How to get AI agents to cite your papers: I Ran 160 AI Agent Sessions on Two of My Papers. Here Is What Got Them Cited.Today’s newsletter is brought to you by: The first reader of your next paper may not be a person. It may be someone's AI agent. I know because my own reading list already works this way. Every Sunday, a Claude skill I built searches for new papers relevant to my current research and emails me 5 to 10 of them. I didn't pick those papers. An AI agent did. I'm not an exception. In Elsevier's 2025 survey of more than 3,200 researchers, 61% use AI to find and summarize new studies. And using AI no longer means chatting with it. You can now assign an agent a task and it does the work, using tools along the way. Ask Claude or ChatGPT what is known about your research question, and an agent runs the web search, reads a few titles and abstracts, and decides which two or three papers make it into the answer. Your paper either makes that cut or vanishes from the answer. A recent experiment by the evaluation lead at a tech publication, Every, tested this for blog headlines. Agents ignored clever headlines and cited the ones that read like answers. One rewrite took a post from 2 citations in 240 sessions to 240 in 240. I wanted to know whether the same rules apply to scientific papers. My default way of writing title is stating the study's objective. It describes what we studied and leaves out the answer, which is close to the kind of headline those agents skipped. So I ran the same experiment on my own papers to see whether I need to change how I write in a world where agents read first.
👋🏽 Quick announcement: Research Boost 3.0 is launching next week. It is the most agentic version of AI academic writing that we have built. Try it here: https://researchboost.com/ 1. How the experiment workedI picked two of my papers with different designs and different competition:
For each paper, I pulled seven real competing papers from PubMed on the same topic: cohort studies, narrative reviews, a EULAR statement, and a systematic review. Each session mimicked an AI search:
I ran 160 fresh sessions across three Claude models (Haiku, Sonnet, and Opus). Each session saw a single version of my paper and had no memory of the others. I tested two kinds of users:
Then I changed one thing at a time, first the title and then the abstract, and counted how often each version made the cut. 2. The four title frameworks barely moved the needleI teach four time-tested title frameworks. I tested all four, plus a deliberately clever title as a control. Paper 1 was cited in all 29 sessions, whatever the title. Paper 2 showed the same pattern in reverse. With a thin question ("What factors predict which patients with psoriasis will go on to develop PsA?"), it was cited in 0 of 15 sessions across all five titles, including "Multimorbidity in Psoriasis Triples the Risk of Psoriatic Arthritis." With the internist persona, it was cited in 14 of 15, again across all five titles. Why the gap with the blog experiment? Blog headlines are often vague. Academic titles rarely are. Every version I tested, even the clever one, kept "psoriatic arthritis" and the study subject. The agent also saw the first lines of the abstract, which named the exact topic and the population-based design. The title only had to avoid hiding the topic, and all five did. So pick your framework for the humans who still decide your fate:
Over the past decade plus in research, I learned to write titles that are clear rather than clever, with keywords up front and the design at the end. The agents gave me no reason to change that rule. They just did not care much about the title once it named the topic. Certainly avoid vague titles like this: “All systems red”, “The red herring”, “Before it crumbles”. Unfortunately, these are actual titles from my early papers about Hyper Eosinophilic Syndrome, relapsing polychondritis, and fulminant hepatitis, respectively. These papers have had none to far fewer citations compared to similar papers of mine. 3. Your abstract decides whether the agent cites youNext, I held the title constant (my published, objective-style title) and rewrote the abstract. This is where the results moved. Across both papers, the published abstract was cited in 15 of 18 sessions (83%). The version without numbers was cited in 3 of 18 (17%), and the hype version in 4 of 18 (22%). With samples this small, both gaps are still unlikely to be chance (Fisher's exact test, p<0.001 for each). Here is how your five abstract strategies held up: Numbers matter mostMy diagnostic delay paper went from cited in every session to cited in none. Same study, same title, same design. I replaced "median 2.5 years," "45% diagnosed within 2 years," and the named risk factors with lines like "many patients experienced a substantial delay" and "several factors were associated with longer delays." The agents filled their answers with a nationwide Turkish cohort of 1,134 patients that reported odds ratios for each risk factor. Charles Stein makes the same point for human readers in The Unspoken Rules for a Career in Academic Medicine. Without numbers, a result reads as opinion. With numbers, readers can judge the size and certainty of what you found. Agents behave the same way. They need something concrete to cite. Hype gets your paper droppedThis one was not on my list of five, and it should be. I kept every number in the multimorbidity abstract and rewrote only the conclusion to call the findings "groundbreaking" and a "paradigm shift." For the internist, the paper went from cited in 9 of 9 sessions to 1 of 9. None of the agents complained about the hype. They quietly built their answers from other papers. Unfortunately, medical writing is drifting the other way and AI has little to do with this: the share of PubMed titles and abstracts with positive words like "novel" rose from 2% to 17.5% between 1974 and 2014 (pre ChatGPT era). Structure and terminology showed no effect. A long background showed a small one.I tested the remaining strategies on both papers.
