Share of Answer vs Share of Voice Across AI Engines (2026)

Share of voice (SOV) tells you how often your brand appears among a fixed peer set. Share of answer (SOA) tells you how much of the answer itself—definitions, shortlists, how-tos, and decision criteria—is sourced from pages you control or aligned third parties. In 2026 those two metrics diverge constantly across Gemini, OpenAI (ChatGPT), Claude, Grok, and DeepSeek. A brand can win mentions while losing the quotable nucleus that buyers remember.

This guide shows how to measure both without inventing vanity traffic numbers, how engine UIs change what you can observe, and how to act when SOV is healthy but SOA is weak (or the reverse).

GEOcheck.ai (ThinkPrompt Co., Ltd) scores visibility across those five engines. Perplexity is a citation target / canary, not a scored engine. Public entry: the homepage analyzer; category context: the homepage AI visibility analyzer. Sister products: Doctranslate.io and Mangaka.app. Not geocheck.cc / GeckoCheck.

Definitions that survive a weekly panel

Metric	What you count	Typical unit	Failure mode if used alone
Share of voice	Mentions (or citations) among a fixed peer set	Brand mentions ÷ peer-set mentions	Celebrates name-drops while rivals own the facts
Share of answer	Fraction of answer units attributable to your sources or verified claims	Units you own ÷ total units in the answer	Harder to score; needs a coding rubric
Citation rate	Answers (or mentions) with a URL/source chip	Citations ÷ prompts	Ignores zero-click UIs
Accuracy rate	Correct parent, product, engines, packaging	Correct ÷ branded prompts	Silent when you are absent

SOV taxonomy detail lives in the AI share-of-voice scorecard. Mention vs citation split: citation vs mention metrics.

What is an “answer unit”?

An answer unit is a discrete claim block the model presents as useful on its own:

1. Definition — “X is …”
2. Shortlist slot — a named product in a “best for” list
3. Attribute table row — engines covered, price band, deployment
4. How-to step — numbered action a buyer might copy
5. Objection handle — “GEO is not just SEO because …”
6. Source chip — explicit URL when the UI shows one

SOA asks: of the units that matter for your category, how many are yours (first-party URL cited, or a third-party page that repeats your verified facts)?

Why SOV and SOA diverge

1. Mention without nucleus — You appear in a peer list; the definition and pricing table come from a rival hub (when AI answers skip citations).
2. Citation without ownership — A review site is cited; your entity page is not (review sites as citation sources).
3. Concentration — Two domains supply most units even when many brands are named (citation concentration risk).
4. Engine UI split — Claude may narrate without chips while Perplexity shows sources; SOV looks similar, SOA observability does not (multi-engine gaps).
5. Stale winning page — You still get mentioned from memory; the retrieved how-to is a competitor’s fresher guide (citation decay).

Measurement protocol (frozen panel)

Use a versioned prompt set (GEO prompt set design). For each prompt × engine × date:

Layer A — SOV fields

• Peer set locked (e.g. Profound, Otterly, Peec, Scrunch + real sales peers)
• Mention Y/N per brand
• Optional: primary vs aside mention
• Roll up: your mentions ÷ all peer-set mentions

Layer B — SOA fields

Code each answer into units (spreadsheet or rubric). For every unit:

• unit_type (definition / shortlist / attribute / how-to / objection / other)
• owner = self_first_party | self_aligned_third_party | peer | unknown | wrong
• evidence = cited URL if any; else “narrative-only”
• accuracy = correct / incorrect / n/a

Then:

SOA = count(units where owner in {self_first_party, self_aligned_third_party} and accuracy ≠ incorrect) ÷ count(all coded units)

Report SOA per engine and per prompt class (branded, category, comparison, how-to, objection). Do not invent session percentages to “explain” SOA.

