluster of results. The old model (“rank #1 → get cited by AI”) does not hold:

  • Google AI Mode vs AI Overviews: for the same query they reach a semantically similar answer 86% of the time (89.7% of pairs scored >0.8 similarity) — but their citations overlap only 13.7% (16.3% for the top 3). Word-level overlap is 16%. (n=730K query pairs for content, 540K for citations; Sept 2025 US.)
  • AI Overviews barely follow the rankings: only 38% of AIO citations come from pages in Google’s top 10; 31.2% from positions 11–100; 31% from outside the top 100. (n=4M AIO URLs / 863K SERPs.)
  • ChatGPT pulls from a parallel index: 28.3% of ChatGPT’s most-cited pages rank for zero Google keywords — cited repeatedly despite no traditional search visibility. (n=top 1,000 cited pages, Sept 2025.)

Implication: each AI surface is its own optimization target. Visibility in ChatGPT does not imply visibility in AI Mode or AIO, and a #1 Google ranking guarantees neither. This is the deepest break from classic SEO in the whole cluster.

2. The CTR collapse is real and accelerating

  • An AI Overview cuts position-1 organic CTR by 58% as of December 2025 — up from 34.5% in April 2025, an eight-month jump. (n=300K keywords, 150K with AIO + 150K control, GSC, Dec 2023 vs Dec 2025.)
  • Yet of the clicks that remain, 96.98% still land in the top 10 (97.56% mobile; never below 95.5% over two years). The top-10 game still matters — for a shrinking pie. (GSC, billions of clicks, US, Aug 2025.)
  • ChatGPT is ~12% of Google’s query volume (2.5B prompts/day vs 13.7B searches) but sends 190× less referral traffic, with a CTR ~96% lower. (Ahrefs Web Analytics, 76K sites; classification disputed — author calls 65% of ChatGPT use search-intent vs OpenAI/Harvard’s 24%.)

Implication: confirms the seo/zero-click-strategy thesis with better-anchored numbers. The CTR direction is independently corroborated by the Pew Research Center clickstream study (8% vs 15% click rate, ~47% reduction) — the institutional anchor for the vendor “58%.“

3. What correlates with AI visibility is not classic SEO

Spearman correlations between brand metrics and AI visibility, ranked (n=75K brands, DR>40, Brand Radar):

FactorCorrelation with AI visibility
YouTube mentions0.737 (strongest of all)
YouTube mention impressions0.717
Branded web mentions0.656–0.709
Branded anchors0.511–0.628
Branded search volume0.352–0.466
Branded traffic0.235–0.357
Domain Rating (DR)0.266–0.326
Backlinks / URL Ratingweak
Number of site pages0.194 (≈ none)

Two more nails in the technical-SEO coffin for AI:

  • Schema markup: no measurable effect (Finding 4).
  • Content length: irrelevant — 0.04 correlation with citation; 53.4% of cited pages are under 1,000 words; average cited length 1,282 words. (n=560K AIOs / 1.68M URLs.)

Implication: the AI-visibility lever is earned brand presence — especially YouTube and unlinked brand mentions — not on-page technical SEO, link volume, or content volume. This refines the wiki’s existing “brand mentions correlate ~3× more than backlinks” claim: brand mentions are strong (0.66–0.71), but YouTube mentions are stronger still (0.737). Correlation, not causation — but the direction matches independent evidence (glossary/e-e-a-t, Chen et al. 2025) that earned third-party authority beats brand-owned signals.

4. Schema markup does not move AI citations

The one quasi-experimental result, and it contradicts widespread GEO advice. Measuring AI-citation change for 1,885 pages that added JSON-LD vs 4,000 matched controls (difference-in-differences, 30-day windows):

  • Google AI Overviews: −4.6% (statistically significant but tiny, and negative)
  • Google AI Mode: +2.4% — indistinguishable from zero
  • ChatGPT: +2.2% — indistinguishable from zero

Ahrefs’ conclusion: “Adding schema produced no major uplift in citations on any platform.” The 53% of AI-cited pages that use schema reflect correlation, not causation — those sites also invest in authority, links, and content quality. A reported mechanism: AI crawlers extract visible HTML during retrieval and largely ignore JSON-LD.

Calibrated takeaway (not “schema is useless”): schema still earns traditional rich results, and author/publisher schema for entity verification is a different mechanism (seo/agentic-search-optimization) with independent support. But schema should no longer be framed as an AI-citation lever. Limits: the study tested only already-cited pages, only JSON-LD, and a 30-day window.

