How Anovox Scores AI Visibility: The Open Model

Why publish the scoring model at all

Most AI-visibility tools show you a score and ask you to trust it. We think that's backwards: a metric you can't audit is a metric you can't act on with confidence. So here is the complete Anovox scoring model — every dimension, every formula, every weighting. If any of it looks wrong for your situation, we'd rather you told us than silently distrusted the dashboard.

Share of Voice: citations divided by answers — with its uncertainty

The headline metric: SOV = (answers that name your brand ÷ total answers collected) × 100, rounded to one decimal. AI answers vary from one run to the next, so one answer per prompt is a noisy sample. Two things keep the number honest. First, it pools a rolling window — the last 4 weekly runs on Starter, the last 7 days on daily plans — and Starter asks each prompt 3 times per run.

Second, every share of voice ships with its likely range: a 95% Wilson score interval, which stays sensible at small sample sizes and at 0% or 100%. 10 of 20 answers is 50%, likely 30–70%; 0 of 20 is 0%, but up to 16% is still plausible. A change is only called a change when a two-proportion test says it is larger than run-to-run noise at 95% confidence; otherwise the dashboard says "no clear change".

Citation detection: names, aliases, domains, rank

A citation counts when your brand name, any alias you registered, or your domain appears in the answer text as a whole word. Multi-word names tolerate spacing variants ("Auto Chase", "Auto-Chase" and "AutoChase" all count). If you typed your brand with capitals, an all-lowercase occurrence doesn't count — so the brand Notion isn't credited for "the notion that…".

We also record rank: the order in which your tracked brands (you and your competitors) are first named in each answer, 1 being first. Competitors are detected with the same matcher, so an overtake (competitor cited, you not) is a directly comparable event, not an inference. The matcher is plain text and can still miss misspellings; add them as aliases.

Sentiment: how you're described, not just whether

Each brand citation is classified positive, neutral, or negative from the sentences that name you — not the whole answer, so a harsh line about a competitor isn't scored against you. A prompt where you're named with caveats ('X is good but pricey') scores the citation but flags the sentiment — because being mentioned badly is not the same as being recommended. Sentiment feeds the gap ranking: a negatively-framed citation outranks a missing one in urgency.

Gap ranking: severity × priority × recency

Aim ranks every open gap by three explainable factors. Severity: how invisible are you (zero citations across all engines outranks cited-once). Priority: the prompt's business weight. Recency: fresh gaps outrank stale ones you've already decided to ignore. An optional fourth factor — relative demand from prompt-volume signals — can be toggled in. No hidden weights: each factor is shown on the worklist next to the score it produced.

AEO Content Score: structure, authority, citability

The site grader scores three axes out of 100. Structure: can a model extract a clean answer (direct-answer-first copy, headings, comparison tables, schema markup). Authority: do independent sources corroborate your claims, and are your pages crawlable by AI agents at all.

Citability: would a synthesis step choose your page — clear claims, defined terms, quotable numbers. Each axis ends in concrete fixes that deep-link into the content generator.

What we deliberately don't score

We don't score 'brand awareness' or 'share of conversation' as vague composites, we don't blend engines into one opaque number without showing the per-engine breakdown underneath, and we don't present relative demand signals as search volume. Where a number is an estimate, it's labeled as one.

The dashboard always lets you click from any aggregate down to the raw stored answer that produced it — the receipts are one click away, always.