This is not an SEO roundup. Nine primary sources — from the Princeton/Georgia Tech GEO study to the official crawler documentation of Google, OpenAI and Perplexity — evaluated, every claim linked directly to its source.
Quick-Check — Executive Summary
- Core finding: Content with quotes, statistics and citations gets cited up to 40% more often in generative search answers, per the Princeton/Georgia Tech study "GEO: Generative Engine Optimization" — plain keyword stuffing actually hurts slightly.
- Trigger frequency: Per Pew Research, an AI Overview summary appeared for 12,593 of 68,879 tracked Google searches (18%) in March 2025; 58% of the 900 US users studied had at least one such search.
- The click drop: With an AI summary present, users click a result only 8% of the time per Pew (vs. 15% without a summary) — and only 1% click any of the sources cited inside the summary itself.
- Measurable CTR erosion: Ahrefs puts the click-rate decline for position 1 with an AI Overview present at 34.5% (March 2025) rising to 58% (December 2025, both vs. keywords without an AI Overview).
- Zero-click becomes the norm: 68.01% of all Google searches in the US ended with zero clicks per SparkToro/Similarweb data (Jan–Apr 2026) — up from 60.45% in 2024.
- Being cited still pays off: a page named as a source inside an AI Overview gets 35% more clicks than a comparably ranked but uncited page per Ahrefs — visibility inside the AI answer becomes the real currency.
Research Case Study
When the search engine already delivers the answer
Since Google rolled out AI Overviews across the US and beyond, and ChatGPT, Perplexity and Copilot established themselves as standalone answer engines, the central question in content marketing has shifted: no longer "how do I rank #1," but "how do I get cited as a source in the generated answer — and is that even worth it when 68% of all searches end without a click anyway." Between SEO blogs selling GEO as a mere buzzword update to SEO, and panic pieces about the "end of the open web," what's usually missing is a solid data basis: what do the studies actually show, with what methodology, and what do the platforms themselves say officially about their crawling and citation mechanisms?
This case study evaluates primary sources only: the original Generative Engine Optimization paper by researchers from Princeton University, Georgia Tech, the Allen Institute for AI and IIT Delhi (arXiv 2311.09735); Pew Research's raw-data study of real click behavior (900 users, 68,879 tracked searches); the CTR analyses from Ahrefs (300,000 keywords) and Amsive (700,000 keywords across 10 websites and 5 industries, reported via Search Engine Land); SparkToro's zero-click study based on Similarweb clickstream data; and the official technical documentation from Google Search Central, OpenAI and Perplexity on their respective crawlers. No secondary citations without checking the original source.
01
Quotes, statistics and citations measurably boost visibility in AI answers — keyword stuffing hurts
The GEO study tested nine content-optimization methods (including quotation addition, statistics addition, cite sources, fluency optimization, authoritative language, technical terms, easy-to-understand, unique words, and keyword stuffing) using GEO-Bench, a purpose-built benchmark of 10,000 queries (8,000 training / 1,000 validation / 1,000 test) across nine datasets including MS MARCO, Natural Questions and ELI5, with GPT-3.5-turbo as the generative engine and validation on the live, production Perplexity.ai. Measured by "position-adjusted word count" (how strongly and how prominently a source appears in the generated answer), the baseline sat at 19.3%. Adding quotations raised it to 27.8%, statistics to 25.9%, source citations to 24.9% — a relative visibility gain of up to 40%. Classic keyword stuffing, by contrast, pushed the value down to 17.8%, below baseline.
Technical Deep-Dive
The most interesting side finding of the GEO study isn't about the optimization methods themselves, but about the distribution effects when multiple sources optimize simultaneously: in the authors' ranking simulations, previously poorly placed sources (rank 5 in classic search results) gained 97 to 115% more visibility in the generated answer through GEO tactics, while previously dominant sources (rank 1) lost between 6% and 30% of their visibility. The authors interpret this as a potential "democratization effect": sites that rank poorly in classic SERPs can catch up in generative answers through structured, citable content — a mechanism that's hard to crack in classic SEO through link authority and domain age. Important caveat: effect strength varied considerably by topic domain in the study (quotations worked best for "people & society," "explanation" and "history," statistics for "law & gov.," "debate" and "opinion") — blanket "add quotes" advice without regard to actual content relevance isn't a free pass.
