Vunnara Methodology
Our recommendations draw on published research into how AI engines choose what to cite — including the Princeton GEO study (KDD 2024) and Ahrefs' large-scale AI citation studies. Below are 10 of the highest-impact methods; effect sizes are reported as measured by each source (some are correlations, not guaranteed lifts, and some methods are directional best practices without a precisely verified figure).
What the Visibility Score measures
The Visibility Score is how often AI answers mention your business when customers ask everyday questions about businesses like yours — never a question that names your business directly. A question that does name it is shown separately, on its own, and is never counted toward this score.
Position (where in the answer you're mentioned) and Source citation (how often a mention links to a source) describe how you appear in the answers that DO mention you — they're shown next to the score, but they never change the score itself. Sentiment and Context aren't measured — no tone or framing classifier exists — so both are marked "Not yet measured" rather than shown as a number.
From 19 Oct 2026, the Visibility Score is the mention rate. Earlier audits are recalculated the same way, every time they're viewed — there's no separate, older scoring method still applying to older audits.
How we gather answers
Each brand gets a fixed set of questions, generated once from its confirmed categories plus a few that name the brand directly. The same questions are reused on every later audit unless the brand's categories, competitors, or location change — so a trend line compares like with like, not a freshly-randomised set each time.
Most engines search the live web for each answer by default, so a response can reflect what's actually online right now rather than only older training data. Each question is asked once per engine per audit — a label like "12 questions × 4 engines = 48 answers" means 48 real conversations, not 48 attempts at fewer questions.
An identical question, to the same engine, under the same web-search setting, asked again shortly after may reuse that very recent answer instead of placing a fresh call — this only applies to a byte-identical repeat, and a web-searched answer is reused for a shorter time than a memory-only one, since it can go stale faster.
A brand's suburb or city is only passed to an engine once it's been confirmed — and not every engine supports it: Gemini's search tool has no location-targeting option at all, and even where location is passed, there's no guarantee an engine treats it as a hard filter rather than light context.
Google AI Overviews and AI Mode
Where a workspace has them, we also check what Google Search shows for the same questions: the AI Overview at the top of the results page (google.com.au) and an AI Mode answer (Google runs AI Mode on google.com, set to Australia). Google's results are collected for us by DataForSEO, a search-data provider, logged out, so nobody's search history or personal results are involved.
Each result is for one place and one device: the business's postcode where we have it, otherwise its suburb, its city or its state's capital (the audit then says "Results for Sydney NSW, not your suburb"), and never the whole of Australia. Each result is labelled with that place, the device and the date.
Google decides, question by question and time to time, whether to show an AI Overview at all. "Not shown" means Google showed none for that question when we checked: it isn't a failure, and it isn't "not mentioned". A mention counts only when it's verified, the same way as for every other engine. Google's answers vary with time and place, so one audit is a snapshot.
Google results are shown in their own block and aren't counted in the Visibility % unless the audit says they are.
What changed
November 2026: detection updated — web-grounded (v2.2).Gemini's answers now come from Gemini 3.5 Flash-Lite, the replacement Google names for Gemini 2.5 Flash, which it retires on 1 December 2026. Every engine's answer can now run to 8,192 tokens, well above the longest answers we've measured. And an audit is compared only with audits that asked the same engines for the same models: a plan change that moves Claude from Haiku to Sonnet, or an engine answering with a different model than the one we asked for, is shown in your trends but never compared.
Audits from before this change are still shown in your trends, marked "Detection updated on" the date of your first audit after it, but they're never compared with the ones after.
November 2026: detection improved — web-grounded (v2.1).We now find your name when an answer writes it with formatting or invisible characters around it (for example in bold, or with a non-breaking hyphen), and keep text written before and after a web search apart. An answer the engine didn't finish, or that its own filter stopped, is no longer scored. A mention is confirmed by your website only when the answer cited it, not when it merely came up in a search.
Because this can find mentions the earlier detection missed, audits from before the change are still shown in your trends, marked "Detection improved on" the date of your first audit after it, but they're never compared with the ones after.
From 7 November 2026, share of voice uses verified mentions on both sides. Yours count when verified, as in Visibility; a competitor's count only when its own website, set in brand settings, is among the sources the answer cited. A competitor with no website set is left out of the share, and named, rather than counted on its name alone. Share of voice is worked out from the stored answers each time it's shown, so this applies to earlier audits too.
