What Actually Earns AI Citations in 2026 (Five Studies, One Playbook)

July 17, 2026·9 min read·Last updated May 28, 2026

Most GEO/AEO advice you'll read this quarter is unsupported by the strongest 2025-2026 studies. We reviewed the largest public controlled tests — Ahrefs (three separate studies covering 1,885, 174,048, and 15,000-prompt datasets), Ziptie, AirOps, searchVIU, and the PikaSEO per-engine compilation — and rebuilt the playbook against real numbers. Five findings changed the way we score sites at [/check](/check). None of them match the current conventional wisdom.

1. Schema doesn't move AI citations on its own

**The number:** Ahrefs added JSON-LD to 1,885 pages and measured citation change against a matched control cohort. ChatGPT moved **+2.2%**, Google AI Mode **+2.4%**, Google AI Overviews **−4.6%**. All statistically indistinguishable from zero. ([study](https://ahrefs.com/blog/schema-ai-citations/))

A companion searchVIU test observed live fetches from ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode. All five parsed **only the rendered HTML** at inference time. JSON-LD in the page head was ignored. Schema still helps discoverability in Google's classic index, rich results, and sitelinks. It should not be sold as a direct citation growth lever.

**What to do:** keep schema on your leaves — it's a durable machine-readability signal and it survives when the page is quoted out of context. Stop selling it as the primary AI-citation lever. See [the full stat](/stats/schema-barely-moves-ai-citations).

2. Front-loading the answer wins — 44% of citations come from the first 30% of the page

**The number:** Ziptie analysed citation-passage positions across ChatGPT, Perplexity, Claude, and Google AI Overviews. **44.2%** of citations pull from the first 30% of body text. An independent 100-page UK replication (AI Boost) landed at **47.3%**. ([Ziptie](https://ziptie.dev/))

LLM extractors work paragraph-by-paragraph on a finite budget. Early paragraphs get scored more often. When two paragraphs cover the same claim, the earlier one is the citation. This is why "answer-first" isn't just readability advice — it's citation economics.

**What to do:** on every leaf page, the first 50-70 words must answer the page's implicit question. Under every H2, put a 40-60 word answer paragraph before any preamble. Kill "welcome to our blog" openers — they occupy prime citation real estate with zero payload. See [the full stat](/stats/first-30-percent-earns-44-percent-of-citations).

3. Word count barely correlates with citation likelihood

**The number:** Ahrefs studied **174,048 pages** surfaced in Google AI Overviews. Spearman rank correlation between word count and citation frequency: **r = 0.04** — statistically indistinguishable from zero. **53.4%** of cited pages are under 1,000 words. ([study](https://ahrefs.com/blog/short-vs-long-content-in-ai-overviews/))

The "publish a 10,000-word ultimate guide" playbook was an SEO tactic aimed at dwell time and internal-linking hubs. Neither signal transfers to AI citation. Worse, mega-guides bury the answer past the first-30% window, dilute the entity graph, and cost more to keep fresh.

**What to do:** split mega-guides into 6-12 short, answer-first pages, one per query. Cross-link them via a hub page. Editorial floor ~350 words, soft ceiling ~1,200. See [the full stat](/stats/short-content-cites-just-as-well).

4. Freshness helps everywhere except Google AI Overviews

**The number:** Ahrefs analysed **17M citations** and found AI assistants cite content **25.7% "fresher"** than organic SERPs on average — 1,064 vs 1,432 days old. ChatGPT shows the strongest bias (citing pages 393-458 days newer than SERPs). Google AI Overviews is the exception: it cites content 16 days *older* than organic results. ([study](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/))

**What to do:** add `dateModified` to Article JSON-LD, render a visible `<time>Reviewed [date]</time>` near the H1, and re-review quarterly. Don't fake it — models cross-check with archive.org. Freshness is the single cheapest lever on ChatGPT, Claude, and Perplexity; it's a nothing-burger on AIO.

