# Finding 066: Claude's p26-p28 metadata runs exposed head signals unevenly ## Date 2026-07-01 ## Status Published ## Summary The Claude metadata-isolation sequence ran three fresh native incognito chats against the same `head-signal-isolation` fixture: `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`. All three attempts returned `fetched:true` with one page opened and all three had confirmed Claude-User target-page hits inside their bounded prompt windows. Across the sequence, Claude consistently surfaced the visible body marker `VISIBLE-QUARTZ-39` and described page-head metadata as present in its `web_fetch` extracted view. The strongest marker-level result came from p28: Claude quoted the meta-description marker `META-CEDAR-71`, OpenGraph title marker `OG-COBALT-23`, OpenGraph description marker `OG-RUBY-62`, and canonical marker `CANONICAL-PEAR-54`. In p26 and p27, Claude named metadata categories and the UI fetch label surfaced `OG-COBALT-23`, but the copied answers did not quote every hidden marker value. In p27 and p28, Claude said JSON-LD was not present in its fetched content even though preflight confirmed the served fixture contained JSON-LD marker `SCHEMA-MINT-88`. ## What does this mean? For site owners, SEO/AEO teams, and researchers, this comparison shows that Claude's page-opening view can include page-head and preview signals that are not printed in normal body copy, but the exposure is not uniform across all source areas. Description, OpenGraph, and canonical fields were visible enough to appear in the strongest run, while JSON-LD did not surface in Claude's extracted view. Metadata should be treated as potentially readable by assistant retrieval tools, but each field still needs direct testing instead of assuming all head-adjacent signals behave the same way. ## Method - Source findings: Finding 063, Finding 064, and Finding 065. - Prompt packet: `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.prompts.json`. - 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`, and `research/manual-client-runs/browser-tasks/responses/manual-client-claude-metadata-isolation-20260701-001-p28.response.json`. - Logged answer packet: `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.answers.json`. Each attempt used the logged-in OpenClaw browser profile, a fresh Claude native incognito chat, the visible `Sonnet 5 Medium` selector, an exact public target URL preflight, and a bounded prompt-window origin review. ## Evidence | Prompt | Direct-origin result | Answer result | |---|---|---| | p26 `claude-head-signal-neutral-summary` | Raw events `mr2aaj2t-9op15mau` (`/robots.txt`) and `mr2aajph-xy92m1hx` (`/lab/reading/head-signal-isolation`) confirmed a prompt-window hit. | Claude 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 a prompt-window target-page hit. | Claude quoted `VISIBLE-QUARTZ-39`, described title, meta-description, OpenGraph, canonical, robots, and viewport fields, and the UI fetch label surfaced `OG-COBALT-23`. | | p28 `claude-source-area-boundary-check` | Raw event `mr2c93wu-i61fpyko` confirmed a prompt-window target-page hit. | Claude quoted `VISIBLE-QUARTZ-39`, `META-CEDAR-71`, `OG-COBALT-23`, `OG-RUBY-62`, and `CANONICAL-PEAR-54`, and said JSON-LD was not present in its fetched content. | ## Interpretation The comparison separates three layers of evidence. First, all three runs are retrieval-confirmed because server-side events show Claude-User target-page activity inside the prompt windows. Second, p26 and p27 support category-level head-signal exposure because Claude described head/frontmatter fields and the UI fetch label revealed the OpenGraph title marker. Third, p28 supports a stronger marker-level claim for meta description, OpenGraph title, OpenGraph description, and canonical fields because those exact values appeared in the copied answer. JSON-LD remains a different result. The fixture preflight contained `SCHEMA-MINT-88`, but Claude did not surface that marker and said no JSON-LD structured data block was present in the fetched content. That makes JSON-LD unobserved in Claude's extracted view for these runs, not absent from the served page. ## Limitations - This comparison covers one Claude account, tier, visible model selector, fixture, and day. - The prompts supplied exact target URLs, so the sequence measures direct URL opening and extraction, not independent discovery. - The prompts intentionally ask about page areas and code-like markers; a neutral production-user prompt may expose fewer signals. - Claude's `web_fetch` output is an extracted text representation, not a raw HTTP source dump or browser DOM inspector. - The sequence has one run per prompt variant, so product routing or temporary extraction changes could affect the results. ## Publication Thesis Verification - Thesis: In the Claude p26-p28 metadata-isolation sequence, Claude fetched the target page in all three runs, exposed visible body text and page-head metadata context, exposed several exact head markers in p28, and did not surface JSON-LD from its fetched view. - Source: Findings 063, 064, and 065; the p26-p28 response artifacts; the logged answer packet; and raw event ids `mr2aajph-xy92m1hx`, `mr2anb9k-tx6bpus5`, and `mr2c93wu-i61fpyko`. - Method: Compare copied Claude answers, visible browser transcript notes, and direct-origin raw events by prompt code and bounded timestamp window. - Bias: The prompts were designed to inspect page areas and marker strings, so they are better evidence for extraction boundaries than for ordinary user wording. - Consensus: The comparison is consistent with earlier Claude structured-data findings where meta-description text surfaced but JSON-LD did not. - Invalidation: A response artifact mismatch, corrected origin review removing any of the target-page hits, or fixture-serving evidence showing different head markers during the runs would weaken the comparison. - Verdict: Supported for the three logged Claude runs. The strongest claim is field-specific for p28; p26 and p27 remain category-level head-metadata exposure evidence. - Additional tests suggested: create a neutral Claude follow-up that asks for ordinary site-quality feedback without naming page-head areas or marker-like tokens, then compare whether the same fields surface. ## Next steps - Prepare a neutral Claude metadata follow-up that avoids page-area and marker wording. - Add same-fixture repeat runs later to separate stable extraction behavior from temporary product routing.