02
AI Overviews appear on nearly one in five Google searches — and users click almost half as often afterward
Pew Research analyzed real browsing behavior from 900 consenting US adults in March 2025, across 68,879 unique Google searches. An AI Overview summary appeared for 12,593 of them (18%); 58% of participants experienced at least one such search during the study month. The key finding on click behavior: for searches with an AI summary, users clicked a result only 8% of the time, compared with 15% for searches without a summary — nearly a halving. Of the sources cited inside the summary itself, only 1% were clicked at all. Users also ended their browsing session immediately after a page with an AI summary (26%) more often than after a page without one (16%).
03
Ahrefs: CTR loss on position 1 nearly doubles within nine months, from 34.5% to 58%
Ahrefs compared 300,000 keywords using aggregated Google Search Console data (150,000 with, 150,000 without an AI Overview, informational queries only, since 99.2% of all AI Overview triggers fall into that intent type). At the first measurement point (March 2024 vs. March 2025, right after the US rollout), average click rate for position 1 fell from 0.056 to 0.031 — for keywords with an AI Overview, the real observed rate (0.026) sat well below the projected rate for that period without an AI Overview (0.040), a 34.5% CTR decline. A follow-up analysis with data through December 2025 shows a further worsening to a 58% CTR decline. The effect thins out with position: position 2 loses 50.8%, position 10 still loses 19.4%. At the same time, Ahrefs found that being cited inside the AI Overview itself brings 35% more organic clicks than an identical position without a citation — citation cushions the loss but doesn't fully offset it.
04
Branded beats non-branded: an Amsive analysis of 700,000 keywords shows a massive asymmetry
Performance agency Amsive analyzed 700,000 keywords across 10 websites in 5 industries and found an average CTR decline of 15.49% in the presence of an AI Overview. The effect is, however, strongly segment-dependent: for non-branded keywords, the decline is 19.98%, rising to 27.04% outside the top-3 positions, and it collapses to a 37.04% drop when combined with a simultaneously shown featured snippet. Branded searches, on the other hand, trigger an AI Overview only 4.79% of the time — and when they do, CTR actually rises 18.68%, likely because users with clear brand intent click through regardless. Informational, non-branded content therefore absorbs the bulk of the traffic loss.
05
Zero-click becomes the norm: 68% of all Google searches end without a click in 2026
SparkToro (Rand Fishkin), together with Similarweb, analyzed the latter's US desktop and mobile clickstream panel for January through April 2026, compared with the same period in 2024. Result: 68.01% of all Google searches ended without any click whatsoever (neither organic, paid, nor onto Google's own properties like Maps), up from 60.45% in 2024 — a 7.56-point rise in two years. The share of searches with at least one click fell 22.9% in relative terms over the same period. The zero-click share differs sharply by device: around 77% on mobile, around 50% on desktop (excluding the separate Google app, which SparkToro expects to run even higher).
06
The zero-click trend isn't (yet) an AI Overview effect — but it's the base one is building on
Search Engine Land's read on the SparkToro study corrects a common misconception: in the studied period (January–April 2026), Google's "AI Mode" accounted for just 0.34% of all searches — meaning the massive zero-click rise is driven mainly by established Google features (featured snippets, knowledge panels, classic AI Overviews), not primarily by new generative search formats. At the same time, Google reportedly told the outlet that AI Mode now reaches over 1 billion monthly users and that query volume more than doubles quarter over quarter — a sign of a possible amplification effect in upcoming measurement periods. The article's practical conclusion: "SEO alone may be insufficient for many publishers seeking to regain historical levels of Google-referred traffic" — the searches that stay profitable are mostly branded, local and transactional with high purchase intent.
