Method
Odds, not ranks.
Pop Sota measures AI visibility as odds. We ask each AI assistant your buyers' real questions 10–20 times, count how often it recommends you, and report that rate with a 95% range. Then we trace which sources drive the answers, fix the ones that matter, and re-test against control clusters we deliberately leave alone.
Four steps
Mine, sample, trace, move
-
01 · Mine
Start from what buyers actually ask
We pull three months of sales calls, support tickets and community threads, redact personal data, and cluster the buyer questions into 8–12 "money clusters" of 5–10 prompts each.
Two clusters become controls: we measure them but never work on them, so we can tell our effect from the market moving.
-
02 · Sample
Ask every assistant, many times
Each prompt runs 10–20 times on each of 7 assistants. We combine consumer-app data from licensed vendors with an official-API baseline, and we store every raw answer.
Odds are reported per cluster, not per prompt, so the samples are large enough to mean something.
-
03 · Trace
Find the sources behind the answers
We rank the URLs and domains each cluster's answers cite, and label each one: your site, review sites, Reddit, YouTube, publishers or competitors.
This is where the work is. Assistants repeat what their sources say.
-
04 · Move
Fix, then re-test against controls
We ship fixes on the highest-ranked sources: answer pages on your site, earned placements, profile updates and correction requests. All offsite work is disclosed and logged.
Every month we re-sample. Every quarter we report the change on worked clusters against the controls, with error bars.
Why odds
Ask twice, get a different shortlist
AI assistants rarely give the same answer twice. In one study, ChatGPT and Google's AI returned the same brand list less than 1% of the time, and the same order less than 0.1% of the time. How often a brand appears was far more stable than where it appeared. So we measure appearance rates across many runs and say how sure we are.
Source: SparkToro and Gumshoe, 2,961 responses from 600 volunteers, January–February 2026.
Same 50% true rate, three sample sizes:
How many runs is enough?
At a 50% recommendation rate, the 95% range is about:
- 30 runs
- ±17 points
- 100 runs
- ±10 points
- 400 runs
- ±5 points
That's why we report clusters of related prompts rather than single prompts, and why a move only counts when it clears the range and beats the controls.
Coverage
7 assistants, measured separately
Usage is splitting across assistants. ChatGPT's share of AI-assistant web visits fell from about 76% to about 53% in a year while Gemini and Claude grew, so one engine is no longer enough.
Source: Similarweb, share of generative-AI web traffic, June 2025 to May 2026.
- ChatGPT OpenAI
- Gemini Google
- Perplexity Perplexity
- Claude Anthropic
- Google AI Mode Google
- Microsoft Copilot Microsoft
- Meta AI Meta
Google AI Overviews are included wherever your buyers' questions trigger them.
Data
Where the numbers come from
- Official APIs and licensed data vendors only. No scripted logins to consumer apps, which their terms prohibit.
- Every raw answer is stored, so any figure can be traced back to the answers behind it.
- Every report states its method: sources, sample sizes and how the ranges are calculated (95% Wilson score intervals).
- Your call transcripts and tickets are redacted before any model sees them, kept in a separate workspace, and deleted after 30 days.
Benchmark
The B2B SaaS AI Shortlist
Our quarterly benchmark will show, category by category, how often each assistant recommends each vendor, with the same error bars we use for clients. Edition 1 publishes in January 2027.
- 5–10 B2B SaaS categories per edition
- All 7 assistants, measured separately
- Odds with 95% ranges, never ranks
- Listed vendors get their full odds card free
See your own odds.
The Odds Baseline runs this method on your buyers' questions in 10 business days.