How we measure what AI says about your brand

AuthorityOn AI tells you what the major language models — ChatGPT, Claude, Gemini, Mistral, Grok — say about your brand when someone asks. We do this by running a curated set of prompts every week, capturing every response verbatim, and reading them carefully