• AI visibility

Guide · AI search

How do AI research tools get found and trusted through AI search?

By being the research assistant an AI answer names when a student, researcher or librarian asks which tool to use, and by publishing the evidence on citation accuracy, sources and data use that this audience checks before trusting anything. AI research tools face an unusual problem: the general assistants that recommend them, such as ChatGPT and Gemini with their own deep research modes, are also their closest substitutes. The tools that win make the difference easy to verify.

By Underneath Team Guide Updated

The short version

  1. Researchers have moved fast. In Elsevier’s 2025 survey (opens in a new tab) of more than 3,200 researchers, 58% use AI tools in their work, up from 37% in 2024, and 61% use them to find and summarize the latest research.
  2. Trust has not kept up: only 22% of those researchers believe AI tools are currently trustworthy, and 59% name transparency and clear citations as the marker that would build confidence.
  3. Citation accuracy is the open wound. In a 2024 test for systematic reviews (opens in a new tab), 28.6% of the references GPT-4 produced were hallucinated; the Tow Center (opens in a new tab) found eight AI search tools gave incorrect answers to more than 60% of queries asking them to identify the source of news excerpts.
  4. The specialist tools have real scale and simple prices: Consensus (opens in a new tab) reports more than 10 million users from more than 12,500 universities, and Elicit’s (opens in a new tab) paid plans run from $11 to $89 per user per month before custom plans for companies and schools.
  5. Institutions are next. Clarivate’s 2026 survey (opens in a new tab) of 1,876 library responses found 46% of academic libraries at some stage of AI implementation, while students’ AI use for assignments rose from 45% in 2024 to 95% in 2026.

Who uses AI research tools, and what is a user worth?

Students, academic researchers, professionals who review evidence, and the libraries and companies that license tools for them.

Students are the largest group and the fastest adopters. Clarivate’s user research found AI use for school assignments rising from 45% to 95% of students surveyed in two years. They mostly start free.

Academic researchers use AI for literature work. Elsevier found researchers use AI tools to find and summarize research (61%), perform literature reviews (51%) and draft grant proposals (41%). Oxford University Press (opens in a new tab), in a survey of over 2,000 researchers, found 76% had used some form of AI tool in their research.

Professionals in fields such as pharma, healthcare and policy run systematic reviews and evidence summaries, often with budgets. Elicit, for example, sells a Pro plan “for systematic reviews” and says researchers report up to 80% time savings on systematic reviews with its product. Biotech partnering is one place this shows: scientists vetting a potential partner use AI tools for literature reviews, as how AI search shapes biotech partnering explains.

Institutions license tools for everyone. Elicit’s Enterprise plan is described as “for companies & schools,” scite (opens in a new tab) sells a team plan at $250 a month that includes 8 seats, and Elsevier’s Scopus AI (opens in a new tab) is offered to institutions alongside a paid Scopus license.

What a user is worth follows that ladder. A student on a free plan is worth little today and may be worth a lot later. A researcher on Elicit’s Pro plan pays $39 per user per month billed annually. A lab, a company or a university library that licenses a tool for hundreds of people is a contract that renews each year if the tool earns its place.

Where does AI search sit in how researchers find tools?

Near the start: researchers already ask AI about the literature, so asking which tool to use is a short step.

No survey we found asks researchers how they discovered the research assistant they use. The surveys do show where they look for guidance: OUP found 54% would look to academic societies for guidance on AI, 43% to their own institution and 27% to publishers. Those are the kinds of independent, trusted sources that AI answers also draw on.

The bigger shift is that general assistants now do research themselves. OpenAI (opens in a new tab) says deep research in ChatGPT can take 5 to 30 minutes to find and synthesize hundreds of online sources into a report, and Google (opens in a new tab) offers a similar Deep Research mode in Gemini. For a specialist research tool, the assistant is both the channel (where a user may ask “what is the best AI tool for a literature review?”) and the alternative (where the same user may simply ask the literature question directly). Writing tools face the same double role; see how AI writing tools compete with ChatGPT.

