Six outcomes, not one

The organizing idea of this report is that “visibility” is not a single outcome. It is a sequence of six, and each can change independently (Figure 1).

• Discoverable: a page can be crawled, indexed and shown with a snippet.
• Retrieved: an AI system pulls the page into the set of sources it reads before answering.
• Cited: the answer links to the page as a source.
• Mentioned or recommended: the answer names the brand, whether or not it links to it.
• Clicked: a person visits the website.
• Converted: the visit produces a lead, booking or sale.

These outcomes are often blurred together. Recent research shows that a website can lose its citation while the brand is still named, and that a brand can lose its mention while its page is still cited (Makosiewicz & Guan, 2026). A study that measures citations says nothing directly about clicks, and a study that measures clicks says nothing directly about revenue. Throughout this report we keep them apart.

Scope and inclusion rules

We set the following protocol before extracting results:

• Core window: January 2025 to 8 November 2026. Earlier work is used only as a baseline.
• Core evidence: original empirical research with identifiable methods and outcomes. This includes vendor research when it describes its data and method.
• Supporting layer: official search-engine documentation, company statements and expert commentary. These help explain requirements but are not treated as independent tests.
• Excluded from the core: recycled statistics pages, unsupported “ranking factor” lists and promotional case studies without adequate methods.

How we read each study

Every study was read at its original source wherever possible. The unit of extraction was one estimate per row, linked to its parent study. For each estimate we recorded the exact outcome and denominator, the population (country, device, platform), the sampling method, the observation period, the design (randomized, controlled, before-and-after or cross-sectional), and whether the data, code or full tables were published.

Where we could not read the primary document, we say so. Two studies in this review rely on secondary reporting: the randomized experiment by Agarwal and Sen (2026), whose working paper could not be retrieved, and Adobe’s 2026 monthly traffic updates, whose PDF reports would not load. Two BrightLocal studies on local AI search were excluded because their article text did not render and secondary summaries disagreed with each other.

Source retrieval and first-pass extraction were assisted by AI research tools. Every figure in this report was checked against the source text, and several discrepancies between studies’ own headlines and their tables are reported below.

Evidence labels

We use five labels for each claim:

What the evidence base looks like

The 36 sources comprise 31 original empirical studies and five supporting documents. Table 1 summarizes them by design. The full register appears in Appendix A.

Three features of this evidence base matter for everything that follows:

1. It is concentrated. Eleven of the 31 core studies come from Ahrefs, which draws on shared research infrastructure. Publication count is not the same as independent replication.
2. It is mostly American and mostly desktop. Only two studies measure Europe directly (SISTRIX in Germany and Ahrefs in France), and the only randomized experiment used US desktop Chrome users.
3. It is strongest in the middle of the chain. Clicks and citations are well studied. Conversions and revenue are not, and almost nothing addresses small or local businesses specifically.

Five different things called “click loss”

1. Position-one click-through rate on affected queries. Ahrefs compared position-one click-through rate (CTR) for 150,000 keywords that showed an AIO with a counterfactual based on 150,000 informational keywords that did not. The estimated loss was 34.5% between March 2024 and March 2025 (Law & Guan, 2025) and 58% between December 2023 and December 2025 (Law & Guan, 2026a). SISTRIX, measuring more than 100 million German keywords, found position-one CTR of over 27% without an AIO and about 11% with one, a loss of almost 60% (Beus, 2026). Two independent programs, one of them covering Germany, reach nearly the same number. Supported across independent sources

2. Outbound organic clicks per search. The only randomized experiment in this field recruited US desktop Chrome users and used a browser extension to hide AIOs for a random group. On searches where an AIO appeared or would have appeared, outbound organic clicks fell from 0.62 to 0.37 per search, a 39.8% reduction. The chance of a zero-click search rose from 0.54 to 0.73 (Agarwal & Sen, 2026, as reported by Rijo, 2026). This measures every organic click on the page, not just the top result, which is why it is smaller than the position-one estimates. Established within this sample