SEO alone may be insufficient for many publishers seeking to regain historical levels of Google-referred traffic. What stays profitable is mostly branded, local and transactional search.
Search Engine Land — commentary on the SparkToro zero-click study, June 202607
Google itself: no special tags needed for AI Overviews — but "query fan-out" changes what gets found at all
Per official Google Search Central documentation, AI Overviews (and AI Mode) work via a "query fan-out" technique: instead of a single search query, the system issues multiple related queries across subtopics and data sources, then merges the results. On the technical requirements, Google is explicit: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." The only condition is that a page is indexed and eligible to appear with a regular snippet in search results. Sites wanting to opt out specifically can use the existing meta tags nosnippet, data-nosnippet, max-snippet or noindex, per the documentation. Google explicitly stresses that no new machine-readable files or special markup formats are needed for AI search features — instead it recommends the same established SEO fundamentals: crawlability and "helpful, reliable, people-first content."
08
OpenAI separates training, search and live fetching into three independently controllable crawlers
OpenAI's official developer documentation distinguishes three bot roles with different functions and different robots.txt effects: GPTBot collects public web content for training the foundation models ("GPTBot is used to make our generative AI foundation models more useful and safe") and can be excluded from training via robots.txt. OAI-SearchBot, independently, builds a search index to surface websites in ChatGPT's search features ("OAI-SearchBot is used to surface websites in search results in ChatGPT's search features") — if this bot is blocked, a site won't appear in ChatGPT search results or live citations, regardless of its GPTBot status. ChatGPT-User, finally, only activates for a specific, user-triggered request and doesn't crawl the web automatically, so classic robots.txt logic only partially applies to it. For visibility in ChatGPT search answers, sites must therefore explicitly allow OAI-SearchBot, independent of whatever decision they've made about training data.
09
Perplexity: the search crawler respects robots.txt, the live fetcher on user requests does not
Perplexity's official documentation describes the same structural split as OpenAI, just under different names: PerplexityBot "is designed to surface and link websites in search results on Perplexity" and, per the documentation, is explicitly not used to train AI foundation models, only for search and citation — sites that block this bot via robots.txt won't be indexed in content terms. Perplexity-User, by contrast, activates "when users ask Perplexity a question" and the answer requires a live fetch of a page; that fetcher "generally ignores robots.txt rules," since the request was triggered directly by the user and counts as an individual fetch rather than automated crawling. For site owners, that means blocking PerplexityBot prevents durable inclusion in the search index but doesn't reliably protect against a one-off live fetch at the moment of a specific user query.
Visibility and click effects in AI search systems compared| Parameter | Baseline | Observed value | Effect | Source |
|---|
| Visibility from citations in content (GEO-Bench) | Baseline position-adjusted word count: 19.3% | With quotation addition: 27.8% | +8.5 points (~40% relative increase) | [1] |
| Click rate on Google search with AI Overview | Without AI Overview: 15% click rate | With AI Overview: 8% click rate | −47% relative | [2] |
| CTR position 1 with AI Overview (Ahrefs, Dec 2025) | Projected CTR without AI Overview | Real CTR with AI Overview | −58% | [3] |
| CTR non-branded vs. branded keywords (Amsive) | Non-branded: −19.98% | Branded: +18.68% | 38.7-point asymmetry | [4] |
| Zero-click share of all Google searches (US) | 2024: 60.45% | Jan–Apr 2026: 68.01% | +7.56 points in 2 years | [5] |
| Rank shift from GEO optimization | Rank-1 source with no countermove | Rank-5 source with GEO tactics | +97 to +115% visibility for rank 5, −6 to −30% for rank 1 | [1] |
Learnings & Actionable Insights
What follows in practice
For companies
Content briefs need to follow the GEO study's lead: concrete statistics, verbatim quotes and traceable citations instead of keyword density. Since Ahrefs shows cited sources get 35% more clicks, visibility inside the AI answer pays off even when overall CTR falls — the goal is citation, not just the blue link anymore.
