The conversation around generative engine optimization moves fast, and a lot of what circulates as "data" is vendor conjecture dressed up as research. This post collects only figures that trace to a named study, a peer-reviewed paper, or a published methodology. Where a figure comes from a simulated experiment or a single vendor's dataset, we say so.
The picture that emerges is consistent: AI-powered search is large enough to matter, the citation mechanics differ meaningfully from classical ranking, and the content tactics that earn AI citations are well-defined, if not yet universally practiced.
How Big Is AI Search, Really?
AI-powered search is growing fast enough that the numbers from eighteen months ago are already outdated. The clearest trend line comes from two directions: AI platform growth and the share of search queries now answered with an AI-generated overview.
On the platform side, Similarweb data shows total AI referral visits across the web grew more than 3x between September 2024 and September 2025. ChatGPT's web visits grew 84% over that same stretch; Claude's grew roughly 770%. Perplexity processed approximately 780 million queries in May 2025, up from 230 million in August 2024, a 239% increase in nine months, according to usage data aggregated by Seoprofy.
Within Google Search specifically, Semrush's AI Overviews study tracked the share of queries triggering an AI Overview throughout 2025. The rate hit a peak of 24.61% in July 2025, pulled back to 15.69% of US queries in November 2025, and the mix of query types shifted dramatically: commercial-intent queries triggering AI Overviews grew from 8.15% in January 2025 to 18.57% by October 2025, while transactional triggers grew from under 2% to nearly 14%. AI Overviews are no longer confined to purely informational queries.
The Gartner prediction from February 2024 has become one of the most-cited projections in this space: traditional search engine volume will drop 25% by 2026 as users shift toward AI-powered conversational assistants. Gartner's longer-term projection suggests organic search traffic could decline 50% or more as generative AI search matures.
Source: Semrush AI Overviews study (semrush.com/blog/semrush-ai-overviews-study/). Monthly figures, January to November 2025.
Note on the chart: the February through October values are interpolated from the Semrush study's reported January, July, and November anchors, and are illustrative of the trend rather than exact monthly readings. The January, July, and November figures are study-reported.
What AI Overviews and AI Search Do to Organic Clicks
The click-impact question is the one every content team is asking, and the answer varies by study design and query type. Here is what the better-controlled research actually shows.
A randomized field experiment by researchers at the Indian School of Business and Carnegie Mellon University is among the most methodologically rigorous work available. The study recruited 1,065 US desktop Chrome users in January and February 2026. Participants were randomly assigned to one of three groups: normal Google Search, a hidden-AI-Overview group (an extension removed AI Overviews in real time), and an AI Mode group. The key finding: removing a top-position AI Overview nearly doubled outbound clicks, from 0.38 to 0.61 per search, representing a 38% reduction when the AI Overview was present. Zero-click searches rose from 54% to 72% when an AI Overview appeared. Importantly, the study found no measurable difference in user satisfaction or ease of finding information between the control and hide-AIO groups, challenging Google's "bounce-click" rationale.
Ahrefs' study using 300,000 keywords (150,000 with an AI Overview, 150,000 informational keywords without) found that AI Overviews reduce position-one click-through rates by 34.5% compared to equivalent queries without an Overview. The study compared March 2024 (pre-US-rollout) to March 2025 (post-rollout).
One counterintuitive finding from the Semrush study: for keywords that newly acquired an AI Overview, zero-click rates actually decreased slightly, from 33.75% to 31.53%. The implication is that the click-suppression effect is real but concentrated in top-of-page placements on established queries, not uniform across all queries with AI Overviews.
The conversion quality picture complicates the volume story. Similarweb reports that ChatGPT referral traffic converts at 7.1%, second only to paid search at 7.8%, and ahead of organic, direct, social, email, and display. AI-referred traffic appears to be higher-intent than most other channels, which matters for how you weigh the volume versus quality tradeoff.
What Actually Drives AI Citations
This is the most actionable section, and also the one with the most variance across studies. Here is what the evidence supports.
The KDD 2024 GEO Paper: Tactic-Level Visibility Lifts
The most rigorous academic work is the GEO paper from KDD 2024 by researchers at Princeton, Georgia Tech, and collaborating institutions. The paper tested specific content modifications on GEO-bench (a simulated generative engine) and on Perplexity.ai (a real deployment). Results on the simulated engine are directional; real-engine results are smaller but consistent in direction.
Source: GEO paper, KDD 2024 (arxiv.org/abs/2311.09735). Simulated engine (GEO-bench); directional figures. Baseline visibility metric: 19.5.
