For most of the history of scholarly publishing, the definition of “end user” was stable enough that no one thought to question it. The user was a researcher, a student, a clinician, a librarian: a human being who navigated a platform, read content, and formed their own judgments. Everything downstream of that assumption, like platform design, metadata strategy, discovery architecture, and usage reporting, was built around it. 

That assumption is no longer sufficient. The shift isn’t just in where users begin their research, it’s in who “users” even refers to, what they’re looking for when they arrive, and what it means to successfully serve them. Drawing on recent conversations with publishers implementing AI discovery tools, library professionals watching student behavior change in real time, technologists building connector infrastructure, and standards bodies working to make usage measurement coherent again, I see a picture emerging that has a clear center of gravity: the user has changed faster than the systems designed to serve them, and the gap between the two is where most of the difficult work is happening. 

A User That Doesn’t Click 

Let’s start with the most concrete version of the change. In the first ten weeks of 2026, roughly half of all inbound traffic to the Silverchair Platform was not coming from human users in any traditional sense; it was bots, crawlers, and automated systems, some indexing content for AI-assisted search, some operating as user agents on behalf of people using chatbots, some attempting to train large language models on whatever they could reach. 

Publishers have been aware of bot traffic for years. What’s different now is the composition and the intent. A significant portion of today’s automated traffic is doing something that looks a great deal like what publishers actually want: finding content, surfacing it where researchers are making decisions, and routing engaged users back to the source. Among publishers on the Silverchair Platform, the JAMA Network was one of the first (in early 2024) to implement a full block on AI crawlers before eventually revising that position in the fall of 2025.

The shift came from working through a meaningful distinction: crawlers that train on content (specific user agents that consume content to improve models with no clear benefit to publishers and no mechanism to incorporate corrections or retractions) are genuinely different from search and user agents, which index and retrieve content in ways that can drive referral traffic. After allowing a curated set of user agents back in, JAMA saw institutional traffic, healthy time-on-site, and GPT referrals climbing immediately toward the same level occupied currently by Google Scholar. 

The implication isn’t that all automated traffic is good. It’s that “should we allow bots?” has become too blunt a question. The meaningful question is which kinds of automated users, with what access conditions, serving which use cases, and that requires engaging with the specifics rather than treating all non-human traffic as a threat. Lou Peck, CEO of The International Bunch, put the strategic reorientation plainly: “Stop treating AI traffic as noise and something to be blocked. Start focusing on generative engine optimization: ensuring your brand and your content are trust anchors, because that’s what AI models rely on.” Look at the SEO analogy: publishers who engaged with Google early, on terms they could influence, were better positioned than those who resisted until the norms had calcified. There’s a reasonable case that AI discovery is at a similar inflection point now. 

The Human User, Sorted 

The automated-traffic story is real, but it can obscure something equally important: the users who are still coming to publisher platforms are more engaged, not less. The data from publishers who have deployed AI discovery tools tells a counterintuitive story about what AI is doing to site traffic. 

Oxford University Press (OUP) has seen average session duration increase by roughly 30% even as overall referrals have declined, a pattern many publishers can likely recognize. The click-through rate from OUP’s AI-assisted search follows a similar arc: 14% on an initial query, rising to 40% as users refine through conversational follow-up, compared to 9% for traditional keyword search. Tanya Laplante, OUP’s Director of Product Platforms, describes this as a sorting effect: “The users who are continuing to come to us are those really engaged researchers who might be getting an answer from ChatGPT, but are then saying to themselves, I need to go and look at the versions of record and peer-reviewed content and make sure that what I’m seeing is accurate.” 

What seems to be happening, at least for publishers with strong brand equity, is less a loss of traffic than a concentration of it: casual searches are increasingly handled by general-purpose AI tools, while users arriving at publisher platforms are arriving with higher intent. Whether that’s a net gain depends on the publisher’s revenue model, but it’s a meaningfully different situation than simply losing users to AI. (I shared my thoughts on what this means for Marketing teams in another recent article.) 

The flip side is that this sorting concentrates value in publishers who have earned trust and creates urgency around demonstrating it. Tasha Mellins-Cohen, Executive Director of COUNTER Metrics, positioned the current moment squarely in the context of earlier syndication decisions: “If you make agreements that benefit you today, you won’t be in a position in five years’ time going cap in hand to these groups asking how much usage you’re getting for your content on their platform.” The history of publisher relationships with EBSCO, ProQuest, and Google Scholar — often built on terms that left publishers with limited visibility into what the syndication was actually worth — is a cautionary precedent. The AI moment is a chance to negotiate before the norms are set, not just around reporting, but for the longevity of your brand. As Paul Gee, Associate Publisher & VP of Product at the JAMA Network, put it: “If your brand isn’t known in the future through some of these tools, where will it be known?” 

What Trust Actually Requires 

Across clinical publishing, academic libraries, large trade publishers, small societies, and beyond, conversations about AI and the user keep arriving at the same word: trust. But what trust requires looks different depending on who’s asking. 

