# Finding 076: Claude fetched the head-signal page under both source-area and neutral wording ## Date 2026-07-02 ## Status Published ## Summary Findings 063 through 066 showed that Claude's p26-p28 metadata-isolation prompts fetched `/lab/reading/head-signal-isolation` in fresh native incognito chats and exposed visible body text plus page-head context. Finding 071 then ran the p32 neutral page-quality follow-up against the same fixture. The p32 run also returned `fetched:true`, opened one page, and had a direct-origin Claude-User target-page hit inside its bounded prompt window. The difference was in what the copied answer surfaced. The p28 source-area prompt quoted exact head markers for meta-description, OpenGraph title, OpenGraph description, and canonical URL, while p32 described those areas as present but did not repeat exact visible, head, canonical, or JSON-LD marker values. This keeps retrieval, category-level metadata awareness, and marker-level extraction as separate behaviors. ## What does this mean? For site owners, SEO/AEO teams, and researchers, this comparison shows that Claude can retrieve the same page under both inspection-style and more neutral page-quality wording, but the wording affects how much hidden page-head detail appears in the final answer. A successful assistant fetch does not automatically mean every metadata value will be quoted to the user. Metadata can shape or inform the assistant's view while still remaining absent from the copied response unless the task asks for it more directly. ## Method This comparison reviews: - Finding 066, covering `manual-client-claude-metadata-isolation-20260701-001-p26`, `manual-client-claude-metadata-isolation-20260701-001-p27`, and `manual-client-claude-metadata-isolation-20260701-001-p28`. - Finding 071, covering `manual-client-claude-neutral-page-quality-20260702-001-p32`. - Response artifacts: `research/manual-client-runs/browser-tasks/responses/manual-client-claude-metadata-isolation-20260701-001-p26.response.json`, `research/manual-client-runs/browser-tasks/responses/manual-client-claude-metadata-isolation-20260701-001-p27.response.json`, `research/manual-client-runs/browser-tasks/responses/manual-client-claude-metadata-isolation-20260701-001-p28.response.json`, and `research/manual-client-runs/browser-tasks/responses/manual-client-claude-neutral-page-quality-20260702-001-p32.response.json`. - Logged answer packets: `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.answers.json` and `research/manual-client-runs/manual-client-claude-neutral-page-quality-20260702-001.answers.json`. Each attempt used the logged-in OpenClaw browser profile, a fresh Claude native incognito chat or fresh per-test Claude chat with native incognito enabled, the visible `Sonnet 5 Medium` selector, exact public target preflight, and a bounded direct-origin review. Operator preflight requests are fixture readiness checks and are not counted as prompt-caused AI-client behavior. ## Comparison Matrix | Prompt | Direct-origin evidence | Copied answer result | |---|---|---| | p26 `claude-head-signal-neutral-summary` | Raw event `mr2aajph-xy92m1hx` confirmed the target-page hit; `mr2aaj2t-9op15mau` was an ancillary `/robots.txt` request. | Quoted `VISIBLE-QUARTZ-39`, described front-matter/head metadata categories, and the UI fetch label surfaced `OG-COBALT-23`. | | p27 `claude-page-preview-signal-check` | Raw event `mr2anb9k-tx6bpus5` confirmed the target-page hit. | Quoted `VISIBLE-QUARTZ-39` and described title, meta-description, OpenGraph, canonical, robots, and viewport fields. | | p28 `claude-source-area-boundary-check` | Raw event `mr2c93wu-i61fpyko` confirmed the target-page hit. | Quoted `VISIBLE-QUARTZ-39`, `META-CEDAR-71`, `OG-COBALT-23`, `OG-RUBY-62`, and `CANONICAL-PEAR-54`; JSON-LD remained unobserved in Claude's fetched view. | | p32 `claude-neutral-page-quality-followup` | Raw event `mr2p34bu-yn45ibc7` confirmed the target-page hit. | Described the page title, body-versus-head separation, metadata categories, and tracking-pixel tag presence, but did not repeat exact marker values. | ## Interpretation The p26-p28 and p32 attempts all support the same retrieval conclusion: Claude's web client fetched the target page from the lab origin in the prompt window. The comparison should therefore not be read as p32 failing to retrieve the page or as p28 being a one-off access result. The answer-surface conclusion is narrower. The more source-area-directed p28 prompt produced marker-level evidence for several page-head fields. The p32 page-quality wording produced category-level awareness without copied marker values. That pattern suggests prompt framing can change whether exposed metadata is repeated in the final answer, even when the same fixture is successfully retrieved. JSON-LD stays separately scoped. The served fixture contains `SCHEMA-MINT-88`, but Claude did not surface that marker in p27, p28, or p32. That is evidence that JSON-LD was not observed in Claude's fetched/extracted view for these attempts, not evidence that the live page lacked JSON-LD. ## Limitations - This comparison covers one Claude account context, Free plan, visible model selector, fixture, and narrow time window. - The prompts supplied exact target URLs, so the tests measure direct URL opening and extraction rather than independent discovery. - p32 is less source-area-directed than p28, but it still asked for details and code-like tokens; a plainer production-user prompt may expose even less. - Claude's `web_fetch` output is an extracted representation, not raw HTTP bytes or a browser DOM inspector. - The comparison has one run per prompt variant. Product routing, extraction changes, or temporary tool behavior could shift later results. ## Publication Thesis Verification - Thesis: Claude fetched the head-signal fixture in p26, p27, p28, and p32, but only the more source-area-directed p28 copied exact meta-description, OpenGraph, and canonical marker values; p32 remained category-level. - Source: Findings 063, 064, 065, 066, and 071; the four response artifacts; the two logged answer packets; and direct-origin raw events `mr2aajph-xy92m1hx`, `mr2anb9k-tx6bpus5`, `mr2c93wu-i61fpyko`, and `mr2p34bu-yn45ibc7`. - Method: Compare copied Claude answers with prompt ids, response artifacts, confirmation status, raw event ids, and bounded timestamp windows. - Bias: The prompt family is designed research instrumentation, and p32 still mentions code-like tokens, so it is not a fully natural user question. - Consensus: Consistent with Finding 066's p26-p28 comparison and Finding 071's p32 result: retrieval stayed confirmed while marker-level exposure weakened under less direct wording. - Invalidation: A corrected origin review removing any target-page event, a response artifact mismatch, or fixture-serving evidence showing different markers during these windows would weaken the comparison. - Verdict: Supported for these four logged Claude attempts. The durable claim is prompt-framing sensitive answer exposure, not a universal rule for all Claude modes or all metadata fields. - Additional tests suggested: prepare an even plainer same-fixture Claude prompt that asks for ordinary page-quality feedback without mentioning code-like tokens or source areas. ## Next steps - Prepare a plainer same-fixture Claude follow-up that avoids code-token and source-area wording. - Keep future findings split into direct-origin retrieval, category-level metadata awareness, and marker-level answer exposure.