3. The chance that a search produces any click. Pew Research Center tracked the browsing of 900 US adults. Users clicked a traditional result on 8% of visits to results pages with an AI summary and 15% without one, and clicked a link inside the summary on 1% (Chapekis & Lieb, 2025). SISTRIX found that 57% of German searches without an AIO led to an organic click, against 33% with one (Beus, 2026). Supported across independent sources

4. Click-through rate for affected keywords or domains. Amsive’s analysis of 700,000 keywords across ten client websites found that keywords triggering AIOs lost 15.5% of their CTR on average and 20.0% for non-branded keywords (Guevara, 2025). In France, domains with high AIO exposure lost a median 23.1% of CTR in the nine days after AIOs launched on 22 July 2026 (Begue & Guan, 2026). These are averages across mixed positions and query types, so they sit below the position-one figures.

5. Losses across a whole keyword portfolio. Most queries do not trigger an AIO, so total losses are far smaller. SISTRIX estimated that AIOs cost 6.6% of all organic clicks in Germany, about 265 million clicks a month (Beus, 2026). Across all 963 French domains, the median change was −5.7% in CTR and −8.8% in clicks (Begue & Guan, 2026).

Other reasons the numbers differ

Outcome definition explains most of the spread, but not all of it. Four other design choices matter:

The Ahrefs versus Seer “disagreement”

The sharpest apparent contradiction is between Ahrefs, which reported in February 2026 that position-one CTR losses had grown from 34.5% to 58%, and Seer Interactive, which reported in April 2026 that CTR on AIO queries had recovered. Seer itself acknowledged the disagreement (Seer Interactive, 2026).

Lined up by month, most of the contradiction disappears. Ahrefs’ latest measurement is December 2025. In Seer’s data, December 2025 is also the low point: organic CTR on queries with an AIO was 1.31%, against 3.16% without one. Seer’s recovery comes from January and February 2026, when CTR on AIO queries rose to 2.36% (McDonald et al., 2026). Ahrefs did not measure those months. The two programs agree on the trough and differ on the months after it.

Two months of recovery from one program, with a model Seer itself describes as directional, is not yet a trend. Unresolved contradiction

When the denominator grows

Seer’s study includes a small table that illustrates a problem affecting many dashboards (Figure 5). For brand-cited organic results, impressions rose from 15.8 million in September 2025 to 39.5 million in November, an increase of 150%. Clicks fell from 398,798 to 301,783, a decrease of 24%. Click-through rate fell from 2.52% to 0.76%, a drop of 70% (McDonald et al., 2026).

The clicks loss is real, but it is a third of the size the CTR change suggests. Most of the CTR “collapse” came from the brand being shown on many more searches. Seer states that it cannot tell, without account-level work, whether this reflects Google changing where AIOs appear or the brands’ own SEO activity.

When one account dominates

Aggregate CTR is weighted by impressions, so a single large account can move the whole number. In Seer’s informational queries where the brand was not cited, one account produced 47% of all impressions at a CTR of 0.20%. With that account included, the segment’s CTR was 0.94%. Without it, the CTR would have been about 1.61% (McDonald et al., 2026). Two brands, Salomon and MongoDB, made up 35% of impressions in the no-AIO informational segment.

When the label comes later

Seer labels each query by whether it showed an AIO in its most recent snapshot, then applies that label to every month (McDonald et al., 2026). A query that gained an AIO in October is counted as an “AIO query” in January as well. This can make AIO queries look as if they performed differently before AIOs existed.

When citations simply come and go

AI answers change from one day to the next, even when nothing on the website changes. Ahrefs asked the same questions of five AI tools every day for 20 days and tracked about 37,500 citations (Figure 6). Half or more of the websites cited on day one were gone by day two: 50% on AI Overviews, 58% on AI Mode, 59% on ChatGPT and more than 80% on Gemini and Copilot (Makosiewicz & Guan, 2026).

