How to measure AI share of voice across 50+ locations

Measure AI share of voice as the percentage of AI answers that name your brand or location across a fixed set of buyer questions, run many times. Never report a single answer or a rank. Answers change on every run, so a 50-location brand needs hundreds of answers per location before one location's number means anything.

This post gives a corporate marketing or ops lead the working method: what to count, how many answers you need per location, how to split local from informational prompts, and what to report to whom. For the wider shift this sits inside, see AEO vs SEO: what changes when search becomes an answer.

Why can't you read an AI answer like a ranking?

Because the list changes every time you ask. SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI results, 2,961 runs in total, collected in November and December 2025. They found a less than 1-in-100 chance that ChatGPT or Google's AI returns the same list of brands in any two responses, and about 1-in-1,000 for the same list in the same order. Gumshoe is an AI tracking company, and the authors disclose it.

The same SparkToro study shows what is stable. ChatGPT named City of Hope hospital in 69 of 71 answers (97%) but ranked it first in only 25. Visibility held up while position did not (SparkToro), which is why the authors call visibility percentage a reasonable metric and ranking position unreliable.

A separate test is more generous about consistency. Detailed.com asked ChatGPT the same five questions 100+ times each, in fresh web conversations from the US, and reported 86% brand consistency, 43% rank consistency and 7% for the same brands in the same order. It tested one engine and five questions, and the page is a live report with no stated year, so read it as direction, not a benchmark. Both studies point the same way: count appearances, ignore order.

How many answers do you need for each location?

About 100 answers per location per engine for a reading within roughly 10 points, and 250 or more if you want 6 points. The arithmetic is the standard margin of error for a proportion, and it is ours, not a published AI benchmark. It assumes each answer is an independent draw, which overstates precision when the prompt set is small, so treat the figures as the best case.

Margins are 95% confidence (our calculation, not a figure from the studies cited here), computed as 1.96 times the square root of p(1-p)/n. The practical meaning: at 100 answers, a location that moves from 40% to 45% has not moved. Most weekly swings you see in a dashboard sit inside the noise.

Tool limits matter here. Ahrefs says most tracking tools cap you at 50 to 100 queries and describes its own approach as aggregation over thousands of prompts, "more like polling." That claim comes from a vendor selling the product and gives no threshold, but the polling framing is right. SparkToro's advice is to run prompts 60 to 100 times and average, and its author says the open question of how many runs are statistically sound is not settled.

What sampling plan works for a 50-location brand?

Use two layers: a brand-level panel read weekly and a location-level panel read monthly. Spreading the same budget evenly across both gives you neither a stable trend nor a usable location number.

For the location layer, a workable design is 20 buyer-phrased prompts per location, each run 5 times per engine. That is 100 answers per location per engine and, across 50 locations, 5,000 answers per engine per monthly period. For three engines, 15,000. These are our planning numbers, not a published standard.

• Prompt set: freeze it for the quarter. Changing prompts changes the denominator, and you cannot compare periods.
• Competitors: fix the list per market, because the competitors in one metro are not the competitors in another.
• Engines: report ChatGPT, Google AI Overviews or AI Mode, and Perplexity separately. A blended number hides which engine moved.
• Fresh sessions: run each answer in a new conversation with default settings, as the studies above did.

Should local and informational prompts be tracked separately?

Yes. They trigger different Google features and name different sources. Whitespark's May 2025 study ran 540 manually collected queries across Houston, Phoenix and Denver and six industries. AI Overviews appeared on 15% of local-intent queries, 92% of informational ones and 97% of hybrid ones. Local packs appeared on 93%, 6% and 17% of the same three groups. The researchers were not physically in those cities, and the sample is one point in time.

Source for the table: Whitespark, 540 queries, May 2025. For a multi-location program the implication is a split scorecard. Local-intent prompts mostly test your Business Profiles and map presence. Informational and hybrid prompts test whether the engine names your brand or a location in an answer. Whitespark's one detailed citation check (same study), nine hybrid plumber queries in Houston, found about 60% of 23 citations went to third-party sites and about 40% to local businesses. That sample is tiny, but it shows why you log the cited sources and not only the mentions.

What should you report, and to whom?

Report brand-level share weekly with the margin printed next to it, location-level share monthly, and a list of locations whose gap to their market's leader is outside the margin. Leadership gets the brand trend, regional leads get their location table, and nobody gets a rank.

Alongside mentions, log which pages and listings the engines cite for each location. Google states in its AI features documentation that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations needed. A page must be indexed and eligible for a snippet. It lists the usual fundamentals, including keeping Business Profile information up to date. Google's Business Profile guidelines add the structural rules a chain needs to follow: one page per location, a real staffed address, and the same name across locations in a country with narrow exceptions. When a location trails its market, check those basics before anything exotic.

Google also says AI Overviews and AI Mode clicks are counted inside the Web search type in Search Console, so you cannot split AI traffic out there. Your own prompt panel is the only direct read on visibility, which is why its design matters.

How to check where you stand

Run this self-test before you buy a tool or read a vendor dashboard. Pick five locations, two strong and three weak by revenue. Write 10 prompts a buyer would type for each, half with the city named. Run every prompt five times in fresh sessions on each engine and record whether each answer names the location. If your tool or your team cannot produce that table with a margin next to each number, you do not yet have a measurement, you have screenshots.

If you want us to run that panel across your locations and show where each one trails its market, book the free AI visibility audit. Our enterprise program page explains how multi-location accounts are scoped.

Frequently asked questions