# Finding 071: Claude's neutral page-quality run fetched the head-signal page without repeating marker values

## Date

2026-07-02

## Status

Published

## Summary

Claude was prompted in a fresh native incognito chat with
`manual-client-claude-neutral-page-quality-20260702-001-p32`, the neutral
page-quality follow-up for the `head-signal-isolation` fixture. The target page
contains visible body marker `VISIBLE-QUARTZ-39`, meta-description marker
`META-CEDAR-71`, OpenGraph markers `OG-COBALT-23` and `OG-RUBY-62`, canonical
marker `CANONICAL-PEAR-54`, and JSON-LD marker `SCHEMA-MINT-88`.

Claude returned `fetched:true`, `pages_opened:1`, and described the page as a
test fixture that separates page-head metadata from visible body copy. Unlike
the earlier source-area prompt, the copied p32 answer did not repeat exact
marker values. It said the visible body included "a marker string", noted that
canonical, meta-description, and OpenGraph tags existed in the fetched markup,
and explicitly avoided confirming or repeating the specific head-metadata code
values as observed truth. Direct-origin raw event `mr2p34bu-yn45ibc7` confirmed
a Claude-User target-page hit inside the bounded prompt window.

## What does this mean?

For site owners, SEO/AEO teams, and researchers, this run shows that neutral wording can still cause Claude to retrieve a page and recognize that page-head metadata exists, but it may not quote the exact hidden marker values unless the prompt asks more directly for source areas or code-like tokens. Retrieval, metadata awareness, and marker-level extraction are separate behaviors, so a page can be fetched successfully without every hidden signal appearing in the assistant's copied answer.

## Method

- Controlled-browser task:
  `research/manual-client-runs/browser-tasks/manual-client-claude-neutral-page-quality-20260702-001-p32.browser-task.json`.
- Prompt packet:
  `research/manual-client-runs/manual-client-claude-neutral-page-quality-20260702-001.prompts.json`.
- Response artifact:
  `research/manual-client-runs/browser-tasks/responses/manual-client-claude-neutral-page-quality-20260702-001-p32.response.json`.
- Logged answer packet:
  `research/manual-client-runs/manual-client-claude-neutral-page-quality-20260702-001.answers.json`.
- Prompt window:
  `2026-07-01T23:16:30.925Z` to `2026-07-01T23:17:07.638Z`.

The operator first preflighted the exact public target URL and confirmed a 200
response containing the expected visible, meta-description, OpenGraph,
canonical, and JSON-LD markers. The prompt was then submitted in a new Claude
native incognito thread under the Kai Free plan with visible model selector
`Sonnet 5 Medium`.

## Evidence

| Evidence | Result |
|---|---|
| Model answer | `fetched:true`, `pages_opened:1`; described the page title, visible body marker presence, page-head metadata categories, and tracking-pixel tag presence without quoting exact marker values. |
| Claude UI fetch state | Showed `Fetching from ai-crawler-lab.kaistone.ai` during the response. |
| Direct-origin target-page event | `mr2p34bu-yn45ibc7` at `2026-07-01T23:16:41.530Z` from `Claude-User/1.0` on `/lab/reading/head-signal-isolation`. |

## Interpretation

The p32 neutral page-quality prompt confirms that Claude could fetch the same
head-signal fixture from a fresh incognito chat and detect the page's broad
body-versus-head separation. It does not provide marker-level extraction
evidence comparable to p28 because the copied answer avoided repeating the
specific head-metadata code values and did not quote `VISIBLE-QUARTZ-39`.

This makes p32 useful as a prompt-framing result. The earlier p28 source-area
prompt showed exact meta-description, OpenGraph, and canonical markers in the
copied answer. The p32 neutral prompt showed retrieval plus category-level
awareness of those areas, but not copied marker values. The difference should
be treated as prompt-variant evidence, not as proof that the underlying page
changed.

## Limitations

- This is one Claude run from one account, tier, visible model selector, and
  time window.
- The prompt supplied the exact target URL, so this measures direct URL
  opening, not independent discovery.
- The answer describes Claude's fetched/extracted view, not raw HTTP bytes or a
  browser-rendered DOM.
- The prompt still asked for "details and code-like tokens", so it is more
  neutral than p28 but not a fully ordinary user question.
- The direct-origin event confirms target-page retrieval, but the copied answer
  remains the source of claims about which fields Claude exposed to the user.

## Publication Thesis Verification

- Thesis: In the p32 neutral page-quality follow-up, Claude fetched the target
  page and recognized the head-signal fixture structure, but the copied answer
  did not repeat exact visible, meta-description, OpenGraph, canonical, or
  JSON-LD marker values.
- Source: The p32 browser-task artifact, response artifact, logged answer
  packet, browser transcript, exact-target preflight output, and direct-origin
  raw event `mr2p34bu-yn45ibc7`.
- Method: Compare the copied Claude answer and visible browser state with the
  exact attempt id, prompt code, timestamp window, and raw server-side event.
- Bias: Single-account, single-run evidence. The prompt is a neutral
  page-quality variant but still mentions code-like tokens, which may shape
  whether Claude describes markers at all.
- Consensus: Consistent with Findings 063-066 on Claude target-page retrieval
  and broad head-signal exposure, while showing weaker marker-level exposure
  under less source-area-directed wording.
- Invalidation: A response-artifact mismatch, corrected origin review removing
  the Claude-User target-page hit, or fixture-serving evidence showing
  different markers during the run would weaken this finding.
- Verdict: Supported for the p32 run. It confirms retrieval and
  category-level head-signal awareness, while keeping exact marker exposure
  unobserved in the copied answer.
- Additional tests suggested: compare p26-p28 with p32 directly and prepare an
  even plainer production-user prompt that avoids "code-like tokens" entirely.

## Next steps

- Publish a focused p26-p28 versus p32 comparison for Claude prompt-framing
  effects.
- Prepare a plainer same-fixture Claude follow-up that asks for ordinary
  page-quality feedback without mentioning code-like tokens.