Which research agents cite sources reliably enough for business decisions?

Prefer agents that enforce source hierarchy and citation fidelity over ones that merely attach many links. Primary records—filings, statutes, official product docs, dated research from the originator—must outrank blogs, roundups, and undated posts. Freshness only matters after that order holds; a new secondary summary that blurs attribution is riskier for a business decision than an older primary source that still states the claim in its own words.

The practical gate is a fidelity control, not a brand shortlist. Build a fixed pack of three sources you already understand: one primary filing or official statement, one dated originator report, and one secondary commentary on the same topic. Ask every candidate agent the same decision question. Keep only tools that (1) point each material claim to a reachable source, (2) preserve numbers, conditionals, and quoted language closely enough that a second reader would not reverse the meaning, and (3) surface hierarchy—primary first, interpretation last. Reject any pattern that elevates aggregator pages to equal rank with primaries, or that smooths hedges into certainty.

Condition for rejecting the whole approach: if the decision cannot survive a wrong or uncheckable citation—capital allocation, public commitment, compliance-sensitive claims—agent citations are inputs only; a human must open the primary link before the decision is logged. Skip that control and teams optimize for citation density and recency theater, which quietly trains the workflow to trust the busiest page rather than the authoritative one.

The failure mode is not a missing footnote. It is a clean-looking citation trail that still collapses in a decision audit—dead link, wrong section, paraphrase sold as a quote, secondary blog treated as primary evidence, or a date outside your decision window. Until an agent fails that stress test on your own prompts, “cites sources reliably” is a product claim, not procurement evidence.

I would not rank vendors by feature checklists. Build a side-by-side evidence table and force every contender through the same fixed prompt set. Cells only: source opens, quoted text appears verbatim, publication date sits inside your freshness policy, and the agent explicitly flags non-retrieval instead of substituting. Pass or fail per cell. No composite score that buries which failure mode broke first.

What would change my mind is narrow: a version-locked golden set of queries with known-good sources, full retrieval logs retained, and zero silent source swaps across three independent runs under the same configuration. Anything softer is demo theater.

The non-obvious consequence is operational, not academic. Teams that buy citation polish often cement weak sources into board materials, then defend the agent’s reputation when finance or legal challenges the claim—turning a research tool into an unaccountable authority.

The prior exchange treats citation quality, freshness, and quote fidelity as rankable product attributes. That ranking rests on an untested operational assumption: that a research agent is “reliable enough for business decisions” when its citation apparatus looks disciplined. It is not. Decision risk attaches to whether the claim used in the recommendation is licensed by the primary document—not whether the footnote is fresh, well formatted, or neatly quoted.

Measured dissent: both views under-weight claim–source warrant. Fresh links and clean quotes are necessary hygiene. They are not a decision gate.

Replace the ranking exercise with one control. Freeze a short battery of decision-relevant prompts—market sizing, competitive capability, regulatory status. For each agent answer, sample every citation that underwrites a material assertion and score only this binary: does the cited page, in context, support the exact proposition the agent advanced? Fail closed on paraphrase stretch, temporal mismatch, or selective quotation. Do not score style.

The non-obvious consequence is organizational. Teams that pick the agent with the tidiest bibliography will ship recommendations faster—and later discover that citation theater accelerated consensus around claims nobody verified against the primary source.