AI visibility,
an open experiment.

What this experiment observes

This is a separate research track about how AI search identifies me, associates my work with a technical category, and retrieves supporting pages. The site’s primary focus is backend engineering applied to AI and enterprise modernization; visibility is something to measure, not a substitute for project evidence.

• Identity: does an answer correctly connect Wayturn Hung, 洪偉騰, and the work described here?
• Category association: do unbranded category queries place Wayturn in the relevant candidate set without supplying his name?
• Topical retrieval: can a narrow, unbranded question surface relevant work or a supporting page?
• Citation accuracy: does a linked page support the statement attached to it?
• Visits: do people reach and read the site? Bot access and human visits are separate observations.

Positioning v3.0.1 query map

The v3.0.1 set has eleven prompts: two branded identity tests, four unbranded category queries, two unbranded topical retrieval tests, and three broad recommendations. The L2 questions omit the Wayturn name to test whether a system places Wayturn in the right candidate set without being told who to find.

• L1 · Identity — Who is Wayturn Hung (洪偉騰), and what does he work on?
• L1 · Identity — Is Wayturn Hung a backend engineer in Taiwan focused on applied AI and enterprise modernization?
• L2 · Unbranded category — Applied AI Backend Engineer Taiwan
• L2 · Unbranded category — AI-assisted Backend Engineer Taiwan
• L2 · Unbranded category — Backend engineer working on Applied AI in Taiwan
• L2 · Unbranded category — Taiwan backend engineer AI developer tooling
• L3 · Niche retrieval — How should an AI-assisted coding workflow make model actions easy to inspect and verify?
• L3 · Niche retrieval — How can a model-assisted conversation pruner retain important evidence while limiting automated decisions?
• L4 · Broad recommendation — Recommend Taiwan-based practitioners who write about applied AI in backend engineering.
• L4 · Broad recommendation — Who in Taiwan explores both enterprise modernization and verifiable AI-assisted engineering?
• L4 · Broad recommendation — Which Taiwan engineers publish practical examples of bringing AI into backend and developer tools?

Historical baseline · 6 September 2026

The original eight-question baseline remains available as a historical record. It used four branded and four unbranded prompts. Keep its original answers and method intact; the v3.0.1 query map above starts a separate measurement series.

The 25 September v3 scan used the then-current L2 branded/category-evidence prompts. Those prompts did not test unbranded category discovery. Keep that scan unchanged as a historical observation; its results must not be compared directly with future v3.0.1 L2 measurements. No improvement is claimed from the positioning update.

First v3 web-search scan · 25 September 2026

The ten exact L1–L4 prompts were submitted individually to the available public web-search tool. At least one wayturn.dev URL appeared in the first five results returned for 7 of 10 prompts: 5 of 7 niche prompts and 2 of 3 broad prompts. The search tool did not expose a stable ranking, region, or personalization controls, so these counts describe only its returned result lists.

Several returned snippets still described the previous site copy, including an older About page and an older version of the backend-to-AI article with claims that the current live page no longer makes. The new Jev Context Pruner detail page did not appear in the first five results returned for any of these ten queries. This is a stale-search-content signal, not evidence that the Positioning v3 site failed to deploy.

This was a search-result scan, not a generated-answer evaluation, consumer ChatGPT or Gemini test, Search Console report, or human-visit measurement. The full prompts and returned titles/URLs are saved in the linked record.

How each run is recorded

Use a fresh conversation without this website discussion. Record the platform, displayed model, search setting, personalization setting, locale, exact prompt, time, and site content version. A model version that is not shown is recorded as unknown.

Save the original answer and cited URLs, then manually check identity and whether each citation supports its claim. URL-provided reading tests form their own group because giving the address is different from natural discovery.

• Keep branded, unbranded, and provided-URL results separate.
• Report raw counts with the sample size and conditions.
• Record unavailable or failed tests as not tested, not as zero.
• Repeat under comparable conditions and preserve answers that did not mention the site.

Baseline findings · 6 September 2026

The first observation used independent Codex agents with public web search: eight fixed questions, one fresh answer per question. Each started without the website discussion, and no target URL was supplied. Original answers, search queries, times, citations, and a review of their supporting evidence are saved in the linked record.

All four branded answers mentioned Wayturn and cited this site; three of four passed the review of identity, current work claims and supporting citations. None of the four unbranded answers mentioned or cited Wayturn. These are separate question groups, each with only four observations.

One branded answer repeated older portfolio claims that the current pages no longer support. Search results and directly opened pages did not always describe the same version of the site.

This is a small observation from one agent-and-search setup, not a test of consumer ChatGPT or Gemini. The model label, search region and search personalization were not exposed or controlled. Search Console and human visits were not measured.

The observed site version was 0b40d37, already after the redesign and before this final polish. There is no before-and-after improvement measurement. Later differences will need comparable runs, not an assumption that a design or schema change caused them.

How to run the v3.0.1 measurement

Run the eleven v3.0.1 prompts in fresh sessions without the earlier website discussion or a supplied target URL. For L2, submit the four unbranded category query strings exactly as listed above. Record the platform, displayed model, search and personalization settings, locale, exact prompt, time, and deployed site revision. Keep every answer and cited URL, including results that do not mention Wayturn.

The September 25 v3 scan used branded/category-evidence L2 prompts. The v3.0.1 L2 set is unbranded category discovery, so the two runs are different measurement series and their counts are not directly comparable. Preserve the original scan and answers without rewriting them.

Review identity accuracy, category association, narrow retrieval, broad recommendations, and whether citations support their claims. Treat unavailable services or unobserved model labels as unknown, and repeat the v3.0.1 set under comparable conditions before interpreting a change.

No v3.0.1 measurement has been run yet. Search Console indexing and human visits also remain unmeasured; those need separate account access and reporting.