Open GEO Lab
Four platforms.
Two separate observations.
The September 17 baseline remains separate from the completed October 16–21 retest of 40/40 answers. Mentions, citations and accurate identification are counted separately; different conditions do not establish effectiveness. Updated 2026-10-22.
Experiment method
40 historical answers, including successes and failures.
Test date: 2026-09-17. Four platforms received five fixed English and five Chinese prompts in separate conversations, without pasted website content. Copilot could not connect and is excluded from the denominator.
Retest status updated on 2026-10-22. Post-update retesting started on 2026-10-16 and reached 40/40 on October 21. Cross-date sanitized observations appear first; the separate September 17 baseline and download remain below. No optimization effect is claimed.
Cross-date retest: October 16–21, 2026
40/40 complete. Unprompted discovery remains a gap.
Claude and Perplexity each contributed ten answers on October 16; ChatGPT ten on October 17; Gemini ten on October 21. Each fixed bilingual prompt used a separate new conversation. Historical answers are not relabeled as today's samples.
| Measure | September 17 baseline | October 16–21 retest |
|---|---|---|
| Valid answers | 40 | 40 |
| Brand mentions | 26/40 | 24/40 |
| Official sources (coding differs) | Original citation coding: 11/40 | Body-linked: 14/40; two additional tool traces |
| Accurate descriptions (coding not identical) | 11 | Strictly accurate: 10; partially accurate: 7 |
| Unbranded appearances | 2/16 | 0/16 |
The retest also contains seven unidentified-brand responses and sixteen without a brand description; strict accuracy among mentions is 10/24. Body-linked sources include official URLs verified by opening citation buttons attached to the answer. Separate tool traces are not double-counted; unopened sources remain unassessed, not zero.
Different conditions do not establish causation
Models, login states, retrieval, sampling dates and deployed versions differ. Gemini used an existing Google login in Auto Gemini 3 mode; ChatGPT was signed out. Unobservable underlying models, regions and personalization remain unknown. Prompted mentions are not unprompted recommendations. An unverified gaming interpretation and guarantees of recognition or removal of hallucinations were coded partially accurate.
These are descriptive observations, not proof of improvement, deterioration or causation from GEO work. No new inquiry or sale was verified. Owned brand profiles are not independent endorsements. No crawling, indexing, ranking, citation or acquisition result is guaranteed. Private conversations and screenshots remain local; the download contains sanitized aggregate counts and conditions, not full answers.
Download sanitized retest counts and conditionsResults
Read the denominators before the percentages.
| Platform | Answers | Name appears | Site cited | Accurate | Confused |
|---|---|---|---|---|---|
| ChatGPT | 10 | 8 | 8 | 8 | 0 |
| Perplexity | 10 | 6 | 2 | 2 | 4 |
| Claude | 10 | 6 | 1 | 1 | 1 |
| Gemini | 10 | 6 | 0 | 0 | 3 |
Names appear in 26/40 answers (65.0%), including wrong identities and prompts that already name the brand. Confusion occurs in 8/26 mentions (30.8%). Both unbranded appearances are ChatGPT provider queries, one in each language.
Grouped by prompt language
| Language | Answers | Name appears | Site cited | Accurate | Confused |
|---|---|---|---|---|---|
| English | 20 | 13 | 5 | 5 | 4 |
| Chinese | 20 | 13 | 6 | 6 | 4 |
Each group contains 20 answers across four platforms. Search conditions and models differ; this is not a controlled language experiment or evidence that one language is easier to index.
Download public prompt-level recordsBrand-authored method archive
Public experiment repository on GitHub — Maintained by the brand, with the sanitized baseline, brand assets and four method articles. The third is the same article as the Substack version and is counted once. This is one external channel, not three platforms or independent third-party endorsement. Publication does not establish search indexation or AI citations.
24 brand mentions do not establish unprompted AI recommendations (Chinese) — The fourth Chinese method article, published in the existing GitHub repository on 2026-10-23, separates 24/40 mentions, 14/40 body-linked official sources and 0/16 unbranded appearances. It retains model, login, cross-date and coding limitations. Brand-authored with AI assistance, not a customer case, independent endorsement or another test round.
Sunpersona brand profile on GitHub — Published with the brand name, logo, service description, official website and business contact email. The organization profile and experiment repository are still one GitHub platform.
Five evidence layers for a new-site GEO experiment — A full Chinese method article published on Substack on 2026-10-09, separating access, indexation, recognition, citations and inquiries, with AI-assistance disclosure and links to source records.
As of 2026-10-24, four distinct external articles appear across two platforms, GitHub and Substack; the fourth is on GitHub only. All are brand-maintained, not independent reporting or evidence of new AI outcomes.
What the samples show
Recognition and platform behavior vary.
- ChatGPT has eight citations in this web-enabled sample, including two provider queries.
- Perplexity correctly recognizes the two domain-led prompts; other branded prompts show confusion.
- Claude reads and cites the site correctly only for the Chinese domain-led prompt.
- Gemini does not correctly identify the brand or domain in this sample and has no site citations.
Conditions and limitations
- Models, login states and search behavior differ; some platforms decide automatically whether to retrieve.
- ChatGPT uses a signed-in account; Perplexity includes signed-out and signed-in samples. Prompt-level conditions are retained.
- Unconfirmed regions are unknown; missing evidence is not converted into success.
- The 40 answers are a convenience sample, not market coverage for every user.
- Private conversation links and personal account details are excluded. Public records retain prompts, conditions, classifications and site citations.
- When checked on 2026-10-21, the GSC report still labeled last updated 2026-10-12 showed 22 indexed pages; robots validation for the legacy /password example was still started. This is not fresh October 22 indexing data, evidence of AI citations or proof that the current release has been recrawled or reindexed.
Changes published on 2026-10-09
Published changes do not establish effectiveness.
- Unified the brand definition across the homepage, services, About and structured data; added service inputs, deliverables and acceptance criteria.
- Published four new bilingual guide topics, updated the original GEO guide and connected language equivalents and internal links.
- Published 40 sanitized historical records, separating the test date, content publication and subsequent retests.
- Replaced the self-reported checklist score with an item count, disclosing that it does not fetch websites or query AI.
- Corrected language-default links and missing-page responses; checked production pages and mobile layouts.
The actual prompts
Use the same prompts for continued retesting.
- What is Sunpersona?
- What services does Sunpersona provide?
- Is sunpersona.com a GEO optimization service?
- Who provides GEO optimization services?
- How can a brand become visible in AI search?
Start with a readiness self-check