There’s been a shift in how people search for answers, and this shift has direct implications for researchers’ work. When students or fellow researchers have questions about a topic you’re an expert in, they’re now less likely to open Google Scholar or knock on your office door. Instead, they’re more likely to type their questions into ChatGPT or Google AI Overviews and receive summarized answers, complete with sources selected by the machine—not by them.
If your research findings aren’t included among the sources selected by these systems, your work risks never reaching the readers who actually need that information—even when it has been published in a highly reputable journal. In an ecosystem like this, scientific publications no longer face only the issue of visibility in academic search engines, but also the question of whether a study can be discovered, retrieved, selected, and then used by generative AI systems to generate answers.
Therefore, researchers’ efforts to increase the visibility of their work through various digital channels are becoming increasingly relevant. Science communication, knowledge translation, and the distribution of information via digital platforms can help expand the possibilities for a study to be discovered and utilized beyond the academic community. However, who actually determines which studies are more visible and subsequently used by generative AI?
New Data, but It Needs to Be Read Carefully
A report released by Meltwater in collaboration with LinkedIn claims to have analyzed 9.5 million answers from six AI models—namely Copilot, Google AI Mode, Google AI Overviews, Claude Sonnet 4, ChatGPT-5, and Gemini 2.5 Pro—to identify which websites are most frequently cited in professional topics.
For the record, this report is not a peer-reviewed academic study, but rather an industry report compiled by a media intelligence company and published on LinkedIn’s marketing blog, Meltwater. Its methodology relies on a proprietary tool called Meltwater GenAI Lens, which is not publicly available, so it cannot be replicated or independently audited.
Nevertheless, I am more interested in examining the claims in the report by comparing them with independent academic literature, to ensure that the implications for researchers—who wish for their work to reach peers, students, and practitioners—remain evidence-based.
What the report claims, and what academic research confirms
In its report, Meltwater identifies five patterns that set LinkedIn apart from other social media platforms and networks.
- LinkedIn is cited as the second most frequently cited source by AI for business-related queries, after YouTube, with its citation share increasing by 26% over the four-week observation period.
- 75% of citations come from individual profiles, not official company pages, meaning individual experts are cited more frequently than institutional accounts.
- Well-structured content, featuring subheadings, bullet points, and concrete figures, is cited far more frequently: the most-cited articles in their sample used bulleted lists, and 92% used clear subheadings.
- LinkedIn far outperforms other similar platforms when it comes to professional topics.
- Fresh and original content appears as a reference far more often than older content or reposts.
So, how can these five claims be tested as patterns consistent with the findings of empirical research? In an early study on Generative Engine Optimization (GEO: a method for structuring content to make it more likely to be cited by AI answer engines), various writing strategies were systematically tested across hundreds of questions spanning various topics. The researchers found that adding source citations, direct quotes, and concrete statistics can increase a piece of content’s visibility in AI-generated answers by more than 40%. These findings reinforce the claims in the Meltwater report regarding the importance of structure and concrete data.
What Meltwater’s findings on LinkedIn reveal about individual experts’ profiles also aligns with a long-studied field: knowledge translation—the process of disseminating scientific evidence so that it is actually utilized by practitioners and the public. Several systematic reviews in the health sector indicate that active and consistent personal accounts of experts on social media tend to be more effective at disseminating evidence compared to formal, rarely updated institutional accounts.
A study of more than 8,500 applied researchers in Germany also found a correlation between activity on Twitter (X)/LinkedIn and bibliometric indicators such as visibility and inter-researcher connectivity. Writing as an individual with expertise has proven relevant to the dissemination of evidence even before the advent of generative AI. AI merely adds a new channel to an already familiar pattern.
But there’s an important caveat the report doesn’t mention
An early study on arXiv from 2024 found that large language models (LLMs) don’t just mimic human citation patterns. They amplify them with a sharper bias toward heavily cited works. The Matthew effect: what’s already popular becomes even easier to find, while new works or those by unknown researchers become increasingly difficult to discover. Consequently, the advantage of individual experts using LinkedIn over institutional accounts, as highlighted in the Meltwater report, isn’t necessarily felt equally by all individuals. Those who truly benefit may be individuals who already had a solid reputation from the start and an established expert persona. It is this complexity that needs to be critically examined in their report.
This concern is reinforced by another study on the use of LinkedIn by academics in the UK, which found that high activity on the platform is concentrated among senior male professors from better-resourced institutions with a focus on industry collaboration, while female researchers and researchers from underfunded universities are underrepresented. The same study concluded that LinkedIn activity data “risks disproportionately representing academic research” if used as a measure of impact.
Not to mention if we raise other questions regarding the use of languages other than English, such as Indonesian or others. There is a high likelihood that this data relies on English-language queries from the Global North market context, as AI systems tend to amplify sources that are already “dominant” on a global scale.
So, what should researchers do?
There are several sensible steps—not as tactics to “outsmart the algorithm”, but as a natural extension of science communication practices that have long been recommended.
- First, occasionally write a clearly articulated summary of your own research findings in the public sphere, complete with subheadings, key points, and concrete figures—as a supplement to, not a substitute for, publications in reputable journals. This aligns with the requirements of many research funders, who now mandate a plain language summary alongside scientific articles.
- Second, consider writing in your personal capacity as an academic, with your position and affiliation clearly stated in your profile, alongside official institutional posts. This is not to build a personal image, but because the voice of an identifiable expert appears to be more easily traceable, both by human readers and by machines.
- Third, update older writings with the latest data rather than leaving them untouched for years. Fresh sources are more likely to be cited by both AI and peers seeking up-to-date information.
- Fourth, encourage journals, faculties, and funding agencies in Indonesia to actively support dissemination in both Indonesian and English, rather than leaving it entirely to individual initiatives. If the discovery of scientific evidence through generative AI becomes a new primary channel, existing language and resource gaps risk being widened—not narrowed—by this technology, unless there is a conscious effort to correct them.
So, after taking these steps, will your research be more easily discovered by AI systems—which have now become the primary intermediaries for the circulation of knowledge? How can we work to reduce disparities in the translation of knowledge? This discussion remains a compelling topic because the race among AI platforms continues to intensify even as we speak. What are your thoughts on this as a researcher?