Students and researchers are also spread across several assistants. By Similarweb’s 2026 count (opens in a new tab), ChatGPT’s slice of visits to generative AI websites slid from about 76% in June 2025 to roughly 53% by May 2026, as Gemini and Claude picked up share. Libraries are also moving content into those assistants: Clarivate found 19% of libraries overall pursue a balanced approach that includes making library content and services available directly in external services such as ChatGPT, Gemini or Claude.

Which questions do people ask about AI research tools?

Questions about accuracy, fit and access: which tool cites real papers, suits a method, and is licensed.

These sample prompts are our own, written to show how students, researchers and librarians might phrase the question; they are not recorded queries:

  • Task fit: “Best AI tool for screening papers for a systematic review in medicine.”
  • Comparison: “Elicit vs Consensus vs Scite for a PhD literature review.”
  • Accuracy: “Which AI research assistant doesn’t make up citations?”
  • Substitution: “Is ChatGPT deep research good enough for a literature review, or do I need a specialist tool?”
  • Access: “Does my university have a license for an AI research assistant?”
  • Data: “Which AI research tools don’t train on my unpublished manuscripts?”

The accuracy and substitution questions decide the category. If a user believes a general assistant is good enough, a specialist tool loses the sale before it is considered. For how AI answers handle unsupported statements generally, see our article on unsupported claims.

How does an AI recommendation turn into users and licenses?

Through a free account that becomes a habit, then a paid plan, then an institutional license that renews.

Named. An answer to a task or comparison question names your tool. Because research users verify by habit, they often open several tools to compare.

Free account. Elicit and Consensus both let users start free. The test is whether results hold up when the user checks the cited papers.

Paid individual or team plan. Heavier work, such as systematic reviews, pushes users to paid tiers. Elicit’s plans are Plus ($11), Pro ($39) and Scale ($89) per user per month, billed annually; scite’s Basic plan is $20 a month.

Institutional license. Libraries and companies buy for whole groups, with security, data and training controls. Elicit lists “No training on your data by default” as an Enterprise feature. This step can take a long time: Clarivate found 33% of libraries still in the exploration and evaluation stage of AI.

We infer that AI visibility matters most at the first and last steps: it brings in individual users, and it shapes the reputation that a librarian or research office checks before signing a license.

Why would an assistant send a researcher to one specialist tool over another?

Platforms document little; studies point to independent sources; researchers add citation accuracy, coverage and data use.

Documented by the platform. ChatGPT, Gemini, Claude and the rest say nothing public about how they pick a literature-review or citation tool to suggest. OpenAI and Google document what their own research modes do, which is the competition specialist tools must answer.

Observed in studies. In a test of 112 Product Hunt startups, Sharma (opens in a new tab) found ChatGPT recognized 99.4% when asked by name but surfaced only 3.32% in discovery questions, and that the number of other sites linking to a product and community discussion predicted visibility in Perplexity. On software ranking questions more broadly, AI search took 72.7% of its sources from independent “earned” sites in work by Chen and colleagues (opens in a new tab), against 45.4% for Google. We infer that coverage in academic, library and scholarly-publishing sources counts for more than claims on a vendor’s own site. Our study of Wikipedia, outside coverage and AI recommendations tests that idea for brands in general.

Trust factors specific to research tools. This audience is trained to check sources, and the record gives it reasons:

  • Citation accuracy. In the 2024 systematic review test, published in the Journal of Medical Internet Research, GPT-3.5 hallucinated 39.6% of references, GPT-4 28.6% and Bard 91.4%. Those models are now old, but the finding shaped how researchers judge every AI tool.
  • Source attribution. The Tow Center’s test of eight AI search tools found error rates ranging from 37% for Perplexity to 94% for Grok 3 when asked to identify the source of news excerpts.
  • Data use. OUP found only 8% of researchers trust AI companies not to use their research data without permission, and only 6% trust them to meet their privacy and security needs.
  • What researchers ask for. Elsevier found the trust markers researchers want are transparency and clear citations (59%), recency and up-to-date literature (55%) and training on high-quality peer-reviewed content (55%).

A specialist tool’s advantage is that it can show its corpus, its citation method and its accuracy tests in public, in a way a general assistant usually does not. Elicit, for example, states that it searches over 138 million academic papers. Our article on whether citations make AI answers more trusted explains why showing sources alone is not proof.

What does it cost a research tool to be left out?

The first account, and the reputation that later decides an institutional license.