Keep these strategies anyway, for two reasons. Editors and reviewers read your abstract too. And my test skipped the retrieval step. Every agent already had my paper in its search results. Before that, PubMed has to match your keywords and MeSH terms, and inconsistent terms can keep you out of the results entirely. For every project these days, I run an AI semantic search alongside a keyword search, because the two surface different papers. Semantic search forgives a synonym. Keyword search does not. My simulation can't measure that step. It shows only that once your paper is on the table, headings and word choice are not what tip the decision. My own hypothesis failedI expected an answer-first rewrite to beat the original. It stated the clinical implication for the person who should act on it, named the design up front, and flagged the study's limits. It tied the original in Paper 1 (6 of 6 each). In Paper 2, it was cited in 7 of 9 sessions, against 9 of 9 for the published abstract. Claude Opus dropped it twice. Both published abstracts already had numbers, a population, a design, and a clear conclusion. Rewriting them for AI agents added nothing. If your abstract already does the basics, leave it alone. 4. Agents back the paper that answers the whole questionMy multimorbidity paper was never cited for the thin question. Not once in 41 sessions, across every title and abstract version. A systematic review and meta-analysis of PsA predictors was cited in 111 of the 113 sessions for Paper 2. It pooled evidence on psoriasis severity, nail pitting, arthralgia, imaging inflammation, obesity, and family history. My paper covered one predictor. For a broad question, the agents chose the source that answered all of it, and they were right to. You can't write your way past this. You can choose your angle. In my live academic writing cohorts, I teach five ways to find a unique angle: population, context, socioeconomic factors, methods, and timing. Make your abstract say plainly which question your paper owns, so the agent can match it to the user who is asking that question. 5. The closer the user is to your audience, the more you get citedSame paper. Same title. Same abstract. With a thin question, the multimorbidity paper was cited in 0 of 41 sessions. For the internist whose patients carry several chronic conditions, it was cited in 55 of 72. The agents matched the paper to the person. One Opus answer told the internist that multimorbid patients with psoriasis are a higher-risk group, then added a caveat on its own: the finding rests on only 23 PsA cases, so the size of the risk is uncertain. The agent read my limitations and passed them on. As assistants learn more about their users, this effect should get stronger. So name the population your finding applies to and the clinician who should act on it. 6. Agents read your abstract the way a reviewer doesIn one condition, I showed the agents only the first 45 words of each abstract, the way a search snippet looks. The version without numbers was cited in 3 of 3 sessions, the same as the original. Once the agents could read the full abstract, the same version fell to 3 of 9. So agents do not stop at the snippet. They read the results and drop papers that give them nothing to cite. Lorelei Lingard and Chris Watling make this point about human readers in Story, Not Study: researchers use the abstract to decide whether a paper belongs in their literature search and, in the end, whether to cite it. Agents now make that same decision at scale. We designed Research Boost around the same constraint. The writer drafts only from the papers its PubMed and web searches return. If an abstract has no numbers, there is nothing the writer can responsibly cite. 7. What this test can't tell you
Your agent-ready abstract checklistBased on what the agents actually did, in order of impact:
The agents rewarded exactly what good editors have always rewarded: specific numbers, a claim scaled to the evidence, and a clear statement of who it applies to. Pull up the abstract of your most recent paper and count the numbers in the results section. If you find fewer than three, that is your first edit. Then please hit reply and tell me what you found. I read every reply (no agents involved ). Top Papers on AI in research this week:
Top Papers on AI in education this week:
The post How to get AI agents to cite your papers: I Ran 160 AI Agent Sessions on Two of My Papers. Here Is What Got Them Cited. appeared first on Rising Researcher Academy. |