?Field guide

AIVS: the spoken recommendation layer of AEO

AEO covers how brands show up in AI answers. AIVS is the spoken slice of that work: when an assistant answers aloud, it usually has room for one name. We treat that layer as measurable, not mystical, and we separate it from outdated Alexa-skills “voice SEO”.

AEO first, then the spoken layer

AEO is our house category for visibility in AI answers, on-screen and spoken. Search engines and answer engines still matter; AEO leads that work because the surface users increasingly meet first is the answer, not the blue link.

Google’s own May 2026 guidance is clear for Google generative features: AI Overviews and AI Mode are rooted in core Search ranking and quality systems, so foundational SEO still applies there. Google also notes that third-party labels such as “AEO” and “GEO” describe AI-search visibility work, while from Google’s perspective that work remains SEO for Search. We keep AEO as our operator category because buyers need language for AI answers across engines, not only Google’s surfaces.

We did not coin AEO. The term traces to the voice-search and featured-snippet era, through the Jason Barnard and Kalicube trail. Secondary sources put formalisation around 2017–2018, with date variance. We use AEO as Question.marketing’s house category for AI-answer visibility work, including the spoken layer this Field Guide calls AIVS.

What Google’s guide does not settle is how ChatGPT, Perplexity, Claude, or non-Google voice stacks choose sources. That boundary matters when someone treats a Google mythbust, for example on llms.txt, as a universal law.

What AIVS adds that generic AEO misses

On a screen, an AI answer can list several brands and a tray of citations. Through a speaker, the same question often returns one name, sometimes two. Second place on screen can still be a mention. Second place aloud is often silence.

That compression is not taste. It follows a tight word budget. Google’s Speakable documentation, in beta, recommends about 20–30 seconds of content per speakable section, roughly two to three sentences, for a workable audio experience. At a conversational text-to-speech pace, an industry convention rather than a Google figure, that is often treated as roughly a 50–75 word envelope. One complete recommendation, frame, name, reason, caveat and next step, can consume most of that envelope. A third brand rarely fits without cutting the reasoning that makes the answer useful.

So AIVS is AEO under a hard constraint: winner-take-most by construction. Optimising for longer comparison tables helps text AI and hurts spoken extractability. The asset that travels both ways is a self-contained, answer-first block of about 40–60 words that a summariser can say without “see below” crutches.

Spoken Shortlist Audit: how we measure AIVS

We run a Spoken Shortlist Audit: a fixed, repeatable protocol for spoken brand visibility. The protocol name and scorecard definitions follow MaxAEO’s published framing of 27 July 2026. We cite that source and do not claim exclusive coinage.

We use that protocol to produce movement numbers without pretending there is a public “voice Search Console”.

Operator steps

1. Build a fixed set of about 40 spoken prompts from real customer language, including support tickets, call notes and site search, split across category discovery, head-to-head, attribute-led and problem-led intents.
2. Freeze variables: device, account state, location and language. Log them. Changing any of them breaks week-to-week comparison.
3. Prefer a clean account. Personalisation from your own history flatters the scorecard.
4. Record every answer. Spoken output vanishes; memory invents favourites.
5. Transcribe verbatim, including hedges. “You might consider Brand X” is not the same as “I’d go with Brand X”.
6. Code each transcript: brands named, order named, whether we were named first, whether a source was attributed aloud, and whether the description was accurate.
7. Ask one standard follow-up, “any others?”, and code the second answer separately.
8. Run the same prompts as text on the same platform. The gap between text shortlist and spoken shortlist is the compression failure list.
9. Repeat on a fixed cadence. Report movement, not a lone absolute.

What to do, and what not to do

Do

• Standardise one compressible category claim across owned and major third-party profiles.
• Write answer-first blocks of about 40–60 words for the prompts where First-Name Rate is weakest.
• Prefer problem language customers say aloud over category jargon.
• Keep foundational SEO and non-commodity expertise for Google AI features; measure other engines separately.
• Treat voice as a downstream surface of AEO work, with Speakable markup only where Google’s documented eligibility fits.

Don’t

• Fund a parallel classic Skills or Actions programme as the main fix for 2026 spoken recommendations.
• Confuse Google’s “AEO/GEO hacks to ignore” list with every non-Google engine’s rules.
• Chase inauthentic mention farms; Google explicitly discounts that pattern for its generative features, and it fails the pub test everywhere else.
• Replace comparison content that wins text AI answers. Keep both shapes: tables for screens, one-sentence claims for speakers.
• Report a single lucky ChatGPT Voice answer as a trend.