# Finding 092: ChatGPT fetched both URL shapes across ten neutral readability runs ## Date 2026-07-03 ## Status Published ## Summary This comparison reviews three completed ChatGPT neutral readability groups for `/lab/reading/hidden-structured-data-conflict`: the p29-p31 full-correlation repeat trio, the p33-p35 short id-only repeat trio, and the p39-p42 mixed-order URL-shape packet. | Series | URL shape | Attempts | Result | Raw event ids | |---|---|---:|---|---| | p29-p31 | full-correlation query string | 3 | all `fetched:true` | `mr2gm1ag-c3itq40v`, `mr2iptim-5hq0zuxp`, `mr2kwg6x-dpw37d24` | | p33-p35 | short id-only query string | 3 | all `fetched:true` | `mr2rhoel-km9938ef`, `mr2rt7zm-2zhwxm1o`, `mr2tebcf-67sgcb24` | | p39-p42 | mixed full/short sequence | 4 | all `fetched:true` | `mr3tu40o-5j6n23nt`, `mr3vxsp5-0t1do4up`, `mr3y3rob-z4q8fk3y`, `mr40985g-99xvskmd` | All ten attempts were submitted in fresh ChatGPT Temporary Chats from the logged-in OpenClaw-controlled browser profile. Every response artifact records `confirmedHitFromPrompt:true` with a prompt-window `ChatGPT-User/1.0` target-page hit. Across the copied answers, ChatGPT reported visible page text or the visible `VISIBLE-PLUM-47` marker, but did not report exact hidden meta-description or JSON-LD marker values. ## What does this mean? For site owners, SEO/AEO teams, publishers, and researchers, these ten runs show that this ChatGPT setup repeatedly opened the same supplied page whether the target URL used a long tracking-style query string or only a short id parameter. URL simplification did not explain the earlier isolated no-hit and did not change the copied answer into a hidden-metadata or structured-data answer. The practical takeaway is that direct URL-opening tests need repeated runs and origin logs before treating URL shape as the cause of a fetch or no-fetch result. ## Method This comparison reviewed the completed finding chain and response artifacts: - Finding 070 for p29-p31 full-correlation repeats. - Finding 075 for p33-p35 short id-only repeats. - Finding 091 for the p39-p42 mixed-order packet. - Response artifacts under `research/manual-client-runs/browser-tasks/responses/` for p29-p31, p33-p35, and p39-p42. - Logged answer packets: `research/manual-client-runs/manual-client-chatgpt-neutral-repeat-20260701-001.answers.json`, `research/manual-client-runs/manual-client-chatgpt-neutral-short-repeat-20260702-001.answers.json`, and `research/manual-client-runs/manual-client-chatgpt-neutral-url-shape-mixed-20260702-001.answers.json`. The review compared URL shape, source prompt id, copied `fetched` status, `pages_opened`, `evidence_quote`, `confirmedHitFromPrompt`, raw event ids, bounded timestamp windows, and answer-side marker exposure. Operator preflight requests remain fixture-readiness checks and are not counted as prompt-caused AI-client behavior. ## Observed Result The full-correlation p29-p31 trio produced three confirmed target hits. The short id-only p33-p35 trio also produced three confirmed target hits. The later mixed-order p39-p42 packet alternated full, short, full, short URL shapes and again produced four confirmed target hits. The copied answer pattern was stable at the level this test can observe. The answers repeatedly described the page as a short reading or fixture page and reported visible content. None of the ten saved answers reported the hidden meta-description marker `META-AMBER-16` or JSON-LD marker `SCHEMA-INDIGO-82`. That answer-side observation is separate from retrieval: the origin evidence confirms page contact, while the copied answers only show what the client chose to expose. One minor variation remains: p39 reported `pages_opened:2`, while p40-p42 reported `pages_opened:1`; the p29-p31 artifacts recorded the opened page URL in `pages_opened`, and the p33-p35 artifacts recorded numeric `pages_opened:1`. Direct-origin review still found one exact target-page hit in every bounded prompt window. ## Interpretation Across these three groups, URL shape was not a simple retrieval divider for ChatGPT. Long full-correlation URLs and short id-only URLs both produced repeat target hits, and the mixed-order packet reduced the chance that the result was only a day-level or ordering artifact. The comparison weakens a narrow hypothesis that the earlier ChatGPT no-hit was caused by the long full-correlation URL alone. It does not prove that URL parameters never matter, that ChatGPT's retrieval backend treated both URL shapes identically, or that every ChatGPT account, tier, mode, region, or future product state will behave the same way. The field-exposure result is also scoped. The copied answers stayed visible page scoped in these ten runs, but that does not prove hidden metadata or structured data were unavailable to every internal component of the retrieval path. ## Limitations - All ten runs used one ChatGPT account context, the Free tier, and native Temporary Chat mode. - The prompt supplied the exact target URL, so this measures direct URL opening rather than independent discovery, search ranking, or crawl traversal. - The short id-only URL attempts still included prompt code and source prompt metadata elsewhere in the submitted prompt text. - The runs occurred across two adjacent days, and product routing or account state could still have shifted during the series. - Copied answer content is not a complete view into the internal retrieved representation. ## Publication Thesis Verification - Thesis: Across p29-p31, p33-p35, and p39-p42, ChatGPT fetched both full-correlation and short id-only target URL shapes under neutral readability wording, while copied answers stayed visible-page scoped. - Source: Findings 070, 075, and 091; the ten response artifacts; the three logged answer packets; and direct-origin raw events `mr2gm1ag-c3itq40v`, `mr2iptim-5hq0zuxp`, `mr2kwg6x-dpw37d24`, `mr2rhoel-km9938ef`, `mr2rt7zm-2zhwxm1o`, `mr2tebcf-67sgcb24`, `mr3tu40o-5j6n23nt`, `mr3vxsp5-0t1do4up`, `mr3y3rob-z4q8fk3y`, and `mr40985g-99xvskmd`. - Method: Compare response artifacts and logged answer packets for URL shape, prompt family, fresh-chat mode, copied answer fields, bounded timestamp windows, confirmation status, raw event ids, and visible versus hidden marker exposure. - Bias: The result is single-client, single-account, direct-URL, and short-window evidence. Product routing, account tier, model selector state, prompt wording, cache behavior, and hidden source metadata in the prompt can all affect outcomes. - Consensus: Consistent with the three lower-level comparison findings and their individual run-level findings. The mixed-order packet agrees with the two earlier same-shape repeat trios. - Invalidation: A corrected origin review that removes any prompt-window raw event, a response artifact mismatch, evidence of prompt contamination, fixture-serving errors, or later same-condition repeats showing stable shape-specific no-hit behavior would weaken this comparison. - Verdict: Supported for these ten ChatGPT Temporary Chat runs. The evidence supports a scoped direct URL-opening conclusion, not a universal claim about all ChatGPT retrieval behavior or all URL shapes. - Additional tests suggested: repeat a mixed-order packet on another day, compare another ChatGPT mode or tier if available, and test whether the same URL-shape pattern appears on a different fixture surface. ## Next steps - Prepare a later mixed-order repeat if ChatGPT retrieval starts varying again. - Keep future URL-shape findings explicit about direct URL opening versus discovery, search, or crawl traversal.