Add Citations to Reliable Sources
Including citations from credible sources is one of three top-performing GEO methods, improving visibility by 30–40% on the Position-Adjusted Word Count metric in the Princeton study. Most effective for factual queries.
Add Relevant Statistics
Adding relevant quantitative data to content was a top-3 GEO method (+30–40% Position-Adjusted Word Count).
Include Credible Expert Quotes
Incorporating credible quotes was the third top-performing GEO method.
Improve Fluency & Readability
Stylistic improvements (Fluency Optimization and Easy-to-Understand) produced a 15–30% visibility boost in the Princeton study — evidence that generative engines value clear, readable writing, not just content additions.
Build YouTube Brand Mentions
Across 75,000 brands, mentions in YouTube video titles, transcripts and descriptions were the single strongest correlate of AI visibility (r≈0.737) — ahead of every other signal. Note: this is a correlation, not a guaranteed lift.
Earn Branded Web Mentions
Branded web mentions correlate strongly with AI visibility (r≈0.66–0.71) across the same 75K-brand study. Earned media outperforms owned-channel content for AI citations. Correlation, not a direct lift.
Add FAQ & Structured Data Schema
FAQ blocks and structured data are commonly recommended for AI search visibility, but a large-scale Ahrefs study (1,885 pages, Aug 2025–Mar 2026) found no statistically significant lift in AI citations from schema alone — evidence across sources is mixed. A supporting factor, not a guarantee.
Add Structured Comparison Tables
AI engines frequently extract well-structured tables for recommendation and comparison answers — a directional best practice without a specific, verified lift figure.
Front-Load the Answer (first 30%)
Analysis of thousands of ChatGPT citations found 44.2% of all LLM citations come from the first 30% of a page. Put your direct answer, key facts and strongest data in the opening third.
Use Clear Heading Structure
Pages with well-organised headings (clean H1–H3 hierarchy, lists) are easier for AI engines to parse and extract from — a directional best practice without a specific, verified lift figure.
AI Discovery Endpoints
Our Technical Audit checks for four endpoints AI crawlers can use to discover a site's policies and content. Only one is grounded in an external standard; the other three are Vunnara's own recommended format — most sites don't have them yet, and adding them is a forward-looking best practice, not a fix for a standards violation.
- /.well-known/ai.txt (or root /ai.txt) — Emerging standard: the IETF draft draft-car-ai-txt-wellknown-00 specifies a machine-readable AI usage-preferences file served at /.well-known/ai.txt; the older Spawning convention uses root /ai.txt instead. We check both.
- /ai/summary.json, /ai/faq.json, /ai/service.json — Vunnara-recommended structured endpoints for an AI-ready business summary, FAQ content, and a service API/feed. Not yet an industry standard.
AI Crawler Access
- Who we say we are. Every request Vunnara makes to your site identifies itself as
Vunnara-Audit-Bot/1.0 (+https://vunnara.com.au/about). We never send another company's crawler name, so your logs and blocks never blame OpenAI, Anthropic, Google or Perplexity for our visits. - robots.txt.For each AI crawler (GPTBot, ClaudeBot, Google-Extended, PerplexityBot) we read the rules your robots.txt sets for it, as RFC 9309 defines them, and say what they mean: “robots.txt allows GPTBot”, “robots.txt blocks ClaudeBot”, or that it has no rules for that crawler.
- CDNs and firewalls. Whether yours lets an AI crawler through can't be tested from outsidewithout pretending to be it, so we don't. Your own server logs, uploaded to Vunnara, show each crawler's real visits, each checked against the company's published IP ranges or reverse DNS; CDN Shield reports those verified visits beside what robots.txt says.
Research Citations
- Aggarwal et al., GEO (Princeton, KDD 2024) — Generative Engine Optimization — method-level visibility improvements on GEO-bench
- Ahrefs Q1-2026 AI Search Benchmark (75K brands) — Correlation analysis across 75,000 brands — identifies signals associated with AI visibility
- Ahrefs — Schema & AI Citations (1,885 pages, Aug 2025–Mar 2026) — Tracked AI citations before/after adding JSON-LD schema; found no statistically significant lift for ChatGPT or Google AI Mode — schema alone doesn't reliably move AI citations
- Zyppy / Leapd — Citation position analysis — where on a page LLM citations originate