What Claude consults
Claude's commercial-discovery behaviour depends on the surface. In Anthropic's first-party surfaces with retrieval enabled, Claude consults a broad index biased toward editorial and authoritative sources. In enterprise deployments, Claude is increasingly grounded on customer-supplied corpora plus a curated retrieval layer.
For the buyer asking a commercial question through a consumer-grade Claude surface, the model behaves consistently. Three observations recur.
The citation patterns we observe
- Substantive depth is rewarded. Pages that ground their claims with detail are cited at meaningfully higher rates than pages that summarise.
- Named authorship is weighted heavily. A credentialed author is, in Claude's behaviour, near-equivalent to a Tier I citation.
- Editorial caution shows up. Claude is more conservative than its peers in commercial recommendation. It frequently declines to recommend a specific provider where the underlying authority signal is thin.
Claude is the system on which the work of becoming an authoritative-looking subject — institutional voice, named credentials, evidenced argument — produces the most consistent improvement. The system reads as if it were assessing whether to vouch for a source, and refuses where the source has not earned the vouching.
The three moves that work
One — author standing
Each substantive page is authored by a named, credentialed practitioner whose authority profile is verifiable. The author's qualifications, professional memberships, and publication record are visible and machine-readable.
Two — reasoning on display
Pages show the basis for their claims — sources cited, methods described, examples given. The aesthetic is closer to a long-read magazine essay than to a marketing page.
Three — institutional voice consistency
The subject's content reads as if produced by an institution rather than by an outsourced content shop. Voice consistency, editorial standards, and absence of marketing flourishes are signals Claude appears to weight.
What does not work
- Tone drift between marketing pages and the substantive content the model is meant to find.
- Claims unsupported by visible reasoning, even where claims are accurate.
- Volume content under unnamed or generic authorship.
- “Claude-friendly” rephrasing exercises that produce no underlying change in evidence quality.
How the institution measures it
The Index includes a Claude-specific instrumentation in subjects engaged on a Claude-led brief. The institution observes Claude's commercial answers across a stable prompt set quarterly and reports both citation appearance and qualitative recommendation strength.
A closing observation
Claude treats the act of recommending a business as a small editorial decision. The work of being recommendable is therefore the work of becoming the kind of business an editor would put on the record.
Subjects who succeed on Claude tend to be the ones whose internal standards make them the kind of business that would survive a serious profile in the trade press. This is not a coincidence.