The Share of AI Voice Study: Methodology for Measuring AI Citation Rates for Indian Local Businesses

Study Purpose

Share of AI Voice is an emerging measurement concept in the AI search era. This study establishes:

• Baseline SAV rates for Indian businesses across 8 industries
• Engine-specific SAV patterns (which engines cite which business types most)
• Prompt-type SAV variation (recommendation vs informational vs service-specific)
• SAV correlation with AIO Readiness Score (validating the readiness score as a leading indicator)

Status: this is a research protocol, not a completed study. Nothing below is a result. Angryturtle's own Share of AI Voice tracker is built and ready to run measurement for individual businesses today; the aggregate India-wide study described here is a separate, larger undertaking still in the design and sampling phase.

Why measuring SAV matters to a real decision

"Share of voice" has existed in advertising for decades as a way to answer one question: out of all the attention available in a category, how much of it is mine versus my competitors'. Share of AI Voice asks the same question for a newer channel — when someone asks ChatGPT, Perplexity, or Google's AI Overviews a question that could be answered by naming a business in your category and city, how often does the answer name you instead of the three clinics or restaurants down the road.

That number changes a real decision because it's the first metric that tells a business owner whether AI search is a channel worth investing in specifically, as opposed to something that will just happen if the website is fine and the reviews are decent. A business with a 40% SAV in its category cell is already capturing real share of a channel that's growing; a business with 0% SAV for six straight months isn't losing a battle it's fighting, it's absent from a battle entirely. Those are different problems requiring different responses, and neither is visible from Google Search Console or a rank tracker — both of which only measure the traditional search channel.

Why this data doesn't exist yet for Indian businesses

Measuring SAV credibly is more expensive than it sounds, for a reason that's easy to underestimate: it requires running the same prompt across multiple AI engines, multiple times, over multiple months, and manually judging each response for whether the business was actually named and how. There's no API that returns "was my business cited in ChatGPT's answer to this query" — someone has to ask the question and read the answer.

That cost is one reason SAV measurement for Indian businesses specifically hasn't been published anywhere yet. A second, India-specific reason is that a defensible prompt set has to reflect how Indian users actually phrase local queries, including vernacular variants and neighbourhood-level geography rather than city-level geography — "best dermatologist near Indiranagar 100 feet road" behaves differently in AI systems than "best dermatologist in Bengaluru," and a prompt set built by copying US SAV methodology would miss that entirely. Building a prompt set that's genuinely representative of Indian query behaviour, rather than a translated version of a US one, is itself a research task that has to happen before measurement can start.

Study Design

Sample: The SAV study covers a stratified sample of Indian local businesses across 8 industries and 5 cities (Mumbai, Bengaluru, Delhi NCR, Hyderabad, Pune). Within each industry × city cell, 10 businesses are selected: 5 with high AIO Readiness Scores (>70/100) and 5 with low AIO Readiness Scores (<40/100). This design enables analysis of the relationship between readiness and SAV.

Prompt Set Construction

Principles for prompt set design:

Coverage principle: The prompt set must cover the full range of query types that actual users ask about a business category — not just the most common queries. SAV on recommendation queries alone would miss informational and service-specific query exposure.

Natural language principle: Prompts use natural user language, not keyword-optimised language. "Best dermatologist in Koramangala" not "top-rated dermatologist Koramangala review."

Geographic specificity principle: Prompts include specific geographies (neighbourhood, landmark references) that reflect how Indian users actually query — not just city-level.

Vernacular inclusion: Each category's prompt set includes 3–5 Hindi-language prompts to capture vernacular AI search citation rates. See the Vernacular Search & AI Citations in India study for the dedicated deep-dive on this dimension.

