<< All versions

Skill v1.0.0

currentAutomated scan100/100
xuanranl/loamwright-seo-skill/voice-search-optimizer
──Details
PublishedSeptember 29, 2026 at 02:06 AM
Content Hashsha256:ed9014e7b6a38536...
Git SHAcc3f19dac8fe
──Files
Files (1 file, 1.7 KB)
SKILL.md1.7 KBactive
SKILL.md · 38 lines · 1.7 KB

version: "1.0.0" name: voice-search-optimizer description: RETIRED v3.35 (2026-07-04) — voice search optimization has no tactic that answer-first content, FAQ sections, and schema do not already cover; the founding "50% of searches by voice" statistic is a debunked zombie citation. scripts/lint/voice_search_check.py survives as a MANUAL diagnostic only. disable-model-invocation: true user-invocable: false


Voice Search Optimizer — RETIRED

Status: retired 2026-07-04 (v3.35). Do not dispatch. Not an orchestrator stage.

Why retired

  • The "50% of searches will be voice by 2020" statistic that founded VSO is a

mis-attributed zombie citation; it never materialized. 2026 posts still recycling it ("27% of queries are voice") trace to the same debunked lineage.

  • Search Engine Land's own VSO guide (updated Nov 2025) lists question headings,

conversational long-tail, FAQ blocks, snippet formatting, FAQPage/HowTo schema, and page speed — every item is standard SEO this pipeline already executes (FAQ mandatory section + paa-alignment-check, citation capsules, schema-generator, humanizer). No voice-only tactic exists.

  • The genuine shift went to conversational AI (ChatGPT voice, Gemini Live), which

consumes the same answer-first content the GEO stages already optimize for.

What survives

  • scripts/lint/voice_search_check.py — retained as a manual diagnostic CLI

(never a stage; running it as a gate would block on a discipline with no evidence):

``bash python -m scripts.lint.voice_search_check --input memory/workspace/{task}/draft.md --json ``

Evidence base

references/seo/serp-feature-value-2026.md §3.

All versions