Key findings from the paper:
Quotation addition (+41% on simulated engine, +22% on Perplexity.ai). Adding direct quotations from authoritative sources was the single highest-lift tactic. On Perplexity.ai, the subjective impression score improved 32.1%. The hypothesis: generative engines favor content that contains the kind of quotable, attributable text they themselves need to compose a cited answer.
Statistics addition (+40% on simulated engine, +9% on Perplexity.ai). Adding specific, sourced data points produced the second-largest visibility gain. On Perplexity.ai, the subjective impression improvement was 37%, the highest of any individual tactic.
Citing sources (+30% on simulated engine). Content that attributes claims to external sources improved visibility substantially. This mirrors the AI's own output format: systems already designed to compose cited answers appear to favor content structured the same way.
Keyword stuffing (-8% on simulated engine, -10% on Perplexity.ai). This is the most important negative finding. High keyword density actively reduced citation likelihood on both test environments. The GEO playbook inverts one of the most persistent old-school SEO instincts.
The paper also found that lower-ranked pages benefit disproportionately. Pages at rank 5 saw 99-115% visibility improvements from citation-driving tactics, versus smaller or even negative changes for rank-1 pages already earning high baseline visibility.
The SE Ranking Study: Domain Authority and Content Signals
SE Ranking's analysis of 216,524 pages across 129,000 unique domains in 20 niches identified the strongest correlates of ChatGPT citations. The authority signals dominate at the top:
- Sites with 350,000 or more referring domains averaged 8.4 ChatGPT citations per response. Sites with 2,500 referring domains averaged 1.6 to 1.8. The relationship is roughly linear with a sharp threshold effect around 32,000 referring domains, where citations nearly doubled.
- Domain Trust scores above 97 averaged 8.4 citations. Scores below 43 averaged 1.6. The gap is fivefold.
- Content updated within three months averaged 6 citations. Outdated content averaged 3.6. Freshness roughly doubles citation rate.
- Pages with 19 or more statistical data points averaged 5.4 citations. Pages with minimal data averaged 2.8. This corroborates the GEO paper's statistics-addition finding with a real-engine dataset.
- Pages with attributed expert quotes averaged 4.1 citations versus 2.4 without. Quotes nearly double citation rate, again consistent with the GEO paper.
- Word count over 2,900 averaged 5.1 citations versus 3.2 for pages under 800 words.
One surprising finding: FAQ schema markup correlated with slightly fewer citations (3.6 versus 4.2 without), and question-style headings slightly underperformed straightforward declarative headings. ChatGPT appears to favor topic-clear, descriptive content over explicitly question-formatted sections.
The Ahrefs ChatGPT Study: How Citations Actually Get Selected
The Ahrefs analysis of 1.4 million ChatGPT prompts provides the clearest picture of the mechanical citation selection process.
The source-type breakdown is striking. Of all URLs ChatGPT retrieved across 1.4 million prompts, 88.46% of the cited URLs came from ChatGPT's standard search index. News sources accounted for 12.01% of citations. Reddit was retrieved in volume (over 16 million entries) but cited only 1.93% of the time. YouTube URLs had a 0.51% citation rate.
The semantic relevance finding explains why the standard best practices work: the cosine similarity between a prompt and a cited page title was 0.602, versus 0.484 for non-cited pages. When fanout sub-queries (the internal related questions ChatGPT generates) were used as the comparison, cited titles scored 0.656. Content that directly answers the sub-questions a generative engine asks internally is materially more likely to be cited.
URL structure matters at the margins. Natural-language URL slugs had an 89.78% citation rate; opaque URLs had an 81.11% citation rate. An 8.7-percentage-point gap from a one-time decision.
Platform Differences: ChatGPT, Perplexity, and AI Overviews Do Not Cite the Same Pages
This is an underappreciated point in most GEO content. The platforms are not interchangeable citation pools.
An analysis of 680 million citations found that only 11% of domains are cited by both ChatGPT and Perplexity. Optimizing for one platform does not automatically generalize to the other.
Perplexity references significantly more sources per response. One analysis found Perplexity averaged 21.9 citations per response versus ChatGPT's 10.4. Perplexity's citation rate (the share of retrieved pages that actually get cited) was measured at 15.43% in one study, versus 2.78% for ChatGPT. The difference in citation density reflects different answer formats: Perplexity returns a source-heavy response by default; ChatGPT is more selective.
The ranking-to-citation correlation has shifted substantially for Google AI Overviews. Ahrefs analyzed 863,000 keywords and 4 million AI Overview URLs and found that only 38% of cited pages rank in the top 10 for the same query, down from 76% in July 2025. About 31% of citations come from pages ranking positions 11 to 100, and 31% from beyond the top 100. Google's query fan-out process is drawing from a wider pool than classical top-10 optimization would predict.