For agencies
Reporting KPIs need to move past classic ranking positions: Amsive shows non-branded informational content can lose up to 37% CTR, while branded search actually gains. Client content strategy should be split by intent type rather than promising blanket "more AI Overview visibility."
For developers & creators
GPTBot, OAI-SearchBot and PerplexityBot are technically separate crawlers with different robots.txt effects and different purposes (training vs. search vs. live fetch). Anyone who wants to be cited in ChatGPT Search or Perplexity must explicitly allow OAI-SearchBot or PerplexityBot — independent of the separate decision about GPTBot for model training.
For investors & analysts
A 68% zero-click share and a CTR loss of up to 58% on position 1 are a structural risk for any business model that depends directly on search-referral traffic. What matters for valuation is how far a publisher portfolio has already diversified into branded search, owned-audience channels and citation visibility instead of pure click traffic.
Open Questions
- How stable are the effects measured in the GEO study across newer model generations (GPT-5, Gemini 3, Claude), given the original study still used GPT-3.5-turbo?
- Does the "democratization effect" the GEO study observed (rank 5 gaining more than rank 1 loses) eventually homogenize content, if every publisher copies the same citation and statistics tactics?
- How does the split shift between Google AI Overviews, ChatGPT Search and Perplexity, if AI Mode — already over 1 billion monthly users per Google — pushes its share of all searches well past the current 0.34%?
- Does the 35% citation click bonus Ahrefs measured hold up as far more publishers deliberately optimize for GEO tactics, or does the marginal benefit shrink as the tactic saturates?
- How does the asymmetric robots.txt handling (automated crawlers respect it, user-triggered live fetchers like ChatGPT-User or Perplexity-User largely ignore it) affect publishers' control over their own content in the long run?
Verified Sources & Studies
- GEO: Generative Engine Optimization (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande)Princeton University / Georgia Tech / Allen Institute for AI / IIT Delhi, arXiv 2311.09735, published at KDD 2024The original study coining GEO, with the GEO-Bench benchmark (10,000 queries) and concrete percentage figures per optimization method.
- Google users are less likely to click on links when an AI summary appears in the resultsPew Research Center, July 22, 2025The only study based on real, tracked browsing data (900 users, 68,879 searches) rather than panel surveys.
- AI Overviews Reduce Clicks (Ahrefs study)Ahrefs, analysis based on 300,000 keywords and aggregated Google Search Console dataA time-series comparison of March 2024 vs. March 2025 vs. December 2025 with concrete CTR values per position.
- New data: Google AI Overviews are hurting click-through ratesSearch Engine Land, reporting on the Amsive study (700,000 keywords, 10 websites, 5 industries)Segments by branded/non-branded and position — shows the asymmetry of CTR effects.
- In 2026, Less than One Third of Google Searches Still Send a ClickSparkToro (Rand Fishkin) with Similarweb clickstream data, June 2026The most current available zero-click rate with transparent methodology (Similarweb panel, device split).
- Google zero-click searches reach 68% in early 2026: StudySearch Engine Land, June 2026Contextualizes the SparkToro numbers and separates the effect from Google AI Mode's still-small share (0.34%).
- AI Features and Your WebsiteGoogle Search Central, official developer documentationGoogle's own binding statement on requirements, meta tags and the "query fan-out" mechanics of AI Overviews.
- Overview of OpenAI CrawlersOpenAI, official developer documentationDefinitively defines which bot (GPTBot, OAI-SearchBot, ChatGPT-User) is responsible for training, search, or live fetch.
- Perplexity CrawlersPerplexity AI, official documentationClarifies the robots.txt handling of PerplexityBot vs. Perplexity-User straight from the source.