# Finding 080: Gemini fetched the plain head-signal page and quoted the visible marker

## Date

2026-07-02

## Status

Published

## Summary

Gemini was prompted in a fresh new chat with
`manual-client-gemini-plain-head-signal-20260702-001-p37`, the plain
site-owner review prompt for the `/lab/reading/head-signal-isolation` fixture.
The prompt asked for a brief page review and did not ask Gemini to inspect
source areas, page-head metadata, structured data, code-like tokens, marker
values, or hidden fields.

Gemini returned `fetched:true`, `pages_opened:1`, described the page as a
simulated AI Crawler Lab head-signal test fixture, and quoted the visible body
marker `VISIBLE-QUARTZ-39`. The copied answer mentioned page-head signal
categories such as meta descriptions, OpenGraph, canonical links, and JSON-LD
at a category level, but it did not repeat exact page-head marker values such
as `META-CEDAR-71`, `OG-COBALT-23`, `OG-RUBY-62`, `CANONICAL-PEAR-54`, or
`SCHEMA-MINT-88`. Direct-origin raw event `mr3ahz4q-7kfpdrht` confirmed a
target-page hit inside the bounded prompt window.

## What does this mean?

For site owners, SEO/AEO teams, publishers, and researchers, this run shows that Gemini can retrieve a directly supplied diagnostic page under ordinary site-owner review wording, and its answer may include exact visible body text even when exact hidden page-head marker values are not copied back. That distinction matters because a successful fetch, category-level awareness of page-head signals, visible marker quotation, and hidden marker disclosure are different behaviors that should not be collapsed into one claim.

## Method

- Controlled-browser task:
  `research/manual-client-runs/browser-tasks/manual-client-gemini-plain-head-signal-20260702-001-p37.browser-task.json`.
- Prompt packet:
  `research/manual-client-runs/manual-client-gemini-plain-head-signal-20260702-001.prompts.json`.
- Response artifact:
  `research/manual-client-runs/browser-tasks/responses/manual-client-gemini-plain-head-signal-20260702-001-p37.response.json`.
- Logged answer packet:
  `research/manual-client-runs/manual-client-gemini-plain-head-signal-20260702-001.answers.json`.
- Prompt window:
  `2026-07-02T09:15:58.000Z` to `2026-07-02T09:16:14.310Z`.

The exact public target URL had been prepared in the p37 browser-task packet.
The prompt was submitted in a fresh Gemini new chat under the kai bot work
account, with a visible Pro badge and the model selector showing `Flash`.
OpenClaw browser `act` calls failed before submission with
`action targetId must match request targetId`, so the prompt was submitted by
direct Chrome DevTools Protocol operations against a fresh Gemini tab.

## Evidence

| Evidence | Result |
|---|---|
| Model answer | `fetched:true`, `pages_opened:1`; described the page as an AI Crawler Lab head-signal fixture. |
| Gemini UI state | Fresh new Gemini chat; visible account `kai bot`; Pro badge; model selector `Flash`. |
| Direct-origin target-page event | `mr3ahz4q-7kfpdrht` at `2026-07-02T09:16:06.640Z` from user-agent `Google` on `/lab/reading/head-signal-isolation?id=manual-client-gemini-plain-head-signal-20260702-001-p37`. |
| Network identity | Reverse DNS returned `rate-limited-proxy-108-177-76-168.google.com` and forward-confirmed `108.177.76.168`; classifier provider hint was `Google` with low identity confidence. |

```json
{
  "timestampWindow": {
    "startedAt": "2026-07-02T09:15:58.000Z",
    "endedAt": "2026-07-02T09:16:14.310Z"
  },
  "confirmedHitFromPrompt": true,
  "rawEventIds": [
    "mr3ahz4q-7kfpdrht"
  ]
}
```

## Interpretation

This p37 run confirms that Gemini retrieved the head-signal fixture under
plain site-owner review wording. Unlike ChatGPT p37, Gemini copied the exact
visible body marker into the response. Like ChatGPT p37 and Claude p36,
Gemini did not repeat exact hidden page-head marker values in the copied
answer.

The result should be read as a single Gemini controlled-browser observation,
not a broad Gemini product claim. It is useful as a cross-client comparison
point for the p36/p37 plain head-signal series because it separates direct
retrieval, visible-marker quotation, and exact hidden-head-marker copying.

## Limitations

- This is one Gemini run from one account, tier, visible model selector, and
  time window.
- The prompt supplied the exact target URL, so this measures direct URL
  opening rather than independent discovery.
- The answer describes Gemini's fetched/extracted view, not raw HTTP bytes or
  a browser-rendered DOM.
- The fixture intentionally contains diagnostic language and embedded
  AI-reader text, which can make the page's purpose obvious after retrieval.
- The raw event identifies a Google user-agent and forward-confirmed Google
  reverse DNS, but the local classifier kept identity confidence low; the
  client-surface link comes from the controlled Gemini prompt window.

## Publication Thesis Verification

- Thesis: In the p37 plain site-owner review prompt, Gemini fetched the
  head-signal fixture, quoted the visible body marker, and did not repeat exact
  hidden page-head marker values in the copied answer.
- Source: The p37 browser-task artifact, response artifact, logged answer
  packet, Gemini UI snapshot context, and direct-origin raw event
  `mr3ahz4q-7kfpdrht`.
- Method: Compare the copied Gemini 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 from a synthetic fixture. The
  prompt is plain relative to source-area prompts but still names
  AI-assistant readability and an exact lab URL.
- Consensus: Consistent with prior confirmed-hit runs showing that direct
  retrieval and exact hidden marker copying need separate evidence; it differs
  from ChatGPT p37 by copying the visible marker.
- Invalidation: A response-artifact mismatch, corrected origin review removing
  raw event `mr3ahz4q-7kfpdrht` from the prompt window, or fixture-serving
  evidence showing different marker content during the run would weaken this
  finding.
- Verdict: Supported for the p37 run. It confirms direct-origin retrieval and
  visible marker quotation under plain wording, while keeping exact hidden
  page-head marker exposure unobserved in the copied answer.
- Additional tests suggested: run the prepared Perplexity and Copilot/Bing p37
  tasks in fresh client chats, then compare all non-Claude p37 results against
  Claude p36 and ChatGPT p37.

## Next steps

- Continue the p37 cross-client packet with Perplexity and Copilot/Bing.
- Publish a p36/p37 cross-client comparison after the remaining p37 attempts
  have logged answers and direct-origin reviews.