Does ranking highly on Google secure AI citations?
Does a high Google ranking mean the AI summary will cite your website? Google calls this summary an AI Overview. In a paper accepted at a peer-reviewed conference, researchers found that traditional Google results and AI Overviews shared, on average, only 18% of their combined sources. DijitalPi commentary: Report search rankings and citations in AI summaries separately. The study used English queries collected from a US location. It measured neither clicks, website traffic nor sales, and does not establish results for Turkish searches or customer behaviour in Türkiye.
ACADEMIC WORK REVIEWED
How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews
ResearchersRiley Grossman, Songjiang Liu, Michael K. Chen, Mike Smith, Cristian Borcea, Yi Chen
arXiv / ACM SIGIR 2026 · 30 April 2026 · v1; accepted at the conference · Peer-reviewed conference paper · Source language: English
English review: 16 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
For the same search, do traditional Google results, AI Overviews and Gemini cite the same websites?
The researchers examined which sources these three search formats selected and how consistently they selected them. They also investigated which queries produced AI Overviews. This allowed them to assess the relationship between appearing in traditional search and being cited in an AI answer.
Explaining the context · DijitalPi commentary
DijitalPi commentary: A visibility report containing only traditional search rankings may leave the AI answers your customers encounter insufficiently monitored. The practical decision is whether to track source selection separately across query groups. Being included in a source list does not itself tell you whether a customer clicked a link or made a purchase.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The study made an observational, descriptive comparison covering 14,212 queries in total. It was not a randomised experiment. The core dataset contained 11,500 time-independent queries across 9 subsets. These included 5,000 real-user queries from ORCAS, while 3 subsets were generated using Gemini 2.5 Flash. Additional sets covered political queries about US Congress members and 863 Google Trends topics. Core data were collected on 7–8 December 2025 and trending-query data on 19 December 2025. For English queries, traditional Google results and AI Overviews were retrieved through SerpAPI, simulating a mobile device in Newark, New Jersey, USA. The researchers used Gemini 2.5 Flash through its API with access to Google Search enabled. An API lets one software system send requests to another and receive answers.
Understanding the method · DijitalPi commentary
The researchers examined only the first page of traditional Google results. They collected AI Overviews and traditional results using the same timing, device and location conditions. Source comparisons covered the 7,439 queries for which all three systems returned sources. The outcomes were source lists and how often an overview appeared. Users’ clicks and purchases were not tracked.
Jaccard measures the shared URLs in two lists as a proportion of their combined unique URLs: 0 means no shared sources, while 1 means both lists contain the same sources. The researchers also used a measure that gives greater weight to agreement near the top of the lists, repeated queries and tested small wording changes. Traditional Google Search provided the comparison for the website-characteristics analysis. Because the dataset included only sources retrieved by at least one system, those results describe relative source selection across systems rather than the entire web.
03 / RESEARCH FINDINGS
Traditional results and AI answers shared a limited proportion of their sources.
The researchers found average source similarity between pairs of systems ranging from 0.11 to 0.18. On average, only 18% of the combined sources from AI Overviews and traditional results appeared in both lists. AI Overviews and Gemini were the least similar pair. AI Overviews appeared for 51.5% of ORCAS real-user queries and 65.6% of the full core dataset. The rate was 17.4% for product keywords, 88.2% for product comparisons and 92% for product questions. Across two runs of the same query, source similarity was 0.78 for traditional search, 0.66 for AI Overviews and 0.46 for Gemini. Blocking Google-Extended was associated with fewer citations in generative search systems. This cannot be interpreted as an effect caused by blocking.
Shared sources in traditional Google and AI Overviews
18%The average shared proportion of the unique sources in the combined lists. This is neither a click-through rate nor the probability that a website appears in AI answers.
Source similarity when repeating the same query
0.78 / 0.66 / 0.46Jaccard values for traditional Google, AI Overviews and Gemini respectively. Higher values indicate greater overlap between the source lists from two runs.
