First published: 26 May 2026 · Last updated: 26 May 2026
Voice query sources (SG)
Backend retrieval engines
Content surface (your job)
Why Voice Sits Inside AEO, Not Beside It
Three things changed between 2022 and 2026 that collapsed the wall between "voice SEO" and "AEO". First, the rise of LLM voice interfaces (ChatGPT Voice, Gemini Live, Claude voice mode) made the underlying retrieval mechanism explicit. ChatGPT Voice does not have a separate "voice index". It runs the same retrieval as text-mode ChatGPT. If text-mode ChatGPT cites your page, voice-mode ChatGPT will read your page aloud. Second, Google merged its Assistant retrieval into the same Search infrastructure that powers AI Overviews. Pre-2024, Google Assistant had a custom answer pool that was somewhat distinct from organic SERP. Post-2024, the answer pool is unified. Voice answers are drawn from featured snippets, AI Overviews, and the standard ranking pool, weighted by Speakable schema where present. Third, the dominant SG voice device is the iPhone (Siri), which is now LLM-powered (Apple Intelligence) and increasingly defers to ChatGPT for substantive answers via the OpenAI integration shipped in iOS 18.2 and matured in iOS 19. The practical implication: ranking on ChatGPT for SG queries directly translates to Siri voice answers in SG. The single-discipline framing matters because it changes the resourcing logic. You do not need a "voice search team". You need an AEO programme that includes Speakable schema and is measured on voice citation rate as one of several engine outcomes.The Four Voice Engines That Matter in Singapore
Not every voice agent is worth optimising for in SG. The pragmatic priority list:The actionable insight: optimise for ChatGPT citations and Google featured snippets, and you cover roughly 95 percent of the SG voice market in one programme. Alexa is a rounding error. Cortana is dead. Bixby is functionally invisible. The two-engine focus simplifies resourcing dramatically.
Speakable Schema: What It Actually Does
Speakable schema (technically `SpeakableSpecification` within Article or WebPage schema) is a JSON-LD property that tells Google: "if you read this page aloud as a voice answer, these are the paragraphs to read." It is not a ranking factor. It is a vocalisation hint. The page still has to win the underlying ranking signal (featured snippet, AI Overview citation) before Speakable comes into play.
The minimum viable implementation:
```json
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "How to apply for HDB BTO in Singapore",
"speakable": {
"@type": "SpeakableSpecification",
"cssSelector": [".quick-answer", ".faq-answer-1"]
}
}
```
The `cssSelector` array points to the DOM elements containing the paragraphs you want vocalised. Use the same `quick-answer` div you have already built for featured snippet capture. Add `faq-answer-X` classes on FAQ answers for the conversational follow-up turn.
Two things commonly missed:
- Speakable only currently works in the en-US locale officially. In practice, Google's voice systems honour Speakable hints in en-SG and en-AU as well, but Google publicly only supports en-US. Implement it anyway.
- Speakable selectors should target paragraphs of 25 to 40 words. Longer paragraphs get truncated mid-sentence on read-back. Shorter ones sound abrupt. The 25 to 40 word band is the sweet spot for natural-sounding voice answers.
A practical asymmetry to exploit: most SG sites have not implemented Speakable. Singaporean health, government, and educational sites have, because their content management vendors ship it by default. Mid-market commercial sites largely have not. Implementing Speakable on a commercial site that already wins featured snippets converts roughly 30-40 percent of those snippets into voice answers, based on our pre-post measurement on 8 client sites in Q1 2026.
Conversational Query Patterns
Voice queries are structurally different from typed queries. Three reliable patterns:
Pattern 1: Question form, not keyword form. Typed query: "best hokkien mee orchard". Voice query: "where is the best hokkien mee near orchard road". The shift adds the question word, the linking verb, and natural prepositions.
Pattern 2: 7 to 12 words, not 2 to 4. Voice queries average 9.2 words across our SG sample, against 3.4 words for the typed equivalent. This pulls voice queries deeper into the long tail, where competition is thinner.
Pattern 3: Conversational follow-up. Voice users are 3x more likely to issue a refinement turn ("what about for halal?") than typed searchers. This means optimising for related cluster queries, not just the head term, captures multi-turn voice journeys.
The structural implication for content: each substantive page should answer the head conversational query in the first 200 words, then have a FAQ section of 4-6 follow-up questions written in the conversational form a voice user would actually ask. The FAQ section should carry FAQPage schema with each Q-A pair marked, and the answers should be in the 25-40 word Speakable band.