Position still matters for citation probability. Data collected on Google AI Overviews shows position 1 earns a 33.07% citation probability, falling to 13.04% at position 10. But the 38% finding above shows that a majority of citations now come from outside the top 10, which is a meaningful structural shift from where the data stood eighteen months ago.
What the Projections Say
Forward-looking figures deserve extra scrutiny because they are extrapolations, but a few institutional projections have enough specificity to be worth tracking.
Gartner's February 2024 prediction: traditional search engine volume will decline 25% by 2026 due to AI chatbots and other virtual agents. The firm's longer-range view is a potential 50% or more decline in organic search traffic as generative AI search matures. The 25% figure has been widely cited; the underlying survey data showed strong consumer preference for AI-enhanced search.
The Similarweb data on ChatGPT referral volumes shows a step-change that corroborates directional acceleration: after ChatGPT introduced clickable brand links directly in responses (May 7, 2026), total referrals from the platform increased 157.7% week-over-week. Homepage referrals grew 354.7% in the same week. The referral mechanics of AI platforms are still being designed; the current traffic figures are likely not a ceiling.
Semrush's data on AI Overview industry spread indicates the shift is not uniform. Science queries triggered AI Overviews 25.96% of the time in November 2025. Computers and Electronics: 17.92%. Real Estate and Shopping: under 3%. Content teams in high-saturation industries face a more urgent GEO timeline than those in low-saturation ones.
FAQ
What is the strongest single tactic for generative engine optimization?
Based on the KDD 2024 GEO paper, adding direct quotations from authoritative sources produced the largest visibility lift on both the simulated engine (+41%) and Perplexity.ai (+22%). Statistics addition is the second-strongest tactic and also produced the highest improvement in subjective impression scores on Perplexity.ai (+37%). Both tactics align with what SE Ranking's large-scale real-engine study found: pages with attributed expert quotes and 19 or more data points earn substantially more citations.
Does ranking on page one guarantee AI citations?
No, and the correlation is weaker than it was. Ahrefs' analysis of 4 million AI Overview URLs in early 2026 found only 38% of cited pages ranked in the top 10 for the same query, down from 76% in mid-2025. Classical ranking is still the dominant lever for AI citation eligibility, but Google's query fan-out process draws from a wider content pool. For ChatGPT, SE Ranking data shows pages ranking positions 1-45 average 5 citations versus 3.1 for positions 64-75, so ranking still matters, but is not deterministic.
Are the GEO paper statistics reliable?
The KDD 2024 paper's core experiments used a simulated engine (GEO-bench), not live Google or ChatGPT APIs. Real-engine tests on Perplexity.ai showed smaller absolute lifts: +22% for quotation addition versus +41% on the simulated engine. Treat the GEO paper figures as directional signals indicating which tactics help and which hurt, not as guaranteed performance guarantees. The direction of every tactic in the paper has been corroborated by at least one real-engine dataset.
Do different AI platforms cite the same sources?
No. An analysis of 680 million citations found only 11% of domains are cited by both ChatGPT and Perplexity. The platforms have distinct source preferences. Perplexity averages roughly twice as many citations per response as ChatGPT and has a much higher citation rate for retrieved pages (15.43% versus 2.78%). A platform-specific citation strategy, rather than a single universal approach, is more likely to produce results across both systems.
How much does content freshness matter for AI citations?
SE Ranking's study of 216,524 pages found content updated within three months averaged 6 citations per response, versus 3.6 for outdated content. The Ahrefs ChatGPT citation study found cited pages had a median age of roughly 500 days, but generative engines skew toward fresher sources on recency-sensitive queries. Freshness metadata and explicit update dates are worth maintaining.
The data in this post points toward a consistent underlying pattern: AI systems cite content that looks like what they are trying to produce. Cited, structured, statistically grounded, and authoritative content earns inclusion in AI-generated answers for the same reason it earns links from human editors. The citation mechanics are different, but the content standard is not.
We built SparkBlog around exactly this view: that ranking smarter starts with treating content as an engineered system, where every piece is structured for AI extractability and every claim is grounded in a real source. The research here is the foundation. The discipline is applying it consistently across an estate, not just on a single post.
For more on the underlying mechanics, see our pillar on what generative engine optimization actually is and where it came from. If you are working on the specific tactics for getting cited across platforms, see our guide on how to get your content cited by ChatGPT, Perplexity, and AI Overviews. For a comparison of how GEO relates to the broader set of optimization frameworks, see our breakdown of AEO vs GEO vs SEO and our pillar on answer engine optimization.