DijitalPi commentary: Combining product names, comparisons and product questions into a single visibility score could conceal distinctions between query types. Consider separate groups and repeated checks to monitor where your site is cited in the research process. Do not base a Google-Extended blocking decision on a single missing citation either. The study did not test whether removing a block improves visibility.
Low source overlap does not establish an advantage or failure unique to AI: the researchers’ comparison of traditional Google and Bing results also showed that source selection varies between search systems. Fluctuating source lists make it difficult to attribute visibility observed after a content change solely to that change. API collection prevented investigation of personalisation, and the authors note that signed-in users may experience somewhat greater consistency across repeated searches.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The study shows that traditional Google rankings do not directly represent AI Overview or Gemini source lists, and that checking a query once provides limited information. The examined text is arXiv v1 of a peer-reviewed conference paper accepted at ACM SIGIR 2026. Findings are limited to Google’s ecosystem, English queries collected from a US location and descriptive observations from a particular period; Turkish searches and Türkiye were not tested. Clicks, traffic and sales were not measured. The research was supported in part by the NSF, the NIH’s National Center for Advancing Translational Sciences, the Martin Tuchman’62 Chair Endowment and the Leir Foundation. The authors are university affiliated. No company affiliation is reported, and no competing-interest declaration was found in the examined source.
Our open question at DijitalPi: what differences in source selection might repeated checks of your Turkish product, comparison and local-service queries reveal that your traditional search report currently misses?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: product searches and comparison questions
An illustrative home appliance retailer tracks rankings for product names but leaves customer comparison questions out of its reports.
Does appearing in a product search mean I also appear in a comparison answer?
Open the recommendation for example 01+
Track product names and comparison questions as separate query groups.
Create a list of product names, comparisons and usage questions for the same product family. Record traditional results, whether an AI Overview appears and its source URLs separately for each query. Repeat checks with the date, device and location recorded. The paper’s query-group rates are not expected visibility or sales rates for your shop’s Turkish searches.
EXAMPLE 02
Local services: a source missing from one check
An illustrative repair business is considering an immediate content change after failing to appear in an AI summary for a local search.
Does being absent from one search mean my visibility has fallen permanently?
Open the recommendation for example 02+
Compare repeated checks of the same query before deciding on changes.
Keep the service and location wording fixed and record source lists at different times. Initially hold the device and location constant. Compare devices separately. Distinguish an absent overview from an overview that appears without citing the business. Repeated checks may help reveal local variation. The research did not measure customer demand in your area.
EXAMPLE 03
B2B software: reviewing a crawler restriction
An illustrative B2B software company is considering whether removing a Google-Extended block would lead to more AI citations.
Will removing the block get my site into source lists?
Open the recommendation for example 03+
Do not treat this study as a guarantee of what removing the block would achieve.
Record the Google-Extended rule in robots.txt and the reason for blocking it with your technical team. Repeatedly monitor sources for the same queries across Google, AI Overviews and Gemini under the current policy. If the policy changes, retain before-and-after records and account for variation in search outputs themselves. The study found an association between blocking and citations. It did not test the effect of removing a block.
TRY IT WITH YOUR TEAM
Track source visibility separately.
Group searches relevant to your business by what the customer wants to do and prepare a monitoring plan that repeats source checks using the same device and location. Report citations separately from clicks and sales, which require their own data. This monitoring plan is a proposal that DijitalPi has not yet implemented.
When visits from search change, assess your content plan using data from your own site. SEOMACHINE reviews search queries and page performance to help identify which content needs improvement. Explore SEOMACHINE ↗
Riley Grossman, Songjiang Liu, Michael K. Chen, Mike Smith, Cristian Borcea, Yi Chen New Jersey Institute of Technology, Newark, NJ (Riley Grossman, Songjiang Liu, Cristian Borcea, Yi Chen); Nanyang Technological University, Singapore (Michael K. Chen); Indiana University Bloomington (Mike Smith).
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences