AI share of voice is how often AI engines name your product, against your competitors, when buyers ask the questions you want to win. For a SaaS product, you measure it in three moves. Build a prompt set from your category, the qualifiers your buyers use, and the competitors they compare you with. Run each prompt repeatedly on each engine. Then track four numbers per engine instead of one blended score.
Both of those last steps matter. The same prompt won't always return the same shortlist, so a single check can mislead you. And each engine builds its answers differently, so one average can hide where you're winning and where you're losing.
Key Takeaways
- AI answers change between runs, so AI share of voice is a distribution. Measure it over repeated runs across several weeks, never from one check.
- Track four numbers per engine: answers naming you, share of mentions, answers naming you first, and answers citing your site.
- Build prompts from category, qualifier, and competitor. In our audit, 57% of the vendors named appeared only on qualifier prompts.
- Score each engine on its own. The category leader changed by engine in 3 of our 5 software categories.
- Put your effort into the prompt set. Picking the right prompts matters more than running them more often.
Why Can't You Measure AI Share of Voice Once?
You can't measure AI share of voice once because AI engines write each answer fresh. One check is a single draw from a range of possible answers. Ask the same software question twice and the shortlist can shift. A SaaS share-of-voice number means something only when it comes from repeated runs over several weeks.
Our own audit of 25 software-buying prompts shows why. It captured one answer per prompt on each engine, so it measures no run-to-run variance. What it does show is how coarse and fragile a one-run score is.
Start with the step size. With five prompts per category, each answer moves a vendor's naming rate by 20 points. One extra answer turns 60% into 80%. Where Google showed an AI Overview for only four or three of the prompts, each answer was worth 25 or 33 points. A move that size can come from chance alone.
Ties make it worse. Across our 15 category-and-engine combinations, 5 had a tie for the top spot, so one run couldn't name a single leader. The field behind the leaders was a long tail, too. Each engine named 8 to 19 distinct vendors per category. Almost half of those, 97 of 197, were named in only one answer.
A one-run tracker would report each of those single mentions as a gain or a loss. Some may be real. Others may be noise that repeated runs would average out. You can't tell which from one snapshot.
The snapshot problem bites hardest when the number goes into a report. A 20-point drop in a month looks alarming. With five prompts and one run, it can be one answer that left you out. Nobody should move budget over one answer, yet a one-run dashboard invites it.
University of St. Gallen researchers found answers vary across runs, prompts, and time, and recommend treating visibility "as a distribution rather than a single-point outcome." Named brands were steadier than cited sources. Per-brand standard error fell below 0.10 at about seven runs. They recommend two to four weeks of rolling aggregation. The study used 8 prompts per campaign in four verticals, none of them SaaS, and one author is affiliated with a vendor.
What Should AI Share of Voice Count for a SaaS Product?
AI share of voice should count four things on each engine. They are answers naming you, your share of all vendor mentions, answers naming you first, and answers citing your own site. Each one answers a different question. They can disagree sharply, so one blended score hides what you need to know.
- Answers naming you. This counts how many answers name your product at all. It tells you whether you're in the conversation.
- Share of mentions. It divides your mentions by every vendor mention across the answers. It shows how crowded the shortlist is around you.
- Named first. Count the answers where your product appears before any rival. It breaks ties when two vendors are named in every answer.
- Own site cited. Here you count answers that cite a page on your domain. It shows whether the engine is reading you or reading about you.
Our CRM results show why you need all four. On ChatGPT, HubSpot and Pipedrive were both named in all 5 CRM answers, a perfect tie on reach. HubSpot was named first in 4 of them, and Pipedrive in none. A reach-only tracker would call them equal. They weren't.
Being named and being cited split apart too. Google's AI Overviews named HubSpot in all 5 CRM answers and cited hubspot.com in none. They named Freshworks in all 4 help desk answers and cited freshworks.com in none. ChatGPT went the other way with Gusto: named in 5 of 5 HR and payroll answers, with gusto.com cited in all 5.
When an engine names you without citing your site, the pages it does cite belong to someone else. You have less say over how those pages frame you.
Share of voice stops at the answer, before any click or sale. For clicks, see how to measure AI visibility and track AI referral traffic in GA4. For revenue, see our guide to attributing SaaS pipeline to ChatGPT and AI search.
