Prompting Agentic AI for Citable Outputs: 6 Rules That Prevent Fabrication
Prompt engineering for research citations usually means adding "cite your sources" to the end of a query. A 2026 evaluation of citation accuracy across 14 large language models found that instruction is not doing much: link validity and topical relevance both cleared 80%, but factual accuracy — whether the cited source actually supports the claim attached to it — landed at only 39 to 77%, and dropped by roughly 42% on average as the agent's tool calls scaled from 2 to 150 (arXiv:2605.06635). More searching did not produce more accurate citing. That gap is what six specific prompting rules are built to counter.
Why More Retrieval Makes Fabrication Worse
Deep research agents plan sub-questions, run searches against each one, and synthesize a report — the loop that makes them "agentic" rather than a single-shot reply, covered in the comparison of three current research agents. Each additional search pass adds more retrieved text for the model to reconcile into one narrative, and reconciliation is where a model's own training-time assumptions can override what a specific source actually said. An agent with no scope boundary tends to fill the gap with confident synthesis instead of reporting that a claim was not actually found.
Rule 1: Scope-Lock the Query
State the boundary before the agent starts — a date range, a study type, three named databases, or one specific claim to check — instead of an open-ended research question. A bounded query gives the agent fewer places to substitute plausible synthesis for a source it never actually retrieved.
Rule 2: Ask for the Source Before the Claim
Structure the prompt to request the citation first and the claim second: find a source that reports X, then state what it says, rather than tell me about X and cite it. OpenAI's own deep research documentation recommends the equivalent check in reverse — it tells users to click through and verify data points, statistics, and quotes against the original source rather than trusting an inline citation on its face (OpenAI Help Center).
Rule 3: Treat Citation Density as a Red Flag, Not a Quality Signal
A report with fifty or more citations is not automatically more reliable than one with ten. Each additional citation is one more claim requiring the same individual check as a single one would, and a tool built to maximize discovery volume is optimizing for a search feature, not a verification outcome.
Rule 4: Separate "Link Resolves" From "Claim Is True"
Build a two-step check into the workflow after the report comes back: first confirm every link is live and points to a correctly attributed source, then separately confirm the source actually says what the report claims. The 2026 benchmark's own numbers show why collapsing these into one step is risky — link validity cleared 94% while factual accuracy stalled at 39 to 77% on the same set of reports (arXiv:2605.06635).
Rule 5: Cap the Retrieval Budget
If a tool exposes a setting for search depth or tool-call count, use the lower end for citation-sensitive work. The same benchmark found fact-check accuracy fell by roughly 42% on average as tool calls scaled from 2 to 150 across two frontier models — a deeper search pass bought more text to reconcile, not more accurate reconciliation.
Rule 6: Verify After, Not Instead Of
None of the first five rules replace a standalone check once the report is in hand. pip install citecheck runs an open-source, MIT-licensed check of a manuscript's reference list against CrossRef, PubMed, Semantic Scholar, and OpenAlex — confirming a citation is a real, correctly attributed paper, not that the paper supports the sentence sitting next to it. Running citecheck manuscript.docx --format json --output report.json on whatever an agent hands back is the same discipline citation hallucination in AI-drafted manuscripts already requires, regardless of which tool produced the draft, before any of it reaches a submission.