Can AI find a new experiment post?

A small observation of direct access, exact-excerpt discovery and natural discovery across ChatGPT Search and Grok.

  • A supplied URL enabled access in the selected Grok run. Grok retrieved body-specific facts that met the post-access criterion. ChatGPT explicitly said it could not access the actual post. This is one answer per environment, not a platform-wide ranking.
  • The exact excerpt did not produce a qualifying target-post result in either selected B answer. The excerpt matches the captured post character for character. Grok's actual web/X activity was recorded; ChatGPT B's executed search queries were not independently captured.
  • Neither selected natural-query answer found and cited the target X post. Finding other researchers, the author or a GitHub report is a separate outcome. ChatGPT C is a later manual replacement with qualified control evidence.

Collection summary: 6 planned cells · 10 actual submissions · 6 complete selected answers. Four earlier ChatGPT C attempts had technical or body-capture problems before a complete manual replacement was obtained. Those records remain in the dataset.

Completeness describes the selected captures; it does not mean collection was failure-free. No stable retrieval success rate or GEO optimization effect is estimated.

Six-cell comparison

A supplies the URL; B supplies an exact excerpt; C supplies neither identity, URL nor excerpt.

One selected answer per platform × condition. These are observed outcomes, not stable success/failure rates. “Not demonstrated” means the recorded answer did not meet that task's criterion. Missing body capture is never treated as a negative retrieval result.

Condition · environmentTarget outcomeSearch / collection qualificationEvidence
A · ChatGPT SearchNot demonstratedExplicitly could not access the actual post.Full answer and evidence
A · GrokDemonstratedBody-specific details met the access criterion; the added “on the same platforms” qualifier is unsupported.Full answer and evidence
B · ChatGPT SearchNot demonstratedSearch selected; actual query execution was not independently captured.Full answer and evidence
B · GrokNot demonstratedWeb/X activity was recorded, including shorter and punctuation-free variants.Full answer and evidence
C · ChatGPT SearchNot demonstratedLater user-operated C-R5 replacement. Exact question and 29-site search corroborated; model/effort and approximate time user-confirmed; pre-submit temporary/personalization controls not independently recorded.Full answer and evidence
C · GrokNot demonstratedPrivate Expert conversation with observed web/X retrieval; other experiments are not discovery of this target post.Full answer and evidence

ChatGPT B: the native Search chip was selected and the answer describes searches; its actual query execution was not independently captured. ChatGPT C: the complete answer is the later user-operated C-R5 replacement, not an equivalently preflight-captured automated run. Exact prompt and actual search were corroborated, while model/effort and approximate time are user-confirmed and session-isolation controls were not independently recorded.

Methods in brief

  1. Use fresh visible temporary/private conversations and the unchanged A/B/C questions. The target is the original GEO Lab #01 X post.
  2. Evaluate post-body access, exact discovery and unprompted discovery separately. Merely returning the supplied URL, repeating the quote or reading linked GitHub content is insufficient for post access.
  3. Preserve all ten actual submissions and select six complete records. Recovery of an existing message is not another query. All four earlier C technical/partial captures remain visible.

Exact text: the registered B quote is an 86-character continuous fragment from the latter part of a sentence, including the U+2014 em dash and final period. Both B prompts preserve it unchanged. No transcription difference was observed. An exact prompt does not establish the literal query a platform sent to a search service.

Read the original prompts, full scoring rules and controls.

Separate follow-up checks

Native X search · later user report

The user reports finding the original post with these account-scoped queries, using Latest and All users:

  • from:bird_helen67403 "2 of 20"
  • from:bird_helen67403 "GEO Lab"

Evidence level: user-reported, not independently verified. Confirmation recorded: 2026-11-02. Actual search date: Not supplied. No screenshot or retained result list was supplied. The confirmation date is not substituted for the unknown observation date.

These queries disclose the author. They differ from the unbranded AI tasks and do not establish native-X availability on October 31 or explain the original AI outcomes.

Independent access check

On 2026-11-02, a browser recheck timed out and displayed only a Loading placeholder. No result list was observed. This is unresolved access, not a negative native-X search result.

The original six selected A/B/C outcomes remain unchanged. Read the separate follow-up evidence and limitations.

Later B/C retest

The external review proposes a separate November 7, 2026 B/C retest. No completed retest results are included in this revision, and no new test or automation is initiated by the report. Any later batch needs its own timestamps, prompts, conditions and attempt ledger; the October 31 outcomes remain unchanged.

What this cannot show

  • Six selected answers cannot estimate reliable discovery rates, prove indexing failure or show that the post is absent from X.
  • The post was new, and indexing delay is a possible explanation rather than an established cause. A later native-X result would concern another search surface and time.
  • Visible account controls do not audit backend isolation. ChatGPT's manual C replacement has different control evidence; Grok's backend version is not disclosed.
  • This is a small exploratory observation, not evidence that a GEO tactic worked or that either platform is inherently better.

Technical appendix and data

Read the complete technical appendix for all prompts, answers, source roles, supporting C/M/R observation and failure history. All appendix sections remain expanded by default.

Download the redacted public data · Publication scope · Stage 2 data guide

Private conversations, public-share identifiers, screenshots, account UI, local paths and original diagnostics are excluded. Verbatim answer content remains evidence rather than an endorsement; unrelated private passages are marked where omitted. This presentation revision does not alter the original six judgments.

Author: Helen Bird · Experiment date: 2026-10-31 · First published: 2026-11-01 · Last revised: 2026-11-06 · Version: v1.4

Technical maintenance: consistent canonical URLs, page-specific descriptions and social metadata, truthful author/series markup, and links to individual original-answer pages. Original answers, coding, data downloads and research outcomes are unchanged. This bundled update has no measured AI visibility effect. November 6 navigation update: linked the new technical learning record; original experiment data and findings remain unchanged.

Version metadata · Revision history · Cross-report review decisions · Stage 4 review decisions · Technical maintenance