Layer C — Observability flags

• citations_visible Y/N (UI showed sources)
• timeout_or_refusal — exclude from rates or report separately
• Canary prompts on Perplexity kept out of the five-engine product score

Engine-aware reading of SOA

• Gemini — Often tied to fetchable, index-friendly pages; SSR and sitemap hygiene move SOA faster than social posts.
• OpenAI — Live fetch matters for contested claims; thin JS shells suppress first-party units.
• Claude — Strong narrative units; if chips are rare, lean on accuracy coding and branded recovery, not citation vanity.
• Grok / DeepSeek — Compete on dense public pages; vague marketing copy rarely becomes the how-to unit.
• Perplexity (canary) — Best for seeing which URLs win source slots; never fold into SOV/SOA product totals.

Diagnosis matrix

Pattern	Likely bottleneck	Next action
High SOV, low SOA	Mentions without quotable spokes	Ship answer-first hubs; fix entity pages (answer-first structure)
Low SOV, high SOA on branded only	Discovery weak; entity OK	Strengthen category/comparison clusters (topic clusters)
Citations on third parties only	Review/directory gravity	Align G2/Capterra blurbs; add first-party tables
One engine owns your SOA	Crawl or policy asymmetry	Check robots/fetcher Allow and SSR on that path
SOA falling week over week	Decay or rival refresh	Refresh cadence + re-panel (refresh cadence)

Playbook to raise SOA without gaming SOV

1. Nominate units you must own

Pick 8–12 contested units: product definition, engines scored (Gemini, OpenAI, Claude, Grok, DeepSeek—not Perplexity as scored), legal entity, packaging truth, two comparison rows, two how-to steps, one objection. Map each to a canonical URL.

2. Make units extractable

• Lead with the claim in plain HTML
• Tables over prose for attributes
• Stable headings; honest dates
• FAQPage / HowTo / TechArticle JSON-LD that matches visible text
• Internal links with descriptive anchors (internal linking)

3. Align corroborating third parties

Update directory blurbs so they do not contradict your entity page. Aligned third parties can count toward SOA; contradictory ones should count as wrong.

4. Reduce self-competition

Unpublish thin duplicates and probe posts. Soft-404 UUID noise steals crawl budget from the URLs that should supply units (soft-404 UUID pages).

5. Re-measure after infra changes

Crawl fixes change eligibility. Re-run the same panel before celebrating SOV/SOA lifts.

Checklist

• [ ] Peer set versioned; SOV formula documented
• [ ] Answer-unit rubric documented; inter-rater spot checks monthly
• [ ] SOA reported per engine and per prompt class
• [ ] Perplexity canaries separate from five-engine totals
• [ ] Each contested unit mapped to a live SSR URL
• [ ] CTAs only to homepage / homepage AI visibility analyzer
• [ ] Charts annotated when the prompt set or peer set changes

Anti-patterns

1. SOV screenshots as strategy — one lucky shortlist is not a trend.
2. Counting wrong facts as “owned units” — inaccurate SOA is brand damage.
3. Averaging citation UIs into product scores — pollutes packaging claims.
4. Changing the peer set mid-quarter — breaks SOV continuity.
5. Publishing more thin posts to “boost SOA” — units need density, not sprawl.

FAQ

Is SOA just citation rate with a new name?

No. Citation rate is binary presence of a source chip. SOA codes which parts of the answer you own, including narrative-only UIs where chips never appear.

How many units should we code per answer?

Enough to cover the decision surface—often 5–15 for category prompts. Stop when additional units are decorative.

Can we automate SOA coding?

Start manual with a rubric. Automate only after agreement rates are high; models mis-label “owned” vs “peer” constantly.

Does high SOA on how-to prompts mean we win category consideration?

Not necessarily. How-to SOA and category shortlist SOV answer different funnel questions—report both.

Next step

Lock a v1 unit rubric this week. Score last week’s panel for SOV and SOA on Gemini, OpenAI, Claude, Grok, and DeepSeek. If SOV is fine and SOA is weak, ship one denser first-party spoke for the highest-value unit class, align the matching review blurb, then re-run the same prompts.

Start at the homepage AI visibility analyzer and keep peers in view on the homepage AI visibility analyzer.