How we calculate your AI visibility data

1. How we detect a mention

For every tracked question, we send it to each AI engine you've enabled and read back the actual answer text — the same text you can click "view original answer" to see in full. We then check whether your brand's name (or any alias you've added) appears in that text, as a real standalone word — not as a substring buried inside a longer word. "Abb" won't match inside "AbbVie," and "3M" won't match inside "M3M." That's it: a mention is your name genuinely appearing in what the AI actually said, nothing inferred.

We also record where in the answer you were named (first, in the top 3, mentioned further down, or buried), which competitors appeared alongside you, and — when the engine provides them — the sources it cited.

2. Visibility Score

The headline number on your brand page.

Visibility Score is the share of checks, over the last 30 days, where your brand was named. The denominator is every (question × engine) combination we've run in that window — so "62%" means "named in 62% of the checks we actually ran," not a guess extrapolated from a handful of samples.

When you're on a plan that doesn't cover every engine, we say so explicitly ("measured on 3 of 7 engines") rather than quietly present a partial score as if it covered the whole market.

3. Share of Voice & Recommendation Share

Who gets named more — you, or your competitors.

Visibility Score only tells you about yourself. Share of Voice puts you next to your tracked competitors: of every time any tracked name (you or a rival) was mentioned across your answers, what percentage was you. The percentages are normalized to add up to 100% across everyone counted, so "you 42%, Competitor A 31%, Competitor B 17%" is a real split, not three independent rates that happen to overlap.

Recommendation Share is the same comparison, but counts only the times AI actually endorsed a name — see the next section for what that means — rather than every time it was simply said out loud.

4. Recommendation Rate & the Recommendation Gap

Being named isn't the same as being recommended.

"Hoka is one option" and "I'd recommend Hoka" are both mentions — Visibility Score treats them identically. Recommendation Rate doesn't. Every mentioned answer is graded into one of three bands:

• Named only — your brand is stated as a fact, with no endorsement language.
• Recommended — the answer actively suggests you ("I'd recommend," "a great choice for," "best for X").
• Top pick — the answer calls you the best/first choice outright, or pairs recommend-language with being the first thing named.

Recommendation Rate counts "recommended" and "top pick" together, as a share of the same total-checks denominator Visibility Score uses — which is what makes the Recommendation Gap a real, subtractable number: Visibility Score minus Recommendation Rate. A large gap means AI already knows your brand and names it often, but usually reaches for someone else's language of endorsement when it's actually time to pick one — a persuasion problem, not a visibility one.

This grading is a pattern match against the answer text, not a judgment call by a language model. It looks for real, common endorsement phrases ("recommend," "best for," "ideal for," "top pick," and similar) in the text immediately around your name. It runs on every answer for free, with no extra API call — which also means it can miss a genuine endorsement phrased in a way it doesn't recognize, and it currently only reads English. We'd rather under-count a recommendation we can't confidently detect than invent one.

5. Sentiment

Sentiment (positive, neutral, or negative) works the same way as recommendation grading: we read the sentences around your brand's name and count positive versus negative signal words from a fixed list (terms like "reliable," "expensive," "buggy," "innovative"). At least two matching signal words are required before we call a mention non-neutral — a single stray word like "limited" in "limited but focused" isn't enough on its own to tip a verdict. Like recommendation grading, this is a free, deterministic text match, not an AI-generated opinion about your brand.

6. Citations & the Citation Gap

When an engine's answer cites a source — a URL, a review site, a YouTube video — we capture the domain, the page title, and the URL directly from what the engine returned. "Sources AI Cites" ranks those domains by how often they show up across your citation-bearing answers in the last 30 days, and tags each one with whether it names you, a competitor, or both.

The "Where rivals beat you" list and the "Why is [competitor] beating you?" narrative are built the same way as everything above: real counts, stated as sentences, never generated freeform. A line like "Competitor X is cited by 14 sources that don't cite you" is a literal count over your page's own top-15 cited-source list — described as such, not implied to be an exhaustive web-wide count.

7. Opportunity Score

Every recommendation you're shown (what to publish, who to pitch, which gap to close) carries a 0–100 Opportunity Score, so you can sort by likely payoff instead of scanning three separate impact tiers. The score combines four real inputs:

• Impact tier — how dominant the competitor actually is in the gap answers we analysed (repetition and consistency, not a single lucky mention).
• Confidence — what share of the analysed gap responses actually named that rival.
• Real search demand — the query's actual monthly search volume, when we've fetched it. A question nobody searches for scores lower than one people actually ask, and a not-yet-fetched volume is left out of the score entirely rather than counted as zero demand.
• Breadth — whether one piece of advice closes the gap on more than one tracked question at once.

A high-impact recommendation with weak supporting evidence can be outscored by a medium-impact one backed by strong demand and confidence — that's deliberate. The number is meant to reflect real payoff, not just repeat the tier label as a bigger font.

8. Confidence — when a number is too thin to trust

A percentage built from 2 checks swings by roughly 50 points on the very next result. We don't hide that. Wherever a headline number is resting on very few checks, we say so plainly — "Early read: built on 3 checks so far" — rather than paint it green or red as if it were a settled verdict. The same logic gates the newer Recommendation Rate card: it only states a persuasion verdict once there's enough data behind it to mean something.

9. What this data can't tell you yet

In the interest of the same honesty this whole page is trying to model:

• Recommendation and sentiment grading are heuristics, not AI judgments. They're fast, free, deterministic pattern matches — accurate on common phrasing, but they can miss an endorsement worded unusually, and they currently only recognize English text.
• Competitor-side metrics (Recommendation Share, Sentiment, "top pick" status) are graded on a capped recent sample (up to 25 answers), not your full history, and a competitor can never be credited with "top pick" purely by list position the way your own brand can — only by an explicit superlative next to its name. That undercounts a competitor's top picks slightly; it never invents one.
• We don't yet verify factual accuracy — whether an AI's description of your pricing, products or features is actually correct — only whether, and how, you were mentioned.
• Search demand data refreshes monthly, not per-scan, so a very new or newly-trending question may show no volume yet rather than a stale or invented one.