5. Format and freshness beat optimization — the “Best X” listicle

  • “Best X” listicles are 43.8% of ChatGPT’s cited page types (n=750 prompts / 26,283 source URLs) — the single most-cited format. Slightly more prominent still in AI Overviews. Brands in the top third of a comparison article are likelier to be recommended.
  • Freshness is load-bearing: 79.1% of cited lists were updated in 2025; 76.4% of ChatGPT-cited pages were updated within 30 days.
  • See glossary/best-x-listicle for the strategy (own the list, or get placed high in third-party lists).

6. AI answers are volatile on the surface, stable in meaning

  • AI Overviews change every 2.15 days on average; 70% of content differs between consecutive observations; ~45% of citations churn between generations — but semantic similarity stays 0.95. (n=43K keywords, Nov 2025.)

Implication: chasing a citation on a single query is futile — it’ll churn within days. Optimize for topic-level share of voice, measured as glossary/share-of-model, not individual placements.

Bonus: the retrieval-vs-citation gap, and where the AIO hit lands

  • ChatGPT cites only ~50% of the URLs it retrieves (49.98%). Being fetched ≠ being cited. Predictors of citation: title↔prompt semantic match (cited 0.602 vs non-cited 0.484 cosine) and readable URL slugs (89.8% vs 81.1%). See glossary/retrieval-vs-citation.
  • The AIO hit is concentrated: 99.9% of AI Overviews appear on informational queries; shopping triggers one only 3.2% of the time, transactional 2.1%, navigational 0.9% — but YMYL medical hits 44.1%. Bottom-funnel/transactional content is, for now, still click-bearing. (n=146M SERPs.)
  • International: emerging markets saturate faster (Indonesia 37.2%, US 13th at 20.5%, UK 19.1%); English is 52.75% of all AIOs. (n=108M AIO queries.)

The practitioner playbook (what to actually do)

  1. Optimize each AI surface separately. Track ChatGPT, AI Mode, and AI Overviews as distinct targets — citation overlap is ~14%. Don’t assume one win transfers.
  2. Invest in YouTube and earned brand mentions ahead of more on-page content or link volume. The top correlates are off-site and earned. PR is SEO now — see seo/ai-visibility.
  3. Stop treating schema as an AI lever. Keep it for rich results + author-entity verification; don’t budget it as GEO.
  4. Own or rank highly in “Best X” lists in your category, and keep them fresh (sub-30-day updates) — see glossary/best-x-listicle.
  5. Write semantically-matched titles and clean, readable URL slugs — they predict which retrieved pages get cited (glossary/retrieval-vs-citation).
  6. Measure topic-level share of voice, not single citations — answers churn every ~2 days; meaning is stable (glossary/share-of-model).
  7. Defend bottom-funnel content — transactional/commercial queries are still largely AIO-free; the collapse lands on top-funnel informational content.
  8. Don’t chase content length — 0.04 correlation; clarity and direct answers beat word count (seo/ai-seo-content).

Honest limits

  • Single vendor. Every number is Ahrefs-measured; no independent replication. A tool vendor has incentives in “SEO is changing, here’s what to do.”
  • Mostly correlational. Only the schema study isolates a causal effect. The YouTube 0.737 correlate is striking but could be confounded (brands big enough to be on YouTube are big enough to be cited).
  • Snapshots of a moving target. Most data is Sept–Dec 2025; AI surfaces change weekly. Re-check before relying on a specific figure.
  • Definitional disputes. The “ChatGPT = 12% of Google” figure hinges on a search-intent classification the original OpenAI/Harvard researchers would put far lower (24% vs the author’s 65%).
  • Calibrated, not endorsed. Several findings confirm the wiki’s prior direction (earned authority > backlinks; zero-click acceleration); the schema null corrects a prior wiki claim. Both are folded in honestly.

Key Takeaways

  • AI surfaces are separate discovery layers — ranking #1 guarantees neither AIO nor ChatGPT citation (13.7% cross-citation overlap; 28.3% of ChatGPT’s top pages have zero Google visibility).
  • The CTR collapse is accelerating — position-1 −58%, up from −34.5% in 8 months.
  • YouTube mentions (0.737) are the top AI-visibility correlate; technical SEO, link volume, and content length barely move it.
  • Schema markup has no measurable effect on AI citations (the one quasi-experimental finding) — keep it for rich results, drop it as a GEO lever.
  • “Best X” listicles + freshness + clean slugs + semantic titles are the actionable content plays.
  • All Tier 2 single-vendor measurement — strong scale, mostly correlational, not peer-reviewed.

Sources

All 14 studies (Ahrefs, 2026), grouped by finding:

Discovery layers / ranking ≠ citation:

Traffic / CTR:

Correlates / what works:

Format / freshness:

  • Best lists research — “Best X” lists 43.8% of ChatGPT citations (n=750 prompts / 26,283 URLs)

Volatility / triggers / international:

Institutional corroboration (not Ahrefs):