AI Overviews cite different sources than Google search
Google Search, AI Overviews, and Gemini compared across 11,500 queries in a SIGIR 2026 study
Google Search, AI Overviews, and Gemini cite sources overlapping only 18%, and sites blocking the AI crawler get dropped. A SIGIR 2026 study of 11,500 queries.
Google Search, AI Overviews, and Gemini show different internets for the same query. Matching the sources the three engines cited across 11,500 real queries, only 18% overlapped on average (Jaccard 0.11 to 0.18), and AI Overviews appeared on 51.5% of representative real-user queries, sitting above the organic results[1]. The more striking result is that the 21 popular publishers blocking Google's AI crawler were never cited by Gemini, and even AI Overviews, which have access, cited them less[1].
The evidence is a SIGIR 2026 paper from New Jersey Institute of Technology and others[1]. It covers 14,212 queries in total, an 11,500-query time-invariant benchmark plus time-sensitive sets, collected via SerpAPI and the Gemini API from December 7 to 8, 2025, on mobile from Newark, NJ[2]. It studies Google only, excluding ChatGPT, Perplexity, and Bing Copilot, and is descriptive and limited to one point in time[1]. The authors' released benchmark queries and collected data are the primary source here.
We build a product that measures how often a brand is cited in AI search, so we have an interest in this topic. The figures are reported as written, with the authors' reading alongside. Yesterday's post, AI search knows your product but rarely recommends it, noted that recognition and recommendation are different numbers. This post goes one step further: that recommendation comes from entirely different sources depending on the engine.
What queries trigger AI Overviews
AI Overviews appear more for long, question-formatted, information-seeking queries. They were generated on 65.6% of the full benchmark and 51.5% of ORCAS, the representative real-user set (Section 4.1)[1]. The spread across query types was wide. Complex informational queries (ELI5) hit 94.6% while Amazon retail keywords hit 17.4%. Interrogative queries triggered AIOs far more than non-interrogative ones, and reformatting the same natural-language question as keywords lowered the rate (NQ 86.2% versus NQ Keywords 76.5%).
| Query type | AI Overview rate |
|---|---|
| Complex informational (ELI5) | 94.6% |
| Natural-language question (NQ) | 86.2% |
| Keyword form (NQ Keywords) | 76.5% |
| Representative real-user (ORCAS) | 51.5% |
| Amazon retail keyword | 17.4% |
Table 1. AI Overview rate by query type. Source: Grossman et al. 2026, Section 4.1[1].
Shopping split the same way. Amazon keyword queries were at 17.4%, but comparison queries hit 88.2% and product questions 92% (Section 6). Generative search enters the research and comparison stage more than the purchase stage.
The three engines barely cite the same sources
For the same query, the three engines cited almost non-overlapping sources. Average Jaccard similarity sat between 0.11 and 0.18, and even between AI Overviews and traditional search, only 18% of sources were cited by both (Section 4.2)[1]. More unexpected, AI Overviews and Gemini overlapped the least, despite both being built by Google.
| Engine pair | Jaccard | RBO |
|---|---|---|
| AI Overview to Google Search | 0.18 | 0.23 |
| Gemini to Google Search | 0.16 | 0.21 |
| AI Overview to Gemini | 0.11 | 0.15 |
Table 2. Average overlap in cited sources by engine pair (full benchmark). Source: Grossman et al. 2026, Table 2[1].
The gap is not about source counts. The averages were 9.68 for Gemini, 9.24 for AI Overviews, and 8.75 for traditional search, all similar. And the disagreement between engines far exceeded the disagreement between two runs of the same engine, so it is not explained by built-in randomness. The three engines pick a similar number of sources by different methods.
What gets pushed out of citations
Generative search cited popular sites and institutions less and Google-owned content more. The biggest drop was Reddit, down 27.88 percentage points in AI Overviews versus traditional search (Section 4.3)[1]. AI Overviews pulled more from google.com and youtube.com, and Gemini also favored YouTube. Among the top 1,000 most popular domains (Tranco ranking), the share of queries whose top source fell inside that set was 52.7% for traditional search, 40.0% for AI Overviews, and 32.6% for Gemini. Traditional search weighs source reputation and popularity more.
| Engine | Top source in popular top 1,000 |
|---|---|
| Google Search | 52.7% |
| AI Overview | 40.0% |
| Gemini | 32.6% |
Table 3. Share of queries whose top source is in the top 1,000 popular domains. Source: Grossman et al. 2026, Section 4.3[1].
If the ones pushed out are popular and institutional sites, lesser-known providers get an opening instead. A niche provider that does not rank in traditional search has room to appear in generative search, the same direction as the new-product opening in yesterday's discovery gap. The condition is that the difference between GEO, SEO, and AEO has to be read per engine.
Blocking the AI crawler removes you from AI search
This is the most operational finding. The 21 popular publishers that block the Google-Extended bot in robots.txt were never cited by Gemini[1]. The list includes NYTimes, CNN, BBC, Reuters, Nature, NPR, WIRED, National Geographic, and Scientific American. That part is expected, since Gemini cannot access blocked content.
The surprise is next. These publishers were also cited less in AI Overviews, which do have access[1]. In the regression, sites blocking the Google-Extended bot were significantly less likely to be cited by both AI Overviews and Gemini. Google states that blocking this bot does not affect traditional search ranking, yet the loss in generative search is measurable. The authors call it a self-inflicted visibility loss and suggest reconsidering the block. Since a site's crawler policy changes its AI search visibility, it is worth looking at llms.txt and AI crawler settings from the exposure angle.
