# Finding 057: Neutral hidden-signal comparison leaves Claude as the only target-page hit

## Date

2026-07-01

## Status

Published

## Summary

Findings 052-056 completed the p22 neutral hidden-signal prompt family across
ChatGPT, Claude, Gemini, Perplexity, and Copilot/Bing. The p22 prompt reused
the hidden structured-data fixture from p21, but asked for a neutral
site-owner page-quality summary instead of naming meta descriptions, JSON-LD,
or hidden fields in the requested output.

Across the completed p22 set, Claude was the only client with a confirmed
target-page hit. Claude returned `fetched:true`, opened one page, reported
visible code `VISIBLE-PLUM-47`, surfaced meta-description code
`META-AMBER-16`, and left JSON-LD code `SCHEMA-INDIGO-82` unobserved in the
copied answer. ChatGPT, Gemini, and Copilot/Bing were clean bounded no-hits
for the target and related discovery paths. Perplexity did not fetch the
target page, but PerplexityBot requested `/robots.txt` inside the prompt
window.

## What does this mean?

For site owners and researchers, this set shows why retrieval evidence has to come before content-visibility conclusions. In the neutral prompt variant, only Claude actually reached the tested page, and its answer again reflected visible text plus the meta description while leaving JSON-LD out. The other clients did not inspect the page in these attempts, so their answers cannot tell us whether they would have used visible text, metadata, or structured data if retrieval had succeeded.

## Method

This comparison reviews the published p22 finding set:

- ChatGPT: Finding 052,
  `manual-client-chatgpt-20260625-001-p22`.
- Claude: Finding 053,
  `manual-client-claude-20260625-001-p22`.
- Gemini: Finding 054,
  `manual-client-gemini-20260625-001-p22`.
- Perplexity: Finding 055,
  `manual-client-perplexity-20260625-001-p22`.
- Copilot/Bing: Finding 056,
  `manual-client-copilot-bing-20260626-001-p22`.
- Response artifacts:
  `research/manual-client-runs/browser-tasks/responses/*-p22.response.json`.
- Public evidence bundles:
  `public/api/v1/evidence/finding-052.json` through
  `public/api/v1/evidence/finding-056.json`.

Each underlying run used a fresh AI-client chat or native private/temporary
thread, logged the copied model answer with `npm run manual-client:log`, and
reviewed direct-origin events by `promptCode`, `sourcePromptId`,
`confirmedHitFromPrompt`, `rawEventIds`, and bounded timestamp windows.

## Comparison Matrix

| Client | p22 result | Direct-origin evidence | Content implication |
|---|---|---|---|
| ChatGPT | `fetched:false`, `pages_opened:0`; browsing attempt did not retrieve the page. | Clean bounded no-hit for exact attempt and fixture path. | No hidden-signal conclusion; the page was not fetched. |
| Claude | `fetched:true`, `pages_opened:1`; reported `VISIBLE-PLUM-47` and `META-AMBER-16`, with JSON-LD unobserved. | Confirmed `/robots.txt` and exact target-page hits from `Claude-User/1.0`. | Supports visible plus meta-description extraction for this run, but not JSON-LD exposure. |
| Gemini | `fetched:false`, `pages_opened:0`; could not directly open arbitrary external URLs outside its search interface. | Clean bounded no-hit for exact attempt, fixture path, `/robots.txt`, root, and other origin events. | No hidden-signal conclusion; the page was not fetched. |
| Perplexity | `fetched:false`, `pages_opened:0`; `Failed to fetch url content`. | No target-page hit; ancillary PerplexityBot `/robots.txt` request inside the window. | Evidence of prompt-window origin contact, not target-page inspection. |
| Copilot/Bing | `fetched:false`, `pages_opened:0`; client could not open or retrieve the target URL. | Clean bounded no-hit for exact attempt, fixture path, `/robots.txt`, and root. | No hidden-signal conclusion; the page was not fetched. |

## Interpretation

The p22 comparison narrows the hidden-signal claim rather than broadening it.
The neutral wording reduced prompt pressure to inspect named hidden fields, but
it also produced no target-page hit for ChatGPT, Gemini, Perplexity, or
Copilot/Bing. Those four runs are useful retrieval outcomes, not evidence
about which page fields those clients can read.

Claude is the only p22 run that can be compared at the content level. Its
answer matches the earlier p20/p21 pattern: visible page text and
meta-description text surfaced, while JSON-LD did not appear in the copied
answer. Because the p22 prompt did not ask for metadata or JSON-LD fields, the
Claude result strengthens the observation that its extraction view can expose
meta-description text even during a neutral page-quality task.

Perplexity remains important as a separate crawler-contact case. The client
reported a failed fetch and no target-page event appeared, but the origin logs
still captured a PerplexityBot `/robots.txt` request during the bounded prompt
window. That should be treated as ancillary origin activity, not as evidence
that Perplexity inspected the p22 fixture.

## Limitations

- This comparison covers one p22 attempt per client, with the logged account,
  tier, mode, and timestamp windows from Findings 052-056.
- The prompt supplied an exact target URL, so this is a direct-opening test,
  not an independent discovery test.
- The comparison separates answer content from direct-origin hits. A model's
  `fetched:true` or `fetched:false` field is not treated as proof without the
  matching origin review.
- No-hit clients cannot support visible/meta/schema field conclusions from
  this set.
- Claude's JSON-LD absence means the copied answer did not expose JSON-LD in
  this run; it does not prove the underlying service could never inspect
  JSON-LD.

## Publication Thesis Verification

- Thesis: In the completed p22 neutral hidden-signal set, Claude was the only
  client with a confirmed target-page hit; Claude exposed visible body text
  and meta-description text but not JSON-LD, while ChatGPT, Gemini, Perplexity,
  and Copilot/Bing did not fetch the target page.
- Source: Findings 052-056, copied answer artifacts, browser-task response
  files, and public evidence bundles for the same finding range.
- Method: Compare each client's p22 response fields, origin-review status,
  raw event ids, ancillary origin activity, and timestamp windows while
  separating exact target-page hits from `/robots.txt` contact.
- Bias: The comparison uses one run per client, with single accounts and
  observed product modes. Results can change by tier, model, region, time, or
  client retrieval state.
- Consensus: Consistent with the individual p22 thesis-verification sections:
  Claude confirmed the target hit and reported visible plus meta-description
  text; ChatGPT, Gemini, and Copilot/Bing were clean target no-hits; Perplexity
  had a target no-hit with ancillary robots contact.
- Invalidation: A corrected response artifact, mismatched raw event review,
  fixture serving error, or same-condition rerun that changes target-hit
  status or exposes JSON-LD would weaken the comparison.
- Verdict: Supported for the published p22 corpus. The comparison should
  remain scoped to these controlled-browser attempts.
- Additional tests suggested: compare p20, p21, and p22 retrieval transitions
  for ChatGPT and Claude, especially ChatGPT's shift from confirmed p20/p21
  hits to a p22 no-hit under neutral wording.

## Next steps

- Compare ChatGPT's p20/p21 confirmed hits against its p22 no-hit to isolate
  whether neutral prompt wording changed retrieval behavior.
- Keep Perplexity and Copilot/Bing follow-ups focused on retrieval-mode
  reliability until a target-page hit exists.