Standard prompt set structure (25 prompts per business):

Recommendation queries (8 prompts):

• "Best [category] in [specific neighbourhood]"
• "Highly rated [category] near [landmark]"
• "Top [category] in [city] for [specific need]"
• "Which [category] in [area] is best?"
• "Good [category] near me [city]" (vernacular variant: "mere paas best [category] kaun hai")
• "Recommended [category] in [city]" (3 variants covering different geographic references)

Service-specific queries (8 prompts):

• "Where to get [specific service] in [city]"
• "[Specific service] clinic/centre in [neighbourhood]"
• "[Service] cost in [city]" (3 queries covering different services the business offers)
• "Best [service] near [landmark]"
• "[Service] specialist in [area]"

Informational queries (6 prompts):

• "How much does [service] cost in [city]?"
• "How long does [procedure/service] take?"
• "Is [procedure] safe for [Indian context]?"
• "What to expect at first [category] appointment?"
• "[Procedure] process in India step by step"
• "[Service] FAQ [city]"

Comparative queries (3 prompts):

• "[Option A] vs [option B] for [condition/need]"
• "Best [category] for [specific patient/customer type]"
• "Compare [service A] and [service B]"

Measurement Protocol

Measurement conditions:

• All queries run in incognito/private browsing mode to prevent personalisation bias
• One query per fresh browser session for ChatGPT (new conversation each time)
• Queries run on consistent hardware from consistent IP addresses
• All engines queried within a 24-hour window per monthly measurement session to minimise temporal bias

For each prompt × engine combination, analysts record:

• Citation presence: named in response / not named / named with negative framing
• Citation source: which URL/platform was cited (GBP, Practo, business website, JustDial, etc.)
• Citation language: verbatim or paraphrased description of the business in the response
• Engine response type: browsing (live web) / training data (no live sources) / hybrid

Data validation:

• 10% of monthly measurements are independently re-run by a second analyst
• Disagreements > 5% in citation presence classification trigger measurement protocol review

What would invalidate this measurement approach

The single biggest threat to SAV measurement isn't sampling error, it's non-determinism — AI systems can and do give different answers to the identical prompt asked twice in a row, because of temperature settings and live-web variability in what gets retrieved at that moment. A single query run once per business tells you almost nothing; it might catch the business on a day it was cited and miss it on a day it wasn't, for reasons that have nothing to do with the business itself.

That's why this design runs the full 25-prompt set monthly rather than once, and why the competitive SAV extension runs the identical prompt set against the top three competitors in the same cell — a business's SAV number only means something next to what its actual local competitors are getting, not in isolation. A business scoring 20% SAV against competitors averaging 8% is winning its category; the same 20% against competitors averaging 45% is losing badly. The raw number without the competitive context is close to meaningless.

The second real threat is engine drift. ChatGPT, Perplexity, and Google's AI systems change their retrieval and ranking behaviour on their own release schedules. A SAV increase from month 4 to month 5 could reflect a platform-wide change in how often any local business gets cited, not anything the business did. The study logs major known platform changes against the measurement calendar specifically so this can be separated from a genuine business-level trend later.

How a business can run a scaled-down SAV check itself

You don't need 25 prompts and four engines to learn something directionally useful about your own SAV. Pick five to eight of the recommendation and service-specific prompt patterns above, substitute your own category, city, and neighbourhood, and run them in ChatGPT, Perplexity, and Google's AI Overview (via a normal search) in separate incognito sessions. Do this once, then repeat it a week later — the repeat matters, because a single pass can't tell you if what you saw was typical or a fluke.

Record, honestly, whether you were named at all, and if so, whether the source cited was your own website, your GBP, or a directory listing. Then run the identical prompts substituting your top two competitors' names or letting the query stay generic — competitors will surface on their own if they're being cited for the same query. That gives you a rough, self-measured SAV comparison without needing the full four-engine, twenty-five-prompt, twelve-month design this study runs at scale. It won't be statistically defensible, but it will tell you whether you're closer to 0% or 30%, which is usually enough to decide whether AI visibility deserves dedicated budget this quarter.

SAV Calculation Methodology

Standard SAV formula: SAV = Total citation count / (Total prompts × Total engines) × 100

For a 25-prompt × 4-engine study: SAV = Citations / 100 × 100 = Citations%

Engine-specific SAV: Engine SAV = Citations on that engine / Total prompts × 100

Prompt-type SAV: Type SAV = Citations on queries of that type / (Queries of that type × Engines) × 100

Competitive SAV: For each business in the sample, the same prompt set is run for the top 3 competitors in that category × geography cell. Competitive SAV creates a market-share view of AI citations.