For the American Academy of Pediatrics (AAP), building an AI assistant around its Red Book (the standard reference on pediatric infectious disease), the trust requirement is essentially non-negotiable. When AAP surveyed its members, 90% said they would use an on-platform AI tool as long as it drew exclusively from the Red Book. A RAG-based model grounded in 80-plus years of evidence and restricted to vetted content is more than a product feature; it’s the condition under which the tool is professionally usable at all. 

For university libraries, the picture is more contested. Jane Jiang, Director of Libraries at Union College of Union County, described a pattern of young researchers, many of them first-generation college students, beginning with AI tools in large numbers. “Many students are absolutely starting with AI tools first now,” she said, using them to brainstorm and get oriented before they know what to look for. AI lowers the barrier to entry at exactly the moment that barrier feels most daunting. But students are arriving at the reference desk with polished AI-generated overviews and without a clear understanding of whether the sources actually say what the AI claims. Teaching AI literacy (how to evaluate, verify, and trace outputs back to primary sources) has become a core part of the library professional’s role. 

The trust challenge for publishers sits between these two poles. General-purpose AI tools don’t know the difference between a peer-reviewed article and a preprint, a current version and a retracted one, a replicated finding and a preliminary result. Andrew Smeall, VP of Product Innovation at Sage, described this as the core argument for publisher-built connectors: “The AI chat experience right now is still a little bit of a false friend, where it gives such plausible-sounding answers, but the answers are not well-grounded in the scholarly record.” Grounding AI responses in publisher content with proper provenance and metadata is now both a business case as well as a response to a genuine quality problem in how AI currently handles knowledge. 

The User No One Designed For 

The most forward-looking version of the user question involves a type of “user” that publishing platforms weren’t built to serve at all: the AI agent acting autonomously on behalf of a researcher. 

Model Context Protocol (MCP), the open standard that allows AI agents to connect directly to external data sources and retrieve content programmatically, has moved fast. MCP went from 100,000 downloads in its first month to 97 million monthly downloads within a year. Developers are building toward a world where AI tools don’t just assist research but conduct significant portions of it. At the industry level, publisher awareness is still catching up. In a webinar I hosted in May, 47% of the audience said they’d heard of MCP but couldn’t clearly explain how it works, and 27% hadn’t encountered the term before. Only 9% were actively evaluating or implementing integrations. 

That gap between developer adoption and publisher readiness is both a window and a risk. The window is the opportunity to shape how AI agents access scholarly content before the defaults are set by actors with less investment in getting it right. The risk is that if publishers aren’t building MCP endpoints that reflect the values of scholarly publishing (attribution, version control, retraction handling, entitlement management), those defaults will be set without them. 

Jeremy Little, Silverchair’s VP of AI, identified something unique about our current moment: unlike with many technology advancements, of the three barriers to widespread MCP adoption (technology readiness, provider readiness, and user readiness), user readiness may be the least significant obstacle. Researchers are already using AI tools for the tasks MCP was built to improve. The gaps are on the infrastructure and provider sides. Standardization is incomplete, and publishers are at different stages of building out MCP endpoints and working through how to make structured scholarly metadata legible to AI systems that weren’t designed with the scholarly record in mind. 

Making that value legible also requires updated measurement frameworks. COUNTER Metrics has spent much of the past year developing guidance on how to count AI-generated interactions in a COUNTER-compliant way. The frameworks will give libraries, for the first time, a coherent way to see AI-assisted usage alongside traditional human usage. If AI-mediated access to content doesn’t register in usage statistics, the value of publisher content to libraries is systematically undercounted, which is ultimately a problem for publishers too. 

But there are reasons to be optimistic, even confident. The infrastructure that scholarly publishing has built over decades — DOIs, JATS tagging, structured metadata, the commitment to versioning and retractions — is an asset in the AI environment, not a relic. These are exactly the signals that make scholarly content more trustworthy and more useful to AI systems. The problem is that those signals don’t automatically travel with content as it moves through AI-mediated workflows. Getting them to do so requires the kind of collective standards work the industry has historically found difficult but not impossible. Andrew Smeall drew the parallel to JATS directly: just as the industry developed a shared tagging standard that made journal content reusable across platforms and databases, it now needs something analogous for AI communication — a way for AI tools to understand provenance, version status, and institutional context. That standard doesn’t exist in agreed-upon form yet, but the ingredients are largely there. 

Cheryl Firestone, Senior Manager of Digital Publishing at AAP framed the strategy that many in scholarly publishing are beginning to converge on: “I think a lot of the work being done today is really laying the groundwork for future tools. It’s making the initial investment in the infrastructure that we know we can expand.” 

The user has changed. The research workflow has changed. The infrastructure for measuring, attributing, and delivering scholarly content is in the process of changing. What hasn’t changed is the underlying purpose: getting trustworthy knowledge to the people — and increasingly the systems — that need it. 

Author’s note: Many thanks to the contributors, whose insights were pulled from their participation in Silverchair’s 2026 Platform Strategies Webinar Series: Andrew Smeall (Sage), Jane Jiang (Union College of Union County), Jeremy Little (Silverchair), Cheryl Firestone (American Academy of Pediatrics), Tanya Laplante (Oxford University Press), Tasha Mellins-Cohen (COUNTER Metrics), Paul Gee (JAMA Network), Lou Peck (The International Bunch), and Robb Burgess (Silverchair). Explore the recordings and recaps.