Most of that churn reverses on its own. About three-quarters of citations that dropped returned within the 20 days, and about half returned within three days. Only 20% to 27% disappeared for good. Some apparent changes were not changes in source choice at all: Gemini gave no citations in 64% of answers and Copilot in 53%, so many “lost” citations were answers that cited nobody (Makosiewicz & Guan, 2026).

A diagnostic checklist

Before treating a decline as real, check which of four explanations fits:

We also found inconsistencies inside widely cited studies:

• SparkToro’s 2026 post cites a 2024 US zero-click rate of 60.45%, while its own 2024 study reported 58.5% (Fishkin, 2024, 2026).
• Ahrefs’ France study reports a panel-wide CTR change of 3.4% in one section and a median of −5.7% in another (Begue & Guan, 2026).
• Seer’s text describes cited CTR as “2–5x” uncited CTR, while its table shows ratios of about 1.2 to 3.9 (McDonald et al., 2026).

None of these changes the direction of the findings, but each would change a forecast built on them. Established within this sample

Factual questions favor official sources

Ahrefs analyzed 66.6 million free, logged-out ChatGPT responses and about 15 million of the web searches ChatGPT generated to answer them (“fan-out queries”) in August and September 2026 (Linehan, 2026a). The word “official” appeared in 10% to 13% of those searches. In a separate test of 20,000 US prompts, at least one official source was cited in 84.4% of answers to factual questions and 68.3% of answers to comparison questions. Official websites made up 63.4% of all citations for factual questions and 46.4% for comparisons (Figure 7).

Independent research shows the same tilt toward authoritative sources. In Pew’s data, government websites made up 6% of sources in Google’s AI summaries but only 2% of sources in standard results (Chapekis & Lieb, 2025). In a peer-reviewed study of 55,393 trending queries, Xu et al. (2026) found that user-generated content made up 14.2% of AIO references against 49.9% of first-page results. Supported across independent sources

AI systems also reuse a brand’s own website as the source of facts about that brand. When ChatGPT, AI Overviews or AI Mode named a brand, the brand’s own site stayed cited from one day to the next about 75% of the time, against 33% to 40% for other sources (Makosiewicz & Guan, 2026).

Evaluative questions follow off-site discussion

For mentions rather than citations, the pattern is different. Ahrefs compared 10,000 US brands’ visibility across 260 million non-branded prompts in ChatGPT, AI Mode and AI Overviews with how often each brand was discussed elsewhere (Linehan, 2026b). Every type of off-site mention was associated with AI mentions (Figure 8). The strongest association was with Reddit threads that rank in Google (Spearman ρ = 0.75), followed by mentions on other domains and web pages (0.70), news (0.67), YouTube (0.64) and TikTok (0.60).

Two cautions apply. First, these are correlations, and Ahrefs states that it cannot say mentions cause AI visibility. Second, much of the association reflects brand size. When brands were compared within bands of similar search demand, the web-mention correlation fell from 0.70 to between 0.34 and 0.54 (Figure 9). Bigger brands get discussed more and get mentioned more; part of what looks like a mentions effect is a size effect.

Sources shift quickly, and do not simply mirror Google

AI source preferences move fast. Between March and September 2026, the share of ChatGPT answers citing Reddit fell from about one in five to about one in 43, and Wikipedia from about one in seven to about one in 35 (Linehan, 2026a). In Google’s own AI Overviews, 29.8% of cited domains did not appear anywhere on the first page of results for the same query, and only 48% appeared in the top ten (Xu et al., 2026). Ranking well helps (pages ranking in Google were about seven times more likely to stay cited for a week in AI Overviews and AI Mode), but it is not the same as being chosen (Makosiewicz & Guan, 2026).

A source-selection framework

Drawing these findings together, we propose the following working model. It is a synthesis, not a tested rule. Plausible synthesis

What is required

Google’s own documentation states that there are no additional requirements, and no special optimizations, for appearing in AI Overviews or AI Mode. A page must be indexed and eligible to be shown with a snippet, and the usual SEO best practices still apply (Google Search Central, n.d.). Site owners can limit what is used through nosnippet, max-snippet and noindex. This is documentation, not an independent test, but it defines the entry ticket.

