# Finding 078: Claude kept exact head markers out of plainer page-quality answers ## Date 2026-07-02 ## Status Published ## Summary The Claude head-signal series now includes five fresh controlled-browser runs against the same `/lab/reading/head-signal-isolation` fixture: `manual-client-claude-metadata-isolation-20260701-001-p26`, `manual-client-claude-metadata-isolation-20260701-001-p27`, `manual-client-claude-metadata-isolation-20260701-001-p28`, `manual-client-claude-neutral-page-quality-20260702-001-p32`, and `manual-client-claude-plain-page-quality-20260702-001-p36`. All five returned `fetched:true`, opened one page, and had direct-origin `Claude-User/1.0` target-page hits inside their bounded prompt windows. The answer surface changed with prompt wording. The source-area p28 prompt quoted exact meta-description, OpenGraph, and canonical marker values. The less source-directed p32 prompt recognized the page-head/body separation but did not repeat exact marker values. The plainer p36 site-owner review prompt removed source-area, metadata, structured-data, code-like-token, marker-value, and hidden-field wording; Claude still identified the page as a diagnostic head-signal fixture, but again did not copy exact visible, meta-description, OpenGraph, canonical, or JSON-LD marker values. ## What does this mean? For site owners, SEO/AEO teams, and researchers, these Claude runs show that a successful assistant page fetch is not the same thing as exact metadata leakage in the final answer. Claude retrieved the same page under direct inspection, neutral review, and plainer site-owner wording, but exact head-marker values only appeared when the prompt asked much closer to source-area inspection. Ordinary-sounding questions can still let metadata shape the assistant's understanding while leaving specific machine-readable values out of the copied response. ## Method This comparison reviews: - Finding 066, covering p26-p28 metadata-isolation runs. - Finding 071, covering p32 neutral page-quality follow-up. - Finding 076, covering the p26-p28 versus p32 prompt-framing comparison. - Finding 077, covering p36 plain page-quality follow-up. - 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`, `research/manual-client-runs/browser-tasks/responses/manual-client-claude-neutral-page-quality-20260702-001-p32.response.json`, and `research/manual-client-runs/browser-tasks/responses/manual-client-claude-plain-page-quality-20260702-001-p36.response.json`. - Logged answer packets: `research/manual-client-runs/manual-client-claude-metadata-isolation-20260701-001.answers.json`, `research/manual-client-runs/manual-client-claude-neutral-page-quality-20260702-001.answers.json`, and `research/manual-client-runs/manual-client-claude-plain-page-quality-20260702-001.answers.json`. Each attempt used the logged-in OpenClaw browser profile, a fresh Claude native incognito chat or fresh Claude chat with native incognito enabled, visible `Sonnet 5 Medium`, exact public target preflight, and bounded direct-origin review. Operator preflight requests are fixture readiness checks and are not counted as prompt-caused AI-client behavior. ## Comparison Matrix | Prompt | Prompt framing | Direct-origin evidence | Copied answer result | |---|---|---|---| | p26 `claude-head-signal-neutral-summary` | Asked about visible text, metadata, structured data, and code-like tokens. | Raw event `mr2aajph-xy92m1hx` confirmed the target-page hit; `mr2aaj2t-9op15mau` was ancillary `/robots.txt`. | 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` | Asked about page-preview and metadata signals. | 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` | Asked directly about source areas and code-like tokens. | 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. | | p32 `claude-neutral-page-quality-followup` | Asked for site-owner page-quality feedback but still mentioned details and code-like tokens. | Raw event `mr2p34bu-yn45ibc7` confirmed the target-page hit. | Described body-versus-head separation and metadata categories without repeating exact marker values. | | p36 `claude-plain-page-quality` | Asked for ordinary site-owner review without source-area, metadata, structured-data, code-like-token, marker-value, or hidden-field wording. | Raw event `mr31xowy-j9qdrdq1` confirmed the target-page hit; `mr31xohl-eo2hrr9o` was ancillary `/robots.txt`. | Described the page as a diagnostic AI Crawler Lab head-signal fixture and did not repeat exact marker values. | ## Interpretation The retrieval result is stable across this Claude series: all five attempts fetched the target page from the lab origin in the prompt window. The answer-surface result is prompt sensitive. The copied p28 answer contained exact page-head markers, while p32 and p36 moved away from marker-level copying even though Claude still recognized the fixture's body-versus-head diagnostic purpose. The p36 run is the clearest production-wording approximation in this series because it avoids terms that ask Claude to inspect metadata or code-like values. Its answer still showed category-level awareness, but it did not quote the exact visible body marker or the hidden page-head markers. That keeps the durable conclusion narrow: Claude could retrieve and interpret the diagnostic fixture under plainer wording, but the copied response did not expose exact marker values. JSON-LD remains unobserved in Claude's fetched view across this sequence. The served fixture contained `SCHEMA-MINT-88` during preflight, but no copied answer surfaced that marker. This is evidence about Claude's extracted view in these runs, not evidence that the live fixture lacked JSON-LD. ## Limitations - This comparison covers one Claude account context, Free plan, visible model selector, fixture, and narrow set of time windows. - The prompts supplied exact target URLs, so the tests measure direct URL opening and extraction rather than independent discovery. - The fixture is synthetic and diagnostic; p36 shows plainer task wording but still points Claude at an AI Crawler Lab URL. - Claude's fetched view is an extracted representation, not raw HTTP bytes or a browser DOM inspector. - There is one run per prompt variant. Later product routing, model selection, account tier, or tool changes could shift both retrieval and answer-surface behavior. ## Publication Thesis Verification - Thesis: Claude fetched the head-signal fixture in p26, p27, p28, p32, and p36, but exact page-head marker values appeared only in the more source-area-directed p28 copied answer; p32 and p36 stayed category-level. - Source: Findings 066, 071, 076, and 077; the five response artifacts; the three logged answer packets; and direct-origin raw events `mr2aajph-xy92m1hx`, `mr2anb9k-tx6bpus5`, `mr2c93wu-i61fpyko`, `mr2p34bu-yn45ibc7`, and `mr31xowy-j9qdrdq1`. - Method: Compare copied Claude answers with prompt wording, response artifacts, confirmation status, raw event ids, and bounded timestamp windows. - Bias: The fixture is purpose-built for metadata measurement, and even the plainer p36 task uses an exact lab URL that may reveal diagnostic context. - Consensus: Consistent with Finding 076 and Finding 077: direct-origin retrieval stayed confirmed while copied exact marker exposure weakened under less source-directed 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 five 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 matched plain head-signal prompts for ChatGPT, Gemini, Perplexity, and Copilot/Bing, and repeat p36-style wording across multiple Claude runs to separate stable extraction behavior from routing variability. ## Next steps - Prepare matched plain head-signal prompt packets for other AI-client surfaces once a small controlled-browser slot is available. - Add same-wording Claude repeats later to test whether p36's category-level answer surface is stable.