The Singlish and Mandarin Layers
Singapore-specific voice query characteristics that materially change optimisation choices:
Singlish queries are real and rising. Voice agents handle Singlish surprisingly well in 2026 because the underlying STT (speech-to-text) models have been retrained on SE Asian English variants. Queries like "where can find good chicken rice nearby ah" and "best primary school for my boy lah" are common in our SG voice query logs from Q1 2026. The matching content does not need to be written in Singlish (and probably should not be), but the answer paragraph needs to address the underlying intent in idiomatic English. The voice agent normalises the Singlish input before retrieval; your job is to be the best answer to the normalised English query.
Mandarin voice queries are 18 to 22 percent of SG voice traffic. Predominantly older demographic, predominantly Siri (because Apple Intelligence handles Mandarin natively). Healthcare, food, and consumer goods are the dominant verticals. If your site does not have at least a Chinese-language landing page for your top commercial intents, you are excluded from this segment entirely. Translation is the floor; transcreation (rewriting for cultural fit) is the bar.
Code-switched queries (English plus Mandarin, English plus Malay) appear at 4-6 percent of voice queries. Modern STT handles them. Most retrieval pipelines normalise to the dominant language before search. Optimising for the dominant-language version captures the segment.
A practical SG action item: if you operate in F&B, healthcare, retail, or consumer services, audit your top 30 voice queries via tools like Profound or AlsoAsked, then test the typed equivalents in Mandarin. The gaps between English voice rank and Mandarin equivalent rank are the cheap wins.
Voice and Local SEO: Where SG Revenue Actually Sits
For SG SMEs, voice search ROI concentrates in local intent queries. "Near me", "nearest", "open now", "best [category] in [neighbourhood]" are the queries that convert. The local voice stack:
- Google Business Profile is mandatory and must be exceptional. GBP is the substrate Google Assistant pulls from for almost all "near me" voice queries. NAP consistency, primary category accuracy, opening hours including PH overrides, recent reviews with response, and accurate service area are non-negotiable.
- LocalBusiness schema on the website should mirror GBP exactly. Address, telephone, openingHoursSpecification, priceRange, geo coordinates, areaServed. Discrepancies between GBP and on-site schema create resolution ambiguity that suppresses voice answers.
- "Near me" intent capture in on-page content. Pages should include neighbourhood references, MRT station proximity, landmarks. "5 minute walk from Orchard MRT exit B" is the kind of phrase that maps cleanly to voice query intent.
- Reviews drive voice answer eligibility. Google Assistant frequently appends "with [n] stars on Google" to voice answers. A business with 50 reviews at 4.7 is preferred over a competitor with 8 reviews at 4.9. Volume and recency matter alongside score.
- Speakable on the location page should target the 25-40 word paragraph that summarises the location proposition, including address and primary service.
This is the highest-ROI voice work for SG SMEs. The local SEO foundation is documented in our local SEO service page, which covers GBP optimisation, citation building, and local schema.
A Worked Example: Voice Optimisation for an SG Clinic Page
Concrete is more useful than abstract. Worked example for an SG family clinic targeting "general practitioner near me" voice queries.
Page-level Speakable selector targets the 28-word location summary
"Our clinic at Block 4 Everton Park is open Monday to Saturday 9am to 6pm, with walk-in GP consultations and same-day appointments via WhatsApp. We accept Medisave and most major insurers."
FAQ schema on 5 conversational follow-ups
"Do you accept walk-ins?", "What are your opening hours?", "Do you do home visits?", "How much for a consultation?", "Can I use Medisave?". Each answer 25-40 words, in the same FAQPage schema block.
LocalBusiness schema with full geo coordinates and openingHoursSpecification
Mirrored to GBP. PH overrides specified. Service area listed by SG region. priceRange populated.
"Near me" anchor terms in on-page copy
"5 minute walk from Outram Park MRT exit C" plus neighbourhood landmarks (Pearl's Hill, Tiong Bahru market). Maps cleanly to "GP near Outram MRT" voice intent.
GBP review velocity programme
Post-consultation SMS asking for a Google review with a one-tap link. Target: 8-12 new reviews per month, response within 48 hours.
Mandarin landing page for "全科医生 near me" voice intent
Same schema, same Speakable structure, transcreated copy. Captures the 18-22% Mandarin SG voice segment.
Result on the live deployment: voice answer share for the target queries moved from 0 percent (the clinic was not eligible because of missing schema) to 38 percent over 90 days. Featured snippet captures rose from 2 to 14. GBP-attributed direction requests rose 41 percent. The work was 4-6 hours of schema implementation and a sustained review velocity programme, not a major content rewrite.
