Methodology · updated
How we audit, build and measure AI visibility
In short
We treat AI visibility as three steps: be retrievable, get selected as a source, get used in the answer. We audit against Google’s published inputs, fix feeds and pages, and measure with repeated prompt panels reported as a share of citations, with variance shown. We publish the limits too.
The model: retrievable, selected, used
Retrievable
Crawlers allowed in robots.txt and at the CDN, pages indexed and snippet-eligible, content in text. Google lists these as the inputs for AI features (Google).
Selected
The engine picks your page as a source. Selection depends on authority, brand recognition and context, and differs by engine (arXiv 2604.25707).
Used
The engine draws on what it found. Longer, modular pages with definitions, numbers, comparisons and steps get used most (arXiv 2604.25707).
The share-of-citations panel
Single screenshots prove nothing, because LLM answers vary from one run to the next (JH Scherck). We measure with repeated prompt panels, the approach recommended by the selection-and-absorption framework (arXiv 2604.25707).
- Build the prompt set from real buyer questions in your categories: category (“best trail running shoes”), comparison (“X vs Y”), problem (“shoes for flat feet”) and brand questions.
- Agree and freeze it. You approve the panel before the first run. It stays fixed for the reporting period; any change is logged and dated.
- Run each prompt repeatedly on each engine (ChatGPT, Perplexity, Gemini and Google AI Mode) in clean sessions, with location and date recorded.
- Record two things per run: which domains are cited, and whether the answer uses your facts or products.
- Report a share, not a rank, per engine, with run counts and the spread across runs shown next to every number.
| Metric | Definition | Why it matters |
|---|---|---|
| Citation share | Runs in which an engine cites your domain ÷ total runs, per engine | Measures selection |
| Answer influence | Runs in which the answer uses your facts, products or claims ÷ total runs | Measures use, which can happen with or without a link |
| Competitor share | The same two metrics for competitors you name | Context: share is relative |
| First-party AI data | Search Console’s Generative AI performance report (Google) | Google’s own numbers, as the baseline |
| Business context | Branded search trend and revenue | Visibility only matters if it reaches the till |
How we audit
Google says there are no additional technical requirements for AI Overviews or AI Mode: a page must be indexed and eligible to show with a snippet (Google). So our audits start with fundamentals and add the e-commerce layer:
- Crawl access in robots.txt and CDN or firewall rules, for Googlebot, OAI-SearchBot, ChatGPT-User, PerplexityBot and Perplexity-User. Google-Extended only controls training use, not appearance in Search (Search Console Help).
- Indexation, internal links, page experience and text content.
- Structured data checked against what is visible on the page, never added for content that isn’t there.
- Merchant Center and ChatGPT feeds against the ACP feed spec and Google UCP requirements.
- Freshness, semantic HTML and structured data, the pillars most associated with citation in the GEO-16 audit.
Limits we tell every client
- Answers vary. Share of citations is a probability over runs, not a fixed position.
- No one sees Google’s internal metrics. Google says no third-party tool has access (Google).
- Rank trackers are unreliable right now. They are struggling to scrape Google (JH Scherck).
- GA4 “Direct” is inflated by bots (Brodie Clark), and Search Console data has been delayed by more than 65 hours (Glenn Gabe).
- Much GEO advice is vendor marketing. An evidence review says so (Capston); we prefer peer-reviewed or first-party sources.
- Studies have caveats. In one click-through experiment, the “hide AI Overviews” arm only worked about half the time (GetFoundAgent). We cite such studies with their limits.
How we use sources
Every statistic on this site links to its original source next to the number. Practitioner observations are labeled as observations, not findings. We don’t publish client results, logos or testimonials we can’t show you, and we don’t mark up reviews.
See how this applies to each service: AI visibility reporting, AI-commerce readiness audit.
See what AI shopping agents see when they look at your store.
Send your store URL. A person reviews the public signals AI engines rely on and replies with the gaps we would fix first. Free.