# Stage 1 / Stage 2: response to the external review

Reviewed and revised on **2026-11-02**. All recommendations were evaluated regardless of the review's priority labels. No new AI answers were collected. Original prompts, response records, C/M/R coding, selected outcomes and unsuccessful attempts are unchanged.

| Recommendation | Decision and implementation |
| --- | --- |
| Keep the existing experiments and raw evidence | Accepted. Stage 1 retains 12 selected answers, 10 other submissions, 37 captures and 142 entity-by-answer observations. Stage 2 retains six selected cells and all ten submissions, including the four earlier C technical/capture problems. |
| Correct the Q3 creator-overlap label | Accepted. The 2/20 (10%) statistic describes mixed recommended entities/resources, including a paper and community as the two shared resources. The original creator-seeking question remains verbatim. No claim that 20 individual creators were compared. |
| Qualify 97.9% versus 51.3% | Accepted. The numerator is observed own/bylined-content citations among coded Recommendation=true entity-by-answer records: 47/48 and 39/76. Third-party support is outside this indicator; capture and coding coverage limit it. It is not answer accuracy or an inherent platform preference. |
| Rename Repeat Consistency and foreground control differences | Accepted. The reader uses **Observed Repeat Overlap**, with a prominent qualification beside the chart and pair-specific dates/effort differences. ChatGPT Q2 spans dates; Q3 changes date and Extra High to Medium. Grok backend settings were undisclosed. Two repeats do not estimate stable consistency. |
| Add entity-type statistics and avoid interpreting list length as quality | Accepted. An explicit post-hoc mapping covers all 56 strict recommended names, with seven categories. Separate tables show unique entities and entity-by-answer records, for all questions and Q3. Original entity boundaries are retained; the assistant classification does not establish expertise or independence. |
| Audit representative recommendation quality | Accepted with a bounded scope. Ten purposively selected Q3 recommendations, five per platform, were checked against one recorded source page each. Existence, relevance, original observations, methods/sample information and solo-builder fit are separate fields. The sample is not random or a platform-quality ranking; no human validation or replication is claimed. |
| Collapse raw evidence by default | Adapted to the user's existing requirement that headings be expanded. Stage 1 and Stage 2 now have short readers and separate complete technical appendices; evidence remains expanded there. No raw evidence was deleted. |
| Add the later positive X native search observation | Accepted with provenance limits. The user confirmed the two author-scoped searches, Latest and All users on November 2. The actual earlier observation date and screenshot were not supplied. The report says **user-reported follow-up**, not independently verified success on the experiment date. An independent November 2 browser attempt showed only Loading, so it is an unresolved access check, not a negative search result. |
| Keep later X checks out of the October 31 six-cell results | Accepted. A separate post-publication file and section explicitly leave the original A/B/C outcomes unchanged. Author-scoped X search also supplies identity, unlike the AI discovery prompts. It does not explain why the AI answers did not find the post. |
| Put one selected response per cell next to the result table | Accepted. The table's adjacent note states there is one selected complete answer per condition/environment, with no success/failure-rate estimate. The manual C replacement and control differences remain visible. |
| Qualify ChatGPT B actual search execution | Accepted. The Search control and answer's self-report are distinct from independently observed queries. B queries were not captured independently; the reader does not claim both products performed independently verified exact searches. |
| Foreground ChatGPT C manual replacement provenance | Accepted. The later user-operated C-R5 is identified near its result. Exact question, complete body and actual search evidence were corroborated; model/time are user-confirmed and session-isolation preflight was not independently captured. |
| Simplify collection summary and remove ambiguous zero-failure phrasing | Accepted. The reader says six planned cells, ten submissions, six complete selected answers, and four earlier C technical/capture problems. Selected completeness is not confused with the entire collection history. |
| Move technical collection history off the reader's findings | Accepted. Recovery mechanics, full attempts and native evidence remain in the appendix; the reader keeps only the collection differences necessary to assess the outcomes. |
| Move Artur Ferreira C/M/R example to supporting observations | Accepted. The full example remains in the appendix, separately from the core access/exact/natural-discovery findings. |
| Reserve a separate November 7 B/C retest section | Accepted as a **proposal**, with no invented schedule or results. The user supplied no confirmation of this date. A future-check section and empty record are reserved; no automation or new queries were created. Actual later runs must be separately dated/versioned. |
| Connect all three lessons with a simple shared navigation | Accepted. Every reader and appendix links to all three experiments, with a visible current-stage marker. Stage 1 stays at the existing root URL. There is no forced relocation of previously shared links. |
| Record experiment date, first published, last revised, version and revision note | Accepted. A shared version file, per-report note and changelog distinguish these dates. First publication is grounded in the public Git history and retained edition records; older versions are labeled retrospectively. Stage 3's latest change is navigation/version metadata only. |
| Make the initial reading experience concise | Accepted. Each reader leads with the question, three bounded findings, essential figures/table and limitations. All detailed sources, full answers and failures remain accessible through the appendix and downloads. Notes and statistical definitions use blockquotes. |
| Validate ZIPs, internal files and real navigation | Accepted. Archives are checked for integrity and exact byte agreement with the allowlisted public data. Relative links, legacy fragment routing, shared navigation, public downloads and desktop/mobile layouts are verified separately. Verification evidence is retained locally. |
| Preserve redaction and avoid new public personal information | Accepted. Public identity and the Stage 2 target remain as previously authorized. Private conversations, local paths, screenshots containing account UI, credentials and unrelated personal context remain excluded. The source audit adds only public source pages and bounded assistant judgments. |

## Evidence and remaining limits

- [Stage 1 entity-type mapping](../data/stage1-entity-types.json) and [bounded source audit](../data/stage1-source-audit.md).
- [Stage 2 later observations and future-check placeholder](../stages/02-post-discovery/data/post-publication-checks.json).
- [Version metadata](../report-versions.json) and [revision history](../CHANGELOG.md).

> The exact date and capture of the user's earlier native X search remain unavailable. The independent browser attempt did not establish its search results. November 7 is not a confirmed retest. Source spot-checks do not fact-check the whole corpus, validate tool scores, or provide independent human coding. Stage 3's eight independent human-review decisions remain pending. None of these revisions demonstrates GEO effectiveness.