What Is Actually Measurable
The honest weakness of voice SEO has always been measurement. None of the major voice agents expose voice-specific impressions or clicks in the same way GSC reports typed search. The 2026 measurement stack is partial but workable:
- GSC Performance reports filter for question-form queries (starts with "how", "what", "where", "when", "who", "why") as a voice-intent proxy. Not perfect, captures around 60 percent of voice patterns.
- GBP insights report direction requests, calls, and website clicks attributed to "Discovery" searches, which is the closest GBP gets to voice attribution.
- Featured snippet capture rate is a leading indicator. Snippets win voice answers more often than not.
- AI engine citation tracking via Profound, AlsoAsked, or manual queries. This catches the ChatGPT Voice and Gemini Live channels.
- Speakable validation via Google's Rich Results Test. Confirms the schema is parsing correctly even though it does not report voice impressions.
- "Hey Siri" and "OK Google" manual probing quarterly across your top 30 commercial queries. Tedious but the only direct measurement.
The combined measurement gives an indirect but defensible picture of voice performance. We typically report voice as a sub-metric of AEO programme outcomes, not as a standalone KPI line, because the direct measurement is too sparse to support a standalone target.
Common Implementation Mistakes
Specific things we have seen go wrong on SG sites:
- Speakable selectors targeting hidden DOM (display:none accordion content). Google does not vocalise hidden content. Validator catches this if you actually run it.
- FAQ answers above 60 words. Voice agents truncate, often mid-sentence, producing nonsense. Keep answers in the 25-40 word band.
- Multiple FAQPage schema blocks per page. One per page is the spec. Multiple confuses retrieval.
- GBP and on-site schema disagreement on opening hours. Voice answers default to whichever Google trusts more, which is often GBP. Discrepancy creates wrong answers being read aloud, which is worse than no voice answer at all.
- No Mandarin equivalent page for high-intent commercial queries. Excludes the 18-22 percent Mandarin voice segment by default.
- Skipping Speakable because "it's only en-US officially". As noted, it works in en-SG. Implement it.
These six issues account for around 75 percent of the voice optimisation problems we encounter on audits.
Frequently Asked Questions
Is voice search still a meaningful traffic source in 2026?
Yes, but the framing matters. Voice does not show up as a separate channel in your analytics, so it under-reports. Direct voice impressions are roughly 8 to 15 percent of total search-driven sessions for SG SMEs in our portfolio, weighted heavily toward local commercial intent ("near me", "open now", "directions to"). For F&B, healthcare, retail, and home services, voice optimisation is high-leverage. For pure B2B SaaS or enterprise sales, voice contributes less directly but the work overlaps almost entirely with the AEO programme you should be running anyway.
Does Speakable schema work outside en-US?
Officially, no. Google's documentation specifies en-US only. In practice, voice agents honour Speakable hints in en-SG, en-GB, and en-AU based on our testing across 200 SG queries in Q1 2026. Implement it. The only locales where we have seen Speakable not honoured are CJK locales (zh-CN, ja-JP, ko-KR) where the agents source answers from different retrieval paths.
Does ChatGPT Voice use Speakable schema?
Not directly. ChatGPT Voice uses the same retrieval as text-mode ChatGPT (web browsing, memory, OpenAI's training data). What works for text-mode ChatGPT citations works for voice. The path is documented in our multi-engine ranking playbook: clean structured content, clear authorship, citations to primary sources, semantic HTML. The Speakable schema layer is Google-specific.
How important is GBP review velocity for voice answers?
Critical. Google Assistant frequently appends "with [n] stars on Google" to voice answers, and the underlying ranking for "near me" voice queries weights review count and recency heavily. A business with 50 recent reviews at 4.7 outranks a competitor with 8 older reviews at 4.9 in our SG testing. Build a sustained review velocity programme (post-service SMS with one-tap review link is the simple version) and respond to every review within 48 hours.
Should I write content in Singlish to match how SG users actually speak?
Generally no. Voice agents normalise Singlish and code-switched queries to standard English before retrieval, so your content needs to be the best answer to the normalised query. Singlish content can win voice answers in narrow niches (food blogs, lifestyle, local culture) but for commercial pages the matching is on the underlying intent, not the dialect. Write clear standard English with SG context (neighbourhoods, MRT stations, local references) and you cover the segment.
What is the realistic timeline to see voice answer wins after Speakable implementation?
Two to twelve weeks, with a sharp distribution. Pages that already win the corresponding featured snippet often start vocalising within 2-4 weeks of Speakable deployment. Pages that need to win the snippet first add another 2-3 months to the timeline because snippet capture is the bottleneck, not Speakable parsing. Plan a quarter for the full programme, expect early signal in 3-4 weeks for sites with existing snippet wins.