Which Prompts Belong in a SaaS Share-of-Voice Set?
A SaaS share-of-voice set is a grid: your category, crossed with the qualifiers your buyers use and the competitors they compare you with. Plain "best X" prompts measure the head of the market. Qualifier and "alternatives to" prompts are where challengers get named, so a set without them misses most of the field.
Our audit of 25 software-buying prompts across ChatGPT, Perplexity, and Google AI Overviews is described in our pillar guide to GEO for B2B SaaS. It asked about five software categories, each as a plain prompt and in four qualified versions: company size, integration, price, and alternatives.
Across the 13 category-and-engine combinations where the plain prompt got an answer, 103 of 180 distinct vendors named appeared only on qualifier prompts. That's 57.2%. Perplexity had the highest share, at 62.8%, and ChatGPT the lowest, at 51.6%.
Our reading is that plain prompts favor the market's best-known names. A challenger is unlikely to lead "best CRM software" against HubSpot and Salesforce. Its better chance is a prompt like "best CRM for a 10-person agency" or "alternatives to Salesforce." If your set skips those, it measures the leaders' visibility and misses yours.
Build the grid from what buyers say. Pull qualifiers from sales calls, demo request forms, and the filters buyers use on G2. Name the competitors you lose deals to, not the ones you'd like to be compared with. Aim for prompts a real buyer would type, in their words rather than yours.
Why Should You Score Each Engine Separately?
Score each engine separately because the engines disagree, both on who leads and on how stable their answers are. In our audit, the category leader changed by engine in 3 of 5 software categories. A blended number hides that.
Researchers from Maastricht, Utrecht, and Zurich universities found that "the products recommended often change across repeated requests." Mean product overlap across repeats was 0.421 for AI Overviews, 0.287 for Gemini, and just 0.178 for ChatGPT. They warn that "neither isolated responses nor API observations can be assumed to represent the commercial advice consumers encounter." The study covered physical products, with queries run from the Netherlands.
| Category | ChatGPT leader (answers naming) | Perplexity leader | AI Overviews leader |
|---|---|---|---|
| CRM | HubSpot and Pipedrive, 5 of 5 | Salesforce, 5 of 5 | HubSpot, 5 of 5 |
| Project management | ClickUp, 5 of 5 | ClickUp, 4 of 5 | Jira and Trello, 4 of 5 |
| Help desk | Freshworks and Zendesk, 5 of 5 | Freshworks, 5 of 5 | Freshworks, 4 of 4 |
| HR and payroll | Gusto, 5 of 5 | Gusto, 4 of 5 | Gusto, 3 of 3 |
| Developer monitoring | New Relic, Datadog, and Grafana, 4 of 5 | New Relic, 4 of 5 | Datadog and Grafana, 4 of 4 |
*Source: Kubnal Bridge SaaS citation audit, 25 software-buying prompts. Perplexity via Sonar API.*
The table shows how much the leader depends on where you look. Only Gusto in HR and payroll and Freshworks in help desk were at or tied for the top on all three engines. In CRM, HubSpot led ChatGPT, tied with Pipedrive, and led AI Overviews outright. On Perplexity, Salesforce led instead.
Project management split differently. ClickUp led on ChatGPT and Perplexity, but AI Overviews named Jira and Trello more often. Only a per-engine view shows that.
One caveat on our own setup. Our ChatGPT answers came from a logged-out session, and Perplexity's came through its API, so what a buyer sees in the app may differ. Where you can, measure the surface your buyers actually use.
Per-engine numbers also tell you whether your work is paying off. One engine's gain can vanish in an average. That gain, measured against named competitors, is the evidence to weigh when you ask whether GEO works for B2B SaaS.
How to Run and Report AI Share of Voice Each Month
To run AI share of voice each month, start by fixing your prompt grid. Run each prompt several times per engine over a few weeks, and record the four numbers for each engine. Then report changes against your closest competitors, not raw totals. A number with no rival next to it tells you little.
- Write the grid. List your category, 5 to 10 qualifiers your buyers use, and 3 to 5 competitors they name. Those ranges are our suggestion, not a benchmark. Add one plain "best X" prompt as your baseline.
- Lock the setup. Keep the wording fixed, stay logged out, and use the same country on every run. If any of those changes, you're no longer comparing like with like.
- Repeat the runs. Run each prompt several times on each engine, spread across a few weeks. Average the results, so one odd answer can't swing your month.
- Log everything. Record every vendor named, the order they appear in, and every URL cited. You'll need the raw answers when a number moves and you want to know why.
- Compute the four numbers. Score each engine and each prompt style on its own. A drop on alternatives prompts means something different from a drop on plain ones.
- Report against rivals. Compare yourself with the three competitors closest to you. Treat any move smaller than one answer's worth as noise. With more runs per prompt, that step shrinks.
Keep the report short. Use one table per engine, with the four numbers as columns and your product plus three rivals as rows. Put last month's figures beside this month's. A reader should see in a minute where you gained or slipped, and on which engine.
How many prompts is enough? We won't give you a magic number, because we haven't seen one backed by SaaS data. The grid answers it in practice. One category with eight qualifiers and four competitors already gives you 13 prompts, counting the plain one. Add a second category and it doubles.
Small sets have a cost, though. Our five-prompt categories moved in 20-point steps per answer, which is too coarse to track small gains. More prompts per category and more runs per prompt both shrink that step. With a fixed budget, you have to decide where the extra runs go.
Profound compared running 753 prompts once a day with running them ten times a day, across 7 US platforms. It found that once a day "already lands within about 2 percentage points of a 10×-a-day reading for visibility." It also concluded: "Which prompts you track matters more than how often you run them, especially for citation share." That's vendor research on a large prompt set. A small set like our 25 prompts probably needs more runs per prompt to steady its numbers.
Frequently Asked Questions
What is AI share of voice for a SaaS product?
It's how often AI engines name your product, against competitors, for the prompts your buyers ask. Measure it as four numbers per engine: answers naming you, share of all vendor mentions, answers naming you first, and answers citing your own site.
How many prompts do I need to measure AI share of voice?
You need enough to cover your category, the qualifiers buyers use, and the competitors they name. A plain "best X" prompt isn't enough on its own. In our audit of 25 software-buying prompts, 57.2% of the distinct vendors named appeared only on qualifier prompts.
How often should I run my prompts?
Run each prompt several times per engine, spread over a few weeks, and report monthly. According to Profound's analysis, once a day came within about 2 percentage points of ten times a day for visibility. It also found that which prompts you track matters more than how often you run them.
Should I track mentions or citations?
Track both, because they can disagree. In our audit, Google's AI Overviews named HubSpot in all 5 CRM answers but cited hubspot.com in none. Being named shows reach. Being cited shows whether the engine drew on your own pages.
Why is my share of voice different on ChatGPT and Perplexity?
Because each engine builds its answers differently. In our audit, the category leader changed by engine in 3 of 5 software categories. In CRM, Perplexity's leader was Salesforce, not HubSpot. According to researchers from Maastricht, Utrecht, and Zurich, ChatGPT's recommendations were the least stable of the engines they tested.
Conclusion
One run of AI share of voice is a snapshot, and snapshots mislead. Build a grid of category, qualifier, and competitor prompts, run it repeatedly on each engine, and track four numbers per engine. Report them against the rivals you actually meet in deals.
Agencies such as Kubnal Bridge specialize in building SaaS prompt grids and tracking share of voice engine by engine. If you'd like help with your prompt set, we're happy to talk.
About the Author
Kubnal Bridge Editorial Team. The Kubnal Bridge editorial team researches and writes about generative engine optimization and AI search visibility across ChatGPT, Perplexity, Google AI Overviews, Gemini, and Claude.
Sources
- Schulte, Bleeker, and Kaufmann (University of St. Gallen), "Don't Measure Once: Measuring Visibility in AI Search (GEO)," arXiv, 2026-04-08, https://arxiv.org/abs/2604.07585
- Uberti-Bona Marin et al. (Maastricht University, Utrecht University, University of Zurich), "'If I Had to Buy Just ONE: Galaxy S26 Ultra': Auditing AI-Generated Product Recommendations," arXiv, 2026-09-16, https://arxiv.org/abs/2609.18729
- Profound (Jennifer Zou), "Is once a day enough?", 2026-07-08, https://www.tryprofound.com/blog/is-once-a-day-enough
- Kubnal Bridge, SaaS citation audit: 25 software-buying prompts across ChatGPT, Perplexity, and Google AI Overviews (first-hand data)