AI visibility audit

How to measure whether AI assistants mention and cite a brand, score it against competitors, and fix the gaps. Covers query design, detection rules, scoring, per-platform ranking factors, and the GEO optimization playbook.

Query set

Test 20–40 queries. Sources, in order of preference: the user's own list; the brand's known industry keywords plus recent keyword-research output; or auto-discovery ("best {category}", "{category} comparison", "{category} reviews", "{brand} {category}") confirmed with the user. Phrase each query as a natural question a buyer would ask an assistant — "What is the best web scraping API?", not the keyword string "web scraping api". Keyword-style phrasing only belongs on search-box surfaces (Google AI Mode / AI Overviews).

Platforms

Query every AI answer surface you can reach — typically some subset of ChatGPT, Perplexity, Google AI Overviews / AI Mode, Gemini, and Grok — capturing for each query the full answer text and the cited-source list with URLs. Also run a plain Google search per query to record whether it triggers an AI Overview (a distinct surface from AI Mode). Platform rules:

• A platform that errors gets one retry, then is excluded for that query with the gap noted. Never fabricate an answer for an unreachable platform, and never abort the run over one platform.
• A query with no AI answer is data, not an error — record "no AI answer present"; those queries are early positioning opportunities.
• Compute all scores only from platforms that actually returned data, and say which were missing.

Detection rules (per query x platform)

Extract: brand mention (in the answer text), domain citation (in the source list), position of first mention / ordinal position among sources, sentiment of the mention sentence (positive / neutral / negative / unknown — judge only the sentence containing the mention, default to unknown), competitor mentions, competitor domain citations, and a 1–2 sentence excerpt as evidence.

• Brand mention: case-insensitive match on the brand name and common variants.
• False-positive guard for common-word brands (e.g. a brand named "Nimble", "Notion", "Craft"): count a mention only if (a) the name appears within 5 words of a domain-relevant term, or (b) the brand's domain is also cited in the sources. A bare occurrence of the common word is generic usage — record brand_mention: false with a note.
• Domain citation: normalize cited URLs to root domain (strip protocol, www., path) before comparing against brand and competitor domains.
• Citations placed earlier in the answer are stronger signals than late ones — record position when the platform exposes it.

Scoring

• Visibility score = % of queries where the brand has at least one mention or citation on any platform.
• Share of AI voice = brand signals (mentions + citations) / all signals including competitors' x 100. If nobody — brand or competitor — appears anywhere, set it to 0 and report "the space may not yet trigger AI answers"; never divide by zero.
• Citation rate = domain citations / brand mentions. High mentions with low citations means the brand is discussed but not linked — a specific, fixable optimization target.
• Per-platform breakdown: queries with an AI answer, mention count, citation count, coverage rate.
• Competitor gap list: queries where a competitor appears and the brand doesn't — the highest-priority targets, each with the platform and the competitor's citation as evidence.

Delta tracking

Keep a per-brand snapshot of every run (per query, per platform: answer present, mention, citation, sources). On re-runs, lead with: citations gained, citations lost, competitor movements, queries newly triggering (or losing) AI Overviews, and score deltas. First run = full baseline, no delta section.

Per-platform ranking factors

Directional profiles from published research (Princeton GEO study, arXiv:2311.09735; SE Ranking's 400K-page study; ZipTie) — treat as priors to test, not gospel:

• ChatGPT: content-answer fit dominates (~55% of citation likelihood); domain authority matters far less (~12%) than in classic SEO; fresh content is heavily favored; lists, tables, and "X is..." definitional openings get cited over prose.
• Perplexity: FAQ schema is disproportionately rewarded; public PDFs get priority; publishing velocity beats keyword targeting; it prefers citing content that itself cites sources; sources appearing in the answer's first paragraph carry the most visibility value.
• Google AI Overviews / AI Mode: E-E-A-T signals dominate; passage-level optimization beats page-level (the AI extracts passages); featured-snippet winners are ~2x more likely to be cited; pages ranking organically in the top 5 supply ~80% of AI Overview citations — traditional rank is still the entry ticket.
• Gemini: mirrors Google AI patterns and the knowledge graph; structured-data work pays on both at once.
• Grok: real-time and social signals weigh heaviest; strongest recency bias of any platform; least studied — directional only.
• Claude (not directly queryable): extremely selective; factual density with specific numbers is the strongest signal; uses Brave Search for grounding, so estimate visibility via Brave rankings and confirm its crawler isn't blocked.

GEO optimization playbook

Nine methods ranked by measured visibility boost (Princeton GEO study):

Method	Boost
Cite authoritative sources	+40%
Add statistics / specific numbers	+37%
Add attributed expert quotes	+30%
Authoritative tone (no hedging)	+25%
Easy-to-understand language	+20%
Appropriate technical terms	+18%
Vocabulary diversity	+15%
Fluency / readability	+15–30%
Keyword stuffing	−10% — actively hurts

Best combinations: fluency + statistics (highest overall); citations + authoritative tone (B2B); easy language + statistics (consumer). Apply the 3–4 most relevant methods per page, not all nine.

Content block sizing — different surfaces extract at different granularities, and each block must stand alone without surrounding context:

• AI-answer citations (GEO): 134–167-word self-contained passages
• Featured snippets / PAA (AEO): 40–55-word direct answers
• Voice: under 30 words, one clear sentence

Optimizing an existing page: identify its target query → check which platforms currently cite it → score it against the nine methods → add the top-three methods first (citations, statistics, quotes) → restructure into surface-sized blocks → re-query after changes to measure movement.

AI crawler access — check first

If AI bots can't crawl the site, no content optimization matters. Fetch robots.txt and check for: GPTBot, ChatGPT-User, ClaudeBot / anthropic-ai, PerplexityBot, GoogleOther, Googlebot. Report which are allowed vs blocked; recommend allowing all unless there is a documented reason. Also check for /llms.txt at the site root (llmstxt.org) — low priority, forward-looking.

Report shape

TL;DR (visibility %, share of voice vs top competitor, delta line, top findings) → per-platform breakdown with top queries where the brand appears → query-level matrix (query x platform: cited / mentioned / absent) → competitor comparison table (visibility, share of voice, citation rate per brand) → optimization opportunities in three buckets: high priority (competitor visible, brand absent — with the specific competitor citation), quick wins (mentioned but not cited — earn the link), coverage gaps (no AI answer triggered yet — early positioning) → a GEO recommendation per high-priority gap naming the methods to apply → strategic interpretation.