5. Each engine has a different citation graph — optimise per engine

**The numbers:** PikaSEO's Q1 2026 compilation across Bluefish, Tinuiti, Ahrefs (78.6M searches), and SE Ranking (1.3M citations):

**ChatGPT** — Wikipedia = **47.9%** of citations. Wikipedia entity ownership is decisive.
**Perplexity** — Reddit = **24%**, YouTube = **21%**. UGC is a first-class channel. Also closest to Google: **28.6%** overlap with organic top 10.
**Claude** — Brand-owned domains = **70%**. Cleanest, most corporate citation set. This is the engine where your own domain's GEO compliance pays the highest dividend.
**Google AI Overviews** — Reddit = **21%**, top-10 overlap = **37.9%** (down from 76% in July 2025). Rapidly decoupling from its own SERP.

**What to do:** stop treating "AI search" as one channel. Match investment to where your buyers actually ask questions. If ChatGPT dominates your audience, invest in your Wikipedia entity (a 6-12 month project). If Perplexity does, show up on Reddit under a consistent brand-linked handle. If Claude does, obsess over your own site's freshness, structure, and depth. See [the full breakdown](/stats/wikipedia-reddit-brand-per-engine).

Bonus: 80% of AI-cited URLs don't rank in Google's top 100

Ahrefs found **only 12%** of AI citations (ChatGPT/Gemini/Copilot) match Google's top 10 for the same query. **80%** of AI-cited URLs don't rank in the top 100 at all. Google AI Overviews itself has fallen from **76% top-10 overlap (July 2025) to 38% (early 2026)** — the AI layer is decoupling from classic search faster than any GEO tool has priced in. ([study](https://ahrefs.com/blog/ai-search-overlap/))

The rebuilt playbook

In priority order, ranked by evidence strength:

1. **Front-load the answer** — 40-60 word answer under every H2, in the first 30% of the page. Strongest lever, near-zero cost. 2. **Freshness signals** — `dateModified` + visible review date + quarterly re-review. Zero cost on your existing pages. 3. **Per-engine format fit** — listicles for commercial queries, articles for informational. Match format to intent. 4. **Per-engine channel investment** — Wikipedia for ChatGPT, Reddit for Perplexity, own-domain quality for Claude. 5. **Structural fundamentals** — schema, llms.txt, agent-card.json. Necessary hygiene for discoverability; not a citation growth lever on their own. 6. **Word count** — ignore. Optimise for density, not length.

What to do this week

1. Run [/check](/check) on your top five leaf pages. Fix any answer-first failures first. 2. Add `dateModified` and visible review dates to the same five pages. 3. Pick your dominant engine from your analytics referrer data. Match your next quarter's content investment to that engine's citation graph. 4. Stop measuring keyword rankings as a proxy for AI visibility. Track citation share directly ([how](/blog/tracking-ai-search-visibility)).

Sources

Ahrefs — [Schema and AI citations study](https://ahrefs.com/blog/schema-ai-citations/) (1,885 pages)
Ahrefs — [AI assistants prefer fresher content](https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/) (17M citations)
Ahrefs — [Only 12% of AI-cited URLs rank in Google's top 10](https://ahrefs.com/blog/ai-search-overlap/) (15,000 prompts)
Ahrefs — [Short vs long content in AI Overviews](https://ahrefs.com/blog/short-vs-long-content-in-ai-overviews/) (174,048 pages)
Ziptie — [How AI splits your content across multiple answers](https://ziptie.dev/) (April 2026)
AirOps — [Which page types earn the most AI citations](https://www.airops.com/blog/page-types-earn-ai-citations) (June 2026)
searchVIU — [Schema markup and AI in 2025](https://www.searchviu.com/en/schema-markup-ai/)
PikaSEO — [AI citation sources by platform](https://pikaseo.com/articles/ai-citation-sources-by-platform) (Q1 2026 compilation)

License: CC BY 4.0. Cite as "grow.contact, What Actually Earns AI Citations in 2026 (2026-07-17)".