Correlation Analysis: AIO Readiness Score → SAV

A primary research question: does higher AIO Readiness Score predict higher SAV?

Analysis methodology:

• Pearson correlation between AIO Readiness Score (0–100) and SAV (0–100%) at baseline measurement
• Sub-group analysis: correlation by pillar (which pillar explains most variance in SAV?)
• Longitudinal analysis: for businesses that improved AIO Readiness Score by 15+ points over 6 months, what is the observed SAV change?

Hypothesis (to be tested against data): H1: AIO Readiness Score and SAV are positively correlated (r > 0.6) H2: Review Pillar score has the strongest individual correlation with SAV H3: Schema Pillar improvements show SAV impact within 60–90 days of implementation

This hypothesis structure connects directly to the India AI Search Readiness Report, which is where the AIO Readiness Score itself is defined and benchmarked. If H1 holds, the readiness score becomes a genuine leading indicator businesses can act on before waiting for their own SAV numbers to move.

What US-market share-of-voice research doesn't tell us

Share of voice as a concept has a long history in Western digital marketing measurement, but no published US study measures AI citation share using an Indian directory landscape, Indian city geography, or Hindi-language prompts — the underlying source ecosystem AI engines browse is different enough (Practo and JustDial instead of Yelp, WhatsApp Business instead of a US equivalent) that a US SAV benchmark simply isn't a stand-in for what's needed here. This study is being built from Indian observation rather than adapted from an existing dataset, and no specific US SAV figure is cited anywhere in this methodology because none has been verified as applicable to the Indian market.

Limitations and Transparency

AI system non-determinism: AI systems may produce different responses to identical prompts at different times (temperature variance). Monthly measurement protocol mitigates but doesn't eliminate this.

Platform changes: AI systems update frequently. Platform changes may affect citation patterns independent of business-level optimisation. Study notes include major platform changes that may have affected measurement periods.

Sample representativeness: The sample of 10 businesses per industry × city cell is intentionally stratified, not random. Findings are directional, not statistically representative of all Indian businesses in the category.

Commercial context: Angryturtle is a commercial AEO provider. Research findings are published transparently to build credibility and establish industry benchmarks. All findings are based on observed data; no findings are fabricated to support commercial claims.

How and when findings will be published

Aggregate industry- and city-level SAV benchmarks, plus the AIO Readiness → SAV correlation results, will be published as part of Angryturtle's broader India AI search research once the baseline measurement wave and its second confirmatory pass are both complete. Individual business SAV data stays confidential; businesses in the sample get their own number directly. For a business that wants ongoing SAV tracking now rather than waiting for the aggregate study, Share of AI Voice Tracking runs individualised measurement on the same core methodology.

FAQ Section

Q: How can I see my business's SAV in this study? A: Businesses that are part of the research sample receive their individual SAV data as part of Angryturtle's AI Search Readiness Audit. For businesses outside the sample, the industry and city benchmarks in the published report provide the most relevant comparison.

Q: Does Angryturtle publish the full dataset? A: Aggregate industry and city-level SAV benchmarks are published. Individual business data is confidential.

Q: Is the Share of AI Voice tracker live right now? A: The tracking product exists and is built to run this measurement for individual businesses. The aggregate India-wide study described on this page is a separate, larger research effort still in its sampling and design phase.

Get your business's Share of AI Voice measured →

Internal links: Share of AI Voice Tracking · Share of AI Voice glossary · Share of Local Voice glossary · AI Search Readiness Audit · India AI Search Readiness Report · Vernacular Search & AI Citations in India study · Measuring AI Search Visibility blog · AI Citations: ChatGPT, Perplexity, Gemini blog · Prompt Coverage glossary · Prompt Space glossary · Learning Centre: Local SEO Measurement · Perplexity Optimization India · Gemini / Google AI Mode Optimization · Get Cited by ChatGPT