When the question is “which tool should I use for my literature review?” and the answer lists two competitors and a general assistant’s research mode, the user rarely goes looking for a fourth option. The scale of the leaders shows how much is at stake: Consensus reports more than 10 million users, and Elicit says it is trusted by over 5 million researchers. A tool that is missing from answers also misses the word of mouth that follows, because students and researchers recommend tools to each other.

The license stage magnifies the loss. Institutional decisions are slow and deliberate, and a reasonable expectation is that a librarian evaluating options will look at what is said about each tool, including in AI answers. That last point is our inference; no study has measured how libraries use AI answers in procurement.

How does GEO work for an AI research tool?

It makes your tool easy for assistants to describe accurately and for researchers to verify; it cannot promise a recommendation.

For a research assistant aimed at scholars, generative engine optimization (GEO) tends to involve seven pieces of work:

  1. A clear statement of what you search. Name your corpus, its size and update schedule, the fields it covers and what it leaves out, on a public page assistants can cite.
  2. Published accuracy evidence. Describe how you check citations and how often references are wrong, with method and date. Specific, repeatable tests are what this audience trusts.
  3. Independent coverage where scholars look. Library guides, academic society resources, scholarly-publishing newsletters, methods papers and independent reviews. How brands build authority for AI search sets out how that kind of third-party standing is built.
  4. An honest answer to “why not just use ChatGPT?”. A fair page on what a specialist tool does that a general research mode does not, and where the general tool is good enough.
  5. Data and privacy pages. What happens to uploaded manuscripts, whether you train on user data, and what institutions control, in plain words.
  6. Plain pricing and licensing. Individual, team and institutional options in one current place, so an assistant can tell a user whether their library may already pay for it.
  7. Repeated checks across assistants. Put a fixed set of task, comparison, accuracy and access questions to ChatGPT, Gemini, Claude, Perplexity, Copilot and Google on a schedule, and ask new sign-ups how they heard of you. The wider software picture, including subscription revenue, is in our B2B SaaS article.

What don’t we know yet about how researchers pick AI tools?

Nobody has published how researchers settle on a research assistant, or how often an AI answer starts it.

The adoption figures come from publishers and information companies (Elsevier, OUP, Clarivate) that sell research tools themselves. The user counts come from the vendors. The citation-accuracy tests measured general assistants and older models, not the specialist tools, and no independent study we found compares specialist research tools’ citation accuracy with the latest deep research modes. No published study follows an AI recommendation through to a paid plan or an institutional license. For how often people actually open the sources AI cites, see our article on checking sources.

How can a research tool see what assistants tell students, researchers and librarians?

Ask the assistants the comparison and “is ChatGPT enough?” questions your users ask, and read what comes back.

Then audit task, comparison, accuracy and licensing questions across the main assistants and match the results against where your free sign-ups, paid upgrades and library licenses come from, so the gaps that cost the most users and renewals get fixed first. For a second pair of eyes on that audit, talk to us about your research tool’s AI visibility. The generative engine optimization service page explains how we trace which sources shape those answers, then strengthen the corpus, accuracy and licensing evidence librarians check.

Frequently asked questions

Do AI assistants recommend specialist research tools or their own research modes?

There is no published data on how often each happens. General assistants document their own research modes in detail, so specialist tools need equally clear public evidence of what they do better, especially on citation accuracy and coverage.

Does publishing citation-accuracy tests help an AI research tool?

It gives researchers and assistants something specific to check, and researchers say transparency and clear citations are what would raise their trust. Whether it changes how often assistants recommend a tool has not been measured.

Should research tools target students or institutions?

Both, through different evidence. Students respond to recommendations and free access; libraries and research offices need data, security and licensing information. AI answers can reach both, so both kinds of information should be public.

How can a research tool tell whether users come from AI search?

Use three signals together: AI referral traffic, a sign-up question asking how each new researcher heard of you, and regular checks of what assistants say to your key questions. Referral data alone undercounts, because many AI-driven visits arrive as direct traffic.

Sources

Free strategy call

Some questions are easier to answer about your own business.

Bring the one that matters most. On a free 30-minute call we’ll take a first look at it and send you a short written read afterward.

Ask about my businessHow we work