Mechanism
Retrieval and grounding, not memory
For live queries, generative engines do not recall your page from training. They retrieve it. Google describes its own generative features as using retrieval-augmented generation to pull current pages from the Search index, plus query fan-out: the model spawns several related sub-queries and retrieves across all of them before writing an answer.
The practical consequence is that one page ranking for one head term is a weak position. Coverage across the cluster of sub-questions a topic fans out into is a strong one.
Source: Google Search Central, AI features and your website guide, updated 10 July 2026.
Evidence hierarchy
The five best-evidenced citation factors
Cyrus Shepard’s May 2026 Zyppy meta-analysis scored 23 factors across 54 experiments, patents and case studies. The top five by evidence strength were URL accessibility, 9.5/10, classic search rank, 9.4, fan-out rank, 9.3, preview control, 9.2 and query-to-answer match, 9.2. Freshness scored a moderate 7.0. llms.txt scored lowest of all, at 2.0.
Read plainly: engines cite pages they can reach, that already rank, that match the exact question, and that permit a preview.
Source: Zyppy Signal, AI Citation Ranking Factors, 7 May 2026, 54 studies scored 0 to 10.
Rank decoupling
Position one is neither necessary nor sufficient
Ahrefs analysed 863,000 keywords and four million AI Overview URLs in March 2026 and found 38% of AI Overview citations came from pages ranking in Google’s top 10, down from roughly 76% in mid-2025. About 31.2% came from positions 11 to 100, and about 31.0% from beyond position 100.
Other datasets disagree on the exact figure. BrightEdge has reported both much lower and much higher overlap depending on window and industry, which is itself the point: the number is unstable, so treating rank as a proxy for AI visibility is unsafe. Measure the answer surface directly.
Sources: Ahrefs, 863K keywords and 4M AIO URLs, March 2026. BrightEdge rank-overlap tracking, 2026.
Off-page
Mentions correlate about 3 times more strongly than backlinks
Ahrefs’ 75,000-brand study found branded web mentions correlated at r=0.664 with AI Overview visibility, branded anchors at 0.527 and branded search volume at 0.392, while Domain Rating came in at 0.326 and raw backlink count last at r=0.218. The top quartile of brands by web mentions averaged 169 AI Overview mentions against 14 for the next quartile.
The honest caveat the authors themselves add: already-strong brands may simply earn both. Correlation, not proof.
Source: Ahrefs, 75,000-brand Spearman correlation study, 2026.
Freshness, corrected
Cited content is fresher, by 25.7%, not 4.3 times
A widely repeated claim says fresh content is cited 4.3 times more often. That figure traces to no primary study and should not be repeated. The verified measurement, across 16.98 million cited URLs on seven platforms, is that AI-cited content averages 1,064 days old versus 1,432 days for organic top-10 results, a real but moderate 25.7% freshness advantage. ChatGPT skewed freshest at 958 days.
So refresh cadence matters, and it still ranks below accessibility, rank and brand presence in the evidence hierarchy.
Source: Ahrefs content-age study, 16.975M cited URLs, seven platforms, 2026.
Engine divergence
Five engines, five different source pools
A 680-million-citation analysis found only about 11% domain overlap between ChatGPT and Perplexity, and Google’s AI Overviews and AI Mode returned the same URLs only around 13.7% of the time despite reaching similar conclusions. Seer Interactive separately found 87% of SearchGPT citations matched Bing’s top results, while Perplexity performs live retrieval on every query and cites far more sources per answer.
This is why a single blended AI visibility score hides the information you need, and why this practice reports each engine apart.
Sources: AI Platform Citation Source Index 2026, 680M citations. Seer Interactive SearchGPT and Bing study.
What to ignore
Google’s own list of GEO tactics that do nothing
Google’s July 2026 guidance names specific tactics as unnecessary for Google Search and its AI features: llms.txt and other special AI files, artificially chunking content, rewriting content specifically for AI systems, seeking inauthentic mentions, and over-focusing on structured data as if it were a citation requirement.
This site publishes an llms.txt anyway, not as a ranking play, but because it is a useful, human-auditable canonical summary of the entity. That distinction is the whole difference between GEO practice and GEO folklore.
Source: Google Search Central, generative AI features guide, mythbusting section, July 2026.
Format
Editorial depth does the heavy lifting
BuzzStream analysed four million citations from 3,600 prompts across ten industries and found blog and content pages supplied 53.46% of citations, news 14.09% and social 8.71%. With brand-owned queries excluded, earned editorial content accounted for roughly 80%. Syndicated press releases accounted for 0.04% of the entire dataset.
Directional rather than permanent, since the sample covers a single week from late January 2026, but the gap is too wide to be noise.
Source: BuzzStream citation-format analysis, 4M citations, 3,600 prompts, 2026.
Why it is worth it
Citation appears to lift clicks, not just ego
Seer Interactive’s April 2026 study of 53 brands across 5.47 million queries found pages cited in AI Overviews were associated with roughly 120% more organic clicks per impression than uncited pages. SparkToro’s June 2026 study found 68% of US Google searches ended without a click at all.
Read together: fewer searches produce a visit, and the ones that do increasingly favour cited sources. Seer’s authors flag that stronger brands may earn both, again, correlation.
Sources: Seer Interactive, 53 brands and 5.47M queries, April 2026. SparkToro zero-click study, June 2026.