# Finding 061: ChatGPT repeated short-URL retrieval under a neutral support-answer prompt

## Date

2026-07-01

## Status

Published

## Summary

The ChatGPT p25 neutral retrieval-stability rerun used the same
`/lab/reading/hidden-structured-data-conflict` fixture as p22-p24. Like p23,
it used a short id-only target URL, but changed the task framing to a neutral
support-answer request. ChatGPT returned `fetched:true`, `pages_opened:1`,
reported `VISIBLE-PLUM-47`, and said the meta description and JSON-LD source
were not exposed in the retrieved page view.

Direct-origin raw event `mr25un5s-guspwbjl` confirmed the exact target-page hit
from `ChatGPT-User/1.0` inside the bounded
`2026-07-01T14:18:03.653Z` to `2026-07-01T14:18:26.772Z` prompt window. The
operator preflight event for the same attempt occurred at
`2026-07-01T14:14:41.892Z` and is not counted as prompt-caused evidence.

## What does this mean?

For site owners and researchers, ChatGPT reached the same lab page again when the prompt used ordinary support-answer wording and a short URL. Across the latest short-URL reruns, ChatGPT consistently opened the target and reported only visible page text, so the earlier p22 no-hit looks more like retrieval variability or prompt-routing behavior than a stable inability to access this fixture.

## Method

- Prepared packet:
  `research/manual-client-runs/manual-client-chatgpt-neutral-stability-20260701-001.prompts.json`.
- Browser task:
  `research/manual-client-runs/browser-tasks/manual-client-chatgpt-neutral-stability-20260701-001-p25.browser-task.json`.
- Response artifact:
  `research/manual-client-runs/browser-tasks/responses/manual-client-chatgpt-neutral-stability-20260701-001-p25.response.json`.
- Logged answer packet:
  `research/manual-client-runs/manual-client-chatgpt-neutral-stability-20260701-001.answers.json`.
- Prompt code:
  `manual-client-chatgpt-neutral-stability-20260701-001-p25`.
- Source prompt:
  `chatgpt-neutral-support-answer-short-url`.

The exact target URL was preflighted before submission and returned the
expected visible token `VISIBLE-PLUM-47`, meta-description token
`META-AMBER-16`, and JSON-LD token `SCHEMA-INDIGO-82`. The prompt was then
submitted in a fresh native ChatGPT Temporary Chat from the logged-in OpenClaw
browser profile. The visible client context showed the `kai bot` Free account,
the `ChatGPT` model selector, and native Temporary Chat state. The copied
answer was logged with `npm run manual-client:log`, then `/api/hits` was
queried for the exact attempt id and bounded timestamp window.

## Observed Result

ChatGPT returned compact JSON with:

- `fetched:true`
- `pages_opened:1`
- visible token `VISIBLE-PLUM-47`
- evidence quote including `The page quality code is VISIBLE-PLUM-47.`
- limitation text saying only the visible page content available through the
  retrieved view was accessible and metadata or JSON-LD source were not exposed

The reviewed direct-origin event was:

| Raw event id | Timestamp | Path | User agent |
|---|---|---|---|
| `mr25un5s-guspwbjl` | `2026-07-01T14:18:13.335Z` | `/lab/reading/hidden-structured-data-conflict` | `ChatGPT-User/1.0` |

## Interpretation

The p25 result reinforces the p23 short-URL result and the p24 full-URL result:
ChatGPT can retrieve this fixture under neutral wording after the earlier p22
clean no-hit. The copied p25 answer stayed visible-only, matching p21, p23,
and p24, so it strengthens the retrieval-stability interpretation without
changing the hidden-field exposure interpretation.

The p25 raw event came from `ChatGPT-User/1.0` on `20.215.220.176`, which
matched the official OpenAI `ChatGPT-User` range cache. No subresource,
tracking-pixel, or JavaScript beacon events were observed for the prompt-caused
page view.

## Limitations

- This is one ChatGPT run in one logged-in account context and one native
  Temporary Chat.
- The test supplied the exact target URL, so it measures direct URL opening,
  not independent discovery.
- The p25 result differs from p22 in prompt wording and possible product
  routing state, so it does not explain all retrieval variability by itself.
- The answer reports rendered-view limitations. It does not prove that other
  ChatGPT modes, models, account tiers, or future retrieval paths cannot see
  metadata or JSON-LD.

## Publication Thesis Verification

- Thesis: ChatGPT p25 retrieved the hidden-signal target under neutral
  support-answer wording and reported only visible page content.
- Source: The p25 browser-task artifact, response artifact, logged answer
  packet, `/api/hits` raw event `mr25un5s-guspwbjl`, and the preflight record
  for the exact target URL.
- Method: Compare copied model JSON with the bounded direct-origin event
  window, excluding the preflight event because it occurred before prompt
  submission.
- Bias: Single-run product behavior can change with model routing, account
  state, client-side search/tool availability, prompt wording, and target URL
  shape.
- Consensus: Consistent with ChatGPT p21, p23, and p24 confirmed visible-only
  retrieval, and contrary to p22's clean no-hit.
- Invalidation: A corrected event review that removes `mr25un5s-guspwbjl`
  from the prompt window, a response metadata mismatch, fixture-serving error,
  or reruns showing a different stable pattern would weaken the finding.
- Verdict: Supported for this p25 run only. It strengthens the view that p22's
  no-hit was not a stable inability to retrieve this fixture, but it does not
  fully explain why p22 differed.
- Additional tests suggested: publish a focused p22-p25 comparison of ChatGPT
  neutral retrieval stability.

## Next steps

- Compare p22-p25 as a focused ChatGPT retrieval-stability series.