Methodology

This page describes the general approach TheAEOWire uses to measure AI visibility in its original studies. Individual pieces may note deviations or additional specifics; where they do, the study's own methodology section takes precedence over this general description.

What we are trying to measure

AI visibility is not one thing. A brand can be mentioned in an AI-generated answer without being cited as a source, cited without being recommended, or recommended without either. We treat these as distinct outcomes and try to report which one a given result actually reflects:

• Mention — the brand or product name appears anywhere in the generated answer.
• Citation — the answer links to or explicitly names a source (a URL, a publication, a page) tied to the brand.
• Recommendation — the answer suggests the brand or product as an option in response to a comparative or "what should I use" style prompt.

Share of voice, when we report it, is the proportion of sampled answers in which a brand appears (by mention, citation, or recommendation, as specified) relative to competitors named in the same category and prompt set. We state which definition applies in each study.

Prompt sampling

Studies are built around a defined set of prompts representative of real user queries in a category, not a single question repeated. We document the prompt set, how it was constructed, and its size for each study. Where prompts are adapted from real search or chat query data, we note the source.

Engines and locales

We sample across the AI systems relevant to a study's scope, which may include ChatGPT, Google AI Overviews, Gemini, Perplexity, and others as relevant. Each study states which engines were tested, in which locale(s) and language(s), and on which dates. Results from one engine or locale are not assumed to generalize to another, and we avoid aggregating across engines without saying so.

Repeated sampling

AI-generated answers vary between runs, even for an identical prompt, due to model non-determinism, ongoing model updates, and personalization. Because of this, we sample each prompt multiple times rather than once, and report results as a distribution or frequency (for example, "cited in 6 of 10 runs") rather than a single pass/fail outcome. Sample sizes are stated with each result.

Personalization and account state

Answers can differ based on account history, location, device, and session state. Where practical, we sample from logged-out or fresh sessions to reduce personalization effects, and we note when a study could not fully control for this. Findings should be read as a snapshot under the stated conditions, not a claim about what every user sees.

Model updates

The systems we study change frequently and without full changelogs. A result captured on one date may not hold after a model update. Every study is timestamped, and we treat findings as accurate as of their sampling window rather than as permanent facts. Where we re-run a study, we publish it as a new dated piece and link back to the prior result for comparison.

Statistical caveats

Our sample sizes are generally modest relative to total query volume, and results should be read as directional rather than exhaustive. We do not present correlation as causation, and where a study cannot rule out confounding factors — such as brand size, existing search rankings, or content freshness — we say so in the write-up. Raw data and counts are made available with the study wherever practical so readers can assess the underlying numbers themselves.