What Claude Reads and Cites: A 95,000-Record Study

Type: Primary read of a VENDOR study, supplied by Adam. Source: searchable.com/claude-visibility-report.

1. Provenance, and the cap

Searchable is an AEO vendor: it sells the discipline this report promotes. Conflict recorded. Method as disclosed: ~95,000 search-result records across 3,000 unique prompts in Claude, July 2026, single engine, single month. Search-provider behaviour is inferred from index overlap — the study itself states Claude has not disclosed a provider. No confidence intervals published; no rerun counts per prompt disclosed, so we cannot verify any figure clears the rerun-stability floor (Jaccard 0.50–0.61, digest 03 R23).

Tier B, hard cap. Real data, method partially disclosed, vendor-run, no error bars. Nothing below may source a MUST on its own; where a finding corroborates held tier-A work it strengthens that work, and is marked so.

2. Findings

2.1 The retrieval/memory prompt split — the most useful frame in the study

Claude answers some buyer questions by searching (content can win these) and others from parametric memory (no page you publish reaches them in real time). Trigger rates by prompt shape:

Their prescription — classify twenty unbranded buyer questions, split retrieval-shaped from memory-shaped, and let the ratio set the content budget — is cheap, testable, and maps directly onto the parametric-vs-retrieval pathway question already in our open queue (Mallen et al. 2023, digest 03 §6). This is the study’s genuinely new contribution to our library.

2.2 Brave overlap — corroborates a standing REPORTED claim

Pages Claude read overlapped the Brave top ten at 63% against 34% for Google, same queries, same day. This is index-overlap inference, not confirmation — but it is an independent sighting of the Claude→Brave substrate dependency (TechCrunch 2025-03, REPORTED in the canonical substrate module). The module note now records this corroboration. Still not DOCUMENTED: only Anthropic can do that.

2.3 The top-ten gate — rank is binary here, not a gradient

Once a page is read, citation rates are nearly flat across rank: 39.4% (ranks 1–3), 37.2% (4–6), 36.8% (7–10). Position 11+ is rarely read at all. Their framing is right: getting inside the pool is the whole game; exact rank inside it barely matters. Consequence they draw, which we endorse as a test: move ranks 11–30 into the top ten before spending anything on moving rank 7 to rank 3.

2.4 Source-type conversion — press is Claude’s door, and no TechRadar gate exists

Share of READ sources that made the final answer, Claude vs ChatGPT: national press 52.0% vs 19.2% (2.7×), review sites 49.1/44.8, smaller comparison blogs 42.6/28.8, directories 30.3/29.3, Reddit/forums 8.3/12.7. And concentration: TechRadar alone was 22% of ChatGPT’s cited sources in their sample; on Claude no site reached 9% — the opportunity is distributed, which favours a many-small-placements plan over one flagship placement.

2.5 Own-site usage — do NOT conflate with the 2.9% figure

Claude used a brand’s own site in 31.6% of cases vs ChatGPT’s 60.9%. This is a different metric from arXiv 2606.20065’s “own-domain content = 2.9% of content citations” — that was share of ALL citations; this is how often the subject brand’s own site participates when the brand is in play. Both point the same direction (earned over owned, more so on Claude), but the numbers must never be quoted interchangeably.

2.6 Unprompted recall — memory supplies the shortlist

On unbranded questions Claude supplies a brand shortlist from memory, and 86% of memory-supplied brands appeared in the final answer. Consistent with digest 03 §3.6 (follow-up turns drill into turn-1 brands rather than adding new ones). Their recall audit — ten unbranded questions × five fresh chats, count mentions of fifty, track quarterly — is a sound instrument and cheap to reproduce with our own engine set.

2.7 Local runs through listings, not the web

97.7% of local prompts retrieved via a business-listings service; 0.1 conventional web pages read per local query. If replicated, local visibility is a listings-hygiene problem, not a content problem. Caveat from their own testing: unresolved user location can default to London — geo-stratification (R13) applies to local probes doubly.

2.8 Platform change — watermarking

Anthropic: models launched on/after 2026-08-02 mark generated text imperceptibly and attach signed C2PA provenance to supported SVG/PNG/JPG. A detected mark means Claude may have PROCESSED the material, not that it authored it. Provenance context only; treat any “watermark hurts ranking” claim that appears downstream as folklore until measured.

3. Revalidation of held positions

4. What we adopt as TESTS (not rules — tier B cap)

1. Prompt-split audit as a scoping instrument: classify a client’s twenty buyer questions into retrieval vs memory shapes before allocating content budget. Reproduce trigger rates on our own probes before quoting theirs.
2. Rank 11–30 sweep in Brave specifically for Claude-targeted clients.
3. Quarterly recall audit (2.6) added to the measurement stage design.
4. Replicate 2.7 (local/listings) before repositioning any local client’s plan around it.

5. Open queue

1. Reproduce the prompt-split trigger rates in-house (cheap; our probe kit already exists).
2. Trace whether the 63/34 Brave overlap replicates on our query set.
3. Watch for the watermark-as-ranking-signal folklore to appear in quarantine; pre-registered as folklore here.