The same query shifts on a repeat
AI Overviews shifted more when the same query was asked again. Run twice on the same device and location, the overlap in cited sources was Jaccard 0.66 for AI Overviews versus 0.78 for traditional search (Section 4.4)[1]. A cosmetic edit widened the gap. Changing "what is" to "what's," the overlap (RBO) between original and edited query fell to 0.49 for AI Overviews, a 28.99% decline, versus 13.95% for traditional search and 3.85% for Gemini.
When the sources shifted, the generated text shifted too. There was a strong positive correlation between source overlap and generated-text overlap (AI Overviews r = 0.62, Gemini r = 0.55). The same intent with different wording produces different sources and a different answer, which is why rank optimization becomes unstable. That is the case for not judging on one measurement and instead measuring repeatedly under the same conditions.
High-stakes queries appear more and shift more
The problems stacked on sensitive queries. AI Overviews appeared on 93.8% of queries about politicians, and on debate queries 33.4% of AIO summaries led with an affirmative or negative stance, versus 5.6% for Gemini (Section 4.5)[1]. On political queries, generative search cited government resources like congress.gov less, and the share of sources with questionable credibility ratings was 11.4% for AI Overviews and 15.0% for Gemini, against 10.6% for traditional search. AI Overviews were rare for trending queries (8.1%), which the authors read as a guardrail against pulling misinformation on fast-moving events with no consensus. One case study showed an AI Overview giving the wrong answer to a trending question about a boxing result.
What this study did not see
The authors grant limits. The target is Google alone, so ChatGPT, Perplexity, and Bing Copilot may differ. API collection could not capture per-user variation such as signed-in state. The analysis is descriptive and fixed to one point in time (December 2025), so with generative search changing fast it is safer to read the structure than the exact numbers. To that we add one question: whether such different per-engine source selection converges or diverges over time cannot be told from this single snapshot.
Measure per engine, and repeatedly
Three things carry over to practice. First, AI search visibility is a different game per engine. With the three engines overlapping only 18%, a single optimization that works across all generative search is hard, so you measure and prescribe per engine. Second, blocking the AI crawler is worth reconsidering from the exposure side. It is where content protection and AI search visibility conflict, and blocking led to a self-inflicted loss in generative search. Third, with low consistency, do not judge on one measurement.
The measurement tool we build asks non-brand questions per engine, repeatedly, and records how often a brand is cited at the level of question, engine, and time. The cross-engine gap and the consistency problem this study shows are why it has to be read as a distribution, not a single number. Where that distribution is low for an engine, the groundwork of raising entry and rank in the recommendation census comes before content polishing.
Frequently asked questions
What queries trigger an AI Overview most often?
Long, question-formatted, information-seeking queries. In this study AI Overviews appeared on 51.5% of representative real-user queries (ORCAS) and 65.6% of the full benchmark. Complex informational queries (ELI5) hit 94.6%, Amazon retail keywords 17.4%, and queries about politicians 93.8%.
Do Google Search and AI Overviews really show different sites for the same query?
Yes. Across Google Search, AI Overviews, and Gemini, the average overlap in cited sources was Jaccard 0.11 to 0.18. The two Google-built systems, AI Overviews and Gemini, overlapped the least. Source counts were similar, so the difference comes from selection, not quantity.
Does blocking the AI crawler reduce AI search visibility?
It does. The 21 popular publishers that block the Google-Extended bot were never cited by Gemini, and even AI Overviews, which have access, cited them less. The authors call this a self-inflicted visibility loss and suggest reconsidering the block.
So does GEO optimization work?
This study questions it. Each engine cites different sources, so performing well across several generative engines at once is hard, and AI Overviews shift results on repeat runs (0.66 consistency), making rank optimization unreliable. That said, popular sites get pushed down, which opens room for lesser-known providers, and prior work finds third-party reviews are cited heavily.
References
Summary
- Google Search, AI Overviews, and Gemini cite sources that overlap only 18% on average (Jaccard 0.11 to 0.18). The three engines show different internets.
- AI Overviews appear on 51.5% of representative real-user queries, more often for long, information-seeking questions.
- Generative search cites popular and institutional sites less and Google-owned content more. The 21 publishers blocking the AI crawler had zero Gemini citations.
- AI Overviews are less consistent and more sensitive to small query edits, making rank optimization unreliable.
- The working order is to measure per engine, reconsider AI crawler blocks, and measure repeatedly rather than once.
Check this topic against your own brand
TRAIL Search measures how ChatGPT and Perplexity answer your customers' questions and finds and fixes where the brand is missing, turning this post into a diagnosis. Start with 10 questions, no card required.
More posts
AI agents already know when to search the web
LLM agents know whether they need a tool like web search before answering (AUROC 0.89 to 0.96). A UCSD study used that signal to cut tool calls by 48%.
AI search knows your product but rarely recommends it
AI recognizes a product 99% by name but surfaces it only 3% in discovery queries. A 2025 study of 112 startups found GEO scores did not predict it.
Measuring GEO's revenue impact with a marketing mix model
To measure GEO's revenue effect, turn how often AI answers name a brand into expected exposures per market, then feed it to an MMM. A 2026 paper explained.
Keep reading on this topic
- Calibrating MMM with experiments: ROAS as a parameterCalibrate a media mix model by making ROAS a parameter and setting its prior to an incrementality experiment. Simulations cut ROAS uncertainty about fivefold.
- Cutting LLM agent costs: delegate decisions to a 4B modelCutting LLM agent costs: delegating choices to a small decision model saved 81 to 89 percent, but only with thresholds fitted on the agent's own runs.
- How to build a small LLM: Sheared-LLaMA's 3% compute pathHow to build a small LLM, what LLM pruning is, whether small models are enough for AI agents, and the cost of building one, read through Sheared-LLaMA.
This post is part of the Papers category, which collects all 41 posts on the topic. See all posts in Papers