What happens when you add schema

The strongest test of a popular technical tactic is Ahrefs’ difference-in-differences study of 1,885 pages that added JSON-LD structured data between August 2025 and March 2026. Each was compared with matched control pages that had similar citation levels but never added it (Linehan & Guan, 2026). Thirty days after the change:

• AI Mode: +2.4% relative to controls, statistically indistinguishable from zero.
• ChatGPT: +2.2%, statistically indistinguishable from zero.
• AI Overviews: −4.6%, a small but statistically significant decline that the authors could not attribute to schema.

The study has real limits: every page was already heavily cited (100 or more AIO citations), schema types were pooled, and the follow-up lasted only 30 days. It shows that schema is not a lever for pages that are already visible. It does not show that schema is useless for new or uncited pages, or for rich results in ordinary search. Established within this sample

How widely the “AI-ready” signals are actually used

The 2025 Web Almanac crawled more than 16 million websites in July 2025. Only 2.1% had an llms.txt file, which Google says it does not use, and 39.6% of those files were generated automatically by a single WordPress plugin. GPTBot was named in 4.5% of desktop robots.txt files (Chulwuma et al., 2026). Implementation data like this shows what sites do. It cannot show whether doing it helps.

What content changes can do

The academic evidence on generative engine optimization (GEO) is more cautious than marketing claims. A critical survey of 45 GEO studies published between 2023 and July 2026 found that the widely repeated claim that GEO can raise visibility “by up to 40%” comes from a single metric in a simulator where five documents were already in the model’s context. The survey rejects it as a general claim. In one benchmark, only 3 of 54 method-and-domain combinations produced a significant positive effect (Martinez, 2026).

Controlled experiments find small, conditional effects. In a causal audit using frozen search results, rewriting a document in a structured format rather than prose added about half a citation per answer (95% confidence interval +0.20 to +0.84), but the effect on whether the document was cited at all was inconclusive. Fifteen percent of citation decisions changed simply when the same answer was regenerated (Selvam & Ghosh, 2026). An observational study of about 21,000 citations found that pages with statistics, definitions and comparisons were more heavily used in answers than pages formatted as Q&A, but its measure is a proxy and its findings are descriptive (Zhang et al., 2026).

AI-written content

Ahrefs’ two studies of AI-generated content point the same way: detector-classified AI-heavy pages rank throughout the top ten, but perform somewhat worse. Pages with less than 20% AI-detected text made up 55.9% of position-one results and 51.1% of position-ten results; 49.3% of low-AI pages were indexed against 40.4% of very-high-AI pages (Law & Guan, 2026b). An earlier study of about 600,000 pages found a correlation of just 0.011 between AI share and ranking (Ong, 2025). Detectors are imperfect and AI use is confounded with site quality, so the safest reading is the authors’ own: quality, not the tool, appears to drive performance.

A prerequisite-versus-growth matrix

Plausible synthesis

Losses depend on what the visitor came for

SISTRIX’s German data shows how unevenly AI Overviews affect different kinds of sites (Figure 10). Health-information sites lost around 30% of their Google clicks, and parenting and baby portals over 24%. Wikipedia lost 5.56%, or 31.6 million clicks a month. At the other end, Amazon lost 1.73%, recipe site Chefkoch 1.07%, price-comparison site Idealo 0.85%, Booking.com 0.46% and weather site wetter.com 0.18% (Beus, 2026).

The same pattern appears elsewhere. In France, health was the most exposed sector, with an AIO rate of 22.2% and a CTR decline of 20.4% (Begue & Guan, 2026). In the randomized US experiment, the click loss was concentrated on informational queries; for navigational and transactional searches, the estimated effect could not be distinguished from zero (Rijo, 2026). Xu et al. (2026) found AIOs on 64.7% of question-form queries against 9.5% of others.

What winning sites have in common

Zyppy classified more than 400 websites that won or lost Google traffic around the December 2025 core update (Shepard, 2026). Winners were far more likely to have proprietary assets such as databases, tools or original reviews (92.9% versus 57.1%), to let users complete a task (83.7% versus 50.2%), and to offer a product or service (70.2% versus 34.6%) (Figure 11). Sites with four or five of the features won 68% to 70% of the time, against 13.5% with none (Figure 12).

The Zyppy correlations are modest (ρ = 0.21 to 0.39), the sites were selected rather than sampled, and features were coded by hand without published reliability checks. On its own, the study is suggestive. Combined with the SISTRIX category pattern and the experiment’s intent results, it supports a consistent explanation. Plausible synthesis

An answer-substitutability framework

We propose classifying pages by two questions: can a short answer satisfy the need, and does the user need to act on a site? (Figure 13).

For most small and medium businesses, this is encouraging. A plumber, dentist, accountant, retailer or tour operator exists to complete a task: a booking, a quote, a purchase, an appointment. These businesses are typically in the “most resilient” quadrant for their core pages, and in the “high exposure” quadrant only for the informational blog content they publish to attract visitors. Whether adding a tool or booking function causes resilience has not been tested. Hypothesis requiring new data

The scale gap

SE Ranking analyzed Google Analytics data from 101,574 websites from January 2025 to April 2026. AI platforms accounted for 0.32% of all website visits in 2026, about one in every 312. Organic search accounted for 42.75%, and Google alone sent 134 times as many visitors as all AI platforms combined (Deda, 2026; Figure 14). ChatGPT sent 74.78% of AI referrals, followed by Gemini (11.56%), Perplexity (7.23%), Copilot (3.51%) and Claude (2.62%).

Ahrefs’ analysis of 81,947 sites in mid-2025 found the same order of magnitude: AI referrals made up about 0.25% of an average site’s traffic, and Google sent about 210 times as much as the three largest AI assistants (Linehan & Guan, 2025). Clicks that come from Google’s own AI Overviews and AI Mode are counted as Google search traffic in these studies, so “AI traffic” here means visits from standalone assistants. Supported across independent sources

The value gap is reversing

Adobe Analytics, which measures more than a trillion visits to US retail sites, reports AI referral performance relative to other traffic (Figure 15). During the 2025 holiday season, AI referral traffic to retail sites rose 693% year on year and converted 31% better than other traffic. AI-referred shoppers were 33% less likely to leave immediately and spent 45% more time on site (Pandya, 2026). Trade press reporting of Adobe’s later releases puts the conversion advantage at 60% in July 2026 and 90% in September 2026, up from 9% worse in August 2025 (Fernandez, 2026; Haleem, 2026).

Other evidence on value is weaker:

• Ahrefs found that AI search produced 0.5% of its own website’s visitors but 12.1% of its sign-ups. This is one software company’s site (Stox, 2025).
• Semrush reported that an AI search visitor is worth 4.4 times a traditional organic visitor, but did not publish the sample, period or calculation behind the figure (Handley, 2025).
• Google states that total organic clicks have been “relatively stable” and that clicks from pages with AI features are “higher quality,” but has not published the data or method behind either statement (Reid, 2025).

Our assessment: AI referrals converting better than average is supported across sources in direction. The size of the advantage is unresolved and likely varies by sector. Adobe’s figures are relative only, apply to large US retailers, and may reflect who uses AI assistants as much as what the assistants do. Supported across independent sources (direction) Unresolved contradiction (size)

The exchange is uneven

AI companies’ crawlers read far more than they send back. In July 2025, Cloudflare recorded about 38,066 page crawls by Anthropic’s crawlers for every visit referred back, about 1,091 for OpenAI, about 195 for Perplexity and about 5.4 for Google (Tomé, 2025; Figure 16). Cloudflare cautions that native apps often send no referrer, so AI ratios may be overstated. In the same data, 79% of AI crawling was for model training, 17% for search and about 3% for user-initiated actions.