Methods and evidence

How we measure AI visibility

A result is only useful when you know what was asked, which system answered and what counted as a mention. This hub connects each study to its own methods and limitations.

Choose the right source

Published v3 observational audit

The authoritative reanalysis of 2,729 businesses. The earlier 1,004-business sample is included in this total, not an independent replication.

API versus consumer search pilot

Published 20 August 2026. An exploratory comparison of collection routes across 1,000 buying questions. One run per question and route.

V6, Wave 2: methods preview

Planned for 17–21 November 2026. Final collection dates, completed sample, protocol and results are pending. This is a research plan, not a completed study.

What the published audit measures

The v3 paper is dated 2026-05-07. It reports an observational audit of 2,729 businesses across 14 verticals and four metropolitan markets: Los Angeles, New York, Chicago and Sydney.

ChatGPT, Claude and Perplexity have complete coverage in this sample. Gemini and Google AI Overviews cover a subset. The study used one sample per business, model and prompt cell; it did not measure within-cell repeat variability. Collection took place in April–May 2026.

The two sampling phases used different procedures: URL-first for the initial 1,004 businesses and name-first for the 1,725-business expansion. Treat the study as a defined observational sample, not a representative survey of every business.

Read the authoritative v3 paper, including methods and limitations. Status: preprint, not peer reviewed. Earlier paper versions are superseded; the v2 paper and Phase 1 dataset record were withdrawn.

Read a visibility number carefully

Define the denominator. A share of prompts, a share of competitor mentions and a rate across business-market comparisons answer different questions.
Record the collection route. Consumer interfaces and APIs, with or without search, are different measurement conditions.
Keep engines separate. Coverage and associations can differ by system. A blended historical score has a specific recipe and limits.
Separate absence from failure. A valid answer without a target mention is an observed zero. A timeout or unresolved entity match is a missing or uncertain measurement.
Preserve the time window. A result describes the collected answers at that point in time. Repeated measurements and prompt sensitivity need separate reporting.
Keep association distinct from causation. The observational audit cannot establish that changing a signal will cause a visibility increase.

Benchmarks, access and disclosure

The public benchmark browser uses the historical Phase 1 study tables. Its published percentile buckets require at least 10 businesses. Read the sample size for the exact industry, region and signal; the v3 paper’s full sample is a different analysis.

The registered Phase 3 trial did not launch and enrolled no businesses. It supplies no causal findings.

Joel House is the founder of Outrigger. The product’s commercial interests overlap with the research questions. Published summaries and papers are public; underlying business-level study data remains private.

How to read GEO benchmarks · Definitions and denominators · Research inquiries

Sources: v3 paper, sections 3, 5 and 6; August methods pilot; public benchmark browser. Reviewed 3 November 2026.

See where your brand stands

Start with what AI says about your business.

See what AI says about you See pricing