# Finding 051: Structured-data conflict comparison shows visible text dominated ChatGPT while Claude exposed meta text ## Date 2026-07-01 ## Status Published ## Summary Findings 041-050 completed the p20 and p21 structured-data conflict prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Copilot/Bing. The p20 fixture visibly told the assistant that metadata and JSON-LD disagreed with the page body. The p21 fixture removed that visible conflict cue while still serving separate visible, meta-description, and JSON-LD codes in the raw page. Across the two completed sets, ChatGPT and Claude were the only clients with confirmed target-page hits in both variants. ChatGPT reported only the visible page code in both p20 and p21. Claude reported the visible page code and the meta-description code in both p20 and p21, while leaving JSON-LD unobserved in its extracted content. Gemini, Perplexity, and Copilot/Bing did not produce confirmed target-page hits in these p20/p21 runs, so their structured-data treatment cannot be inferred from this pair. ## What does this mean? For site owners and researchers, these paired tests show that a successful assistant page fetch does not mean every hidden page signal influenced the answer. In this lab, ChatGPT stayed on visible body text even when metadata and JSON-LD were present, while Claude also surfaced the meta description but still did not expose JSON-LD. Other clients failed or declined retrieval in these runs, so there was no evidence that they inspected either visible or hidden page content. ## Method This comparison reviews the published p20 and p21 finding set: - p20 visible structured-data conflict: Findings 041-045. - p21 hidden structured-data conflict: Findings 046-050. - Response artifacts: `research/manual-client-runs/browser-tasks/responses/*-p20.response.json` and `research/manual-client-runs/browser-tasks/responses/*-p21.response.json`. - Public evidence bundles: `public/api/v1/evidence/finding-041.json` through `public/api/v1/evidence/finding-050.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 | p20 outcome | p21 outcome | Comparison | |---|---|---|---| | ChatGPT | Confirmed target hit. Reported `VISIBLE-SILVER-30`; metadata and JSON-LD not exposed. | Confirmed target hit. Reported `VISIBLE-PLUM-47`; metadata and JSON-LD not exposed. | Removing the visible conflict cue did not make hidden metadata or JSON-LD appear in the answer. | | Claude | Confirmed target hit plus `/robots.txt`. Reported `VISIBLE-SILVER-30` and `META-TEAL-55`; JSON-LD unobserved. | Confirmed target hit plus `/robots.txt`. Reported `VISIBLE-PLUM-47` and `META-AMBER-16`; JSON-LD unobserved. | Claude exposed meta-description text in both variants, including when the visible page no longer announced a conflict. | | Gemini | Bounded target no-hit. No exact attempt, fixture-path, or `/robots.txt` event during the prompt window. | Bounded target no-hit. No exact attempt, fixture-path, or `/robots.txt` event during the prompt window. | The pair shows retrieval failure or refusal, not structured-data handling. | | Perplexity | Target-page no-hit with ancillary PerplexityBot requests for `/robots.txt` and `/`. | Clean bounded no-hit with no exact target, `/robots.txt`, root, or Perplexity-origin event. | Ancillary origin activity appeared in p20 but not p21; neither run fetched the target page. | | Copilot/Bing | Clean bounded no-hit. The client reported invalid tool invocation during retrieval attempts. | Clean bounded no-hit. The client said it could not open or fetch the target URL. | Both runs are no-retrieval outcomes, so hidden-field behavior remains untested for this surface. | ## Interpretation The useful split is between target-page retrieval and hidden-signal exposure. ChatGPT and Claude both opened the supplied page in both variants, but their answer surfaces differed. ChatGPT stayed with visible body text and did not expose meta-description or JSON-LD codes. Claude exposed the meta-description code as a separate extracted field in both variants, including p21 where the visible copy did not announce a conflict. Neither confirmed-hit client exposed the JSON-LD code. The p21 variant strengthens the interpretation of Claude's meta-description behavior because the visible page no longer tells the assistant that metadata contains a different code. It does not prove raw HTML or DOM access, because Claude described a markdown-like extraction view and left JSON-LD unobserved. For ChatGPT, the p21 result strengthens the visible-content interpretation: the hidden meta and JSON-LD codes were present in preflighted raw HTML, but the answer still reported only the visible code. Gemini, Perplexity, and Copilot/Bing should not be placed in a hidden-signal matrix from these attempts. Their p20/p21 answers and origin reviews did not confirm target-page retrieval. The Perplexity p20 ancillary `/robots.txt` and root activity is evidence of prompt-window origin contact, not evidence that the structured-data target page was fetched or inspected. ## Limitations - This comparison covers one p20 and one p21 attempt per client, with the logged account, tier, mode, and time windows from Findings 041-050. - The prompts supplied exact target URLs, so these are direct-opening tests, not independent discovery tests. - 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. - The p21 prompt still explicitly asks for visible, meta-description, and JSON-LD fields. It removes the visible conflict cue, but it is still a directed hidden-field inspection prompt. - The absence of JSON-LD in answers does not prove JSON-LD was inaccessible to the underlying service in general; it only shows that the copied answers did not expose it in these runs. ## Publication Thesis Verification - Thesis: In the completed p20/p21 structured-data conflict sets, ChatGPT exposed only visible page text on confirmed hits, Claude exposed visible text plus meta-description text, and neither exposed JSON-LD; Gemini, Perplexity, and Copilot/Bing did not produce target-page hits for this pair. - Source: Findings 041-050, copied answer artifacts, browser-task response files, and public evidence bundles for the same finding range. - Method: Compare each client's p20 and p21 response fields, origin-review status, raw event ids, and timestamp windows while separating exact target-page hits from ancillary origin activity. - Bias: The comparison uses one run per client per variant, with single accounts and observed product modes. Results can change by tier, model, region, time, or client retrieval state. - Consensus: Consistent with the individual p20 and p21 thesis-verification sections: ChatGPT p20/p21 both visible-only, Claude p20/p21 both visible plus meta-description, Gemini p20/p21 no-hit, Copilot p20/p21 no-hit, and Perplexity no target-page hit in either run. - Invalidation: A corrected response artifact, mismatched raw event review, fixture serving error, or same-condition rerun that exposes JSON-LD or changes target-hit status would weaken the comparison. - Verdict: Supported for the published p20/p21 corpus. The comparison should remain scoped to these controlled-browser attempts. - Additional tests suggested: add a still more neutral hidden-field prompt that asks for a short page-quality answer without naming metadata or JSON-LD, and rerun only clients that can produce confirmed target-page hits. ## Next steps - Prepare a neutral hidden-signal fixture or prompt variant for ChatGPT and Claude that does not name meta-description or JSON-LD in the prompt. - Keep Perplexity and Copilot/Bing reruns as retrieval-mode checks rather than hidden-field comparisons until they produce target-page hits.