GEO Lab Interim Review · Part 2
Question framing methods and evidence
Methods and collection
Collection ran from November 10 at 23:23 to November 11 at 03:13, 2026, Beijing time. All 120 submitted questions produced analyzable bodies. Grok's minimum observed submission interval was 181.059 seconds. Twenty-four same-prompt cross-environment pairs reversed the planned lead. Block matching uses the planned block and actual timestamps.
Codex semantically coded the full bodies by frame, with a review of evidence excerpts and focal cases. The corpus has one primary coding pass per frame. A name in substantive answer text counts as a mention. Direct personal recommendation requires an affirmative selection of that person. Author naming and explicit, verified attribution are recorded separately. First-party pages verify 432 of 897 recommended resource instances. The other instances have unknown personal attribution.
Binary intervals use the Wilson 95% formula with answers as units. Resource-instance fractions and block contrasts are descriptive. Jaccard overlap is intersection size divided by union size. Each same-prompt comparison uses the 45 pairs of its ten answers. Each different-prompt comparison uses ten same-block pairs. Link comparisons require observable sets on both sides.
One Grok evidence-resource snapshot missed its final word. The expanded capture from that session supplied an exact matching body and the trailing word. The public row includes the recovered text and marks the repair. The answer interpreted GEO in several senses. Only generative-engine optimization resource recommendations enter the primary count.
Per-condition results
Each row switches from a resource prompt to a creator prompt. Brackets show Wilson intervals.
| Environment | Frame | Any name | Any personal recommendation | Mean resources |
|---|---|---|---|---|
| ChatGPT web | Discovery | 6/10 [31%–83%] → 10/10 [72%–100%] | 0/10 → 10/10 | 9.0 → 8.6 |
| ChatGPT web | Explicit recommendation | 6/10 [31%–83%] → 10/10 [72%–100%] | 0/10 → 10/10 | 7.9 → 7.5 |
| ChatGPT web | Original experiments | 9/10 [60%–98%] → 10/10 [72%–100%] | 0/10 → 10/10 | 8.6 → 8.2 |
| Grok Expert | Discovery | 10/10 [72%–100%] → 10/10 [72%–100%] | 0/10 → 10/10 | 8.8 → 6.6 |
| Grok Expert | Explicit recommendation | 10/10 [72%–100%] → 10/10 [72%–100%] | 0/10 → 10/10 | 8.8 → 5.8 |
| Grok Expert | Original experiments | 9/10 [60%–98%] → 10/10 [72%–100%] | 0/10 → 10/10 | 5.4 → 4.5 |
| Environment | Frame | Any recommended resource author named | Shared verified resource instances with author named |
|---|---|---|---|
| ChatGPT web | Discovery | 6/10 [31%–83%] → 10/10 [72%–100%] | 4/9 → 7/9 |
| ChatGPT web | Explicit recommendation | 6/10 [31%–83%] → 10/10 [72%–100%] | 2/5 → 3/5 |
| ChatGPT web | Original experiments | 9/10 [60%–98%] → 10/10 [72%–100%] | 7/12 → 12/12 |
| Grok Expert | Discovery | 9/10 [60%–98%] → 10/10 [72%–100%] | 11/18 → 18/18 |
| Grok Expert | Explicit recommendation | 10/10 [72%–100%] → 10/10 [72%–100%] | 8/12 → 11/12 |
| Grok Expert | Original experiments | 9/10 [60%–98%] → 10/10 [72%–100%] | 7/16 → 16/16 |
Resource lists across frames
These are median overlaps within each platform and target, paired by block.
| Environment | Target | Discovery vs. recommendation | Discovery vs. evidence | Recommendation vs. evidence |
|---|---|---|---|---|
| ChatGPT web | Resource | 14.3% | 6.1% | 6.7% |
| ChatGPT web | Creator | 10.5% | 6.5% | 7.4% |
| Grok Expert | Resource | 27.6% | 10.0% | 8.3% |
| Grok Expert | Creator | 9.5% | 0.0% | 0.0% |
Search and source evidence
Observable answer URLs appeared in 60/60 ChatGPT and 4/60 Grok answers. The conservative identifiable source-URL subset was 47/60 and 1/60. ChatGPT grouped citations were not expanded. Grok exposed a source panel in all 60 runs, with a median of 72 normalized URLs. Panel candidates and page visits have separate fields. Only one matched cross-platform answer pair had identifiable source URLs on both sides.
To inspect the open-ended search path, locate IR2-GROK-WEB-EVIDENCE-CREATOR-B03-A01 in the answer file. Compare the generic search terms, site visit and follow-up query with the displayed answer. IR2-CHATGPT-WEB-EVIDENCE-CREATOR-B08-A01 includes both the hub and GitHub as body links.
Data and reproduction
- Original prompts, settings and collection summary
- 120 redacted answer bodies, body links and visible search evidence
- 120 attempt histories
- Normalized coding, counts and supporting excerpts
- Entity aliases and first-party authorship checks
- Recomputed metrics and block comparisons · Download data archive
Rebuild offline with python3 scripts/build_prompt_framing.py from the repository root. Cite a GitHub commit permalink and the relevant answer ID for a fixed record.
The public export removes private links, credentials, local paths, opaque identifiers, tracking parameters and unrelated same-name search hits. It retains aggregate source-domain sets used by the calculations and omits source-panel previews. The curated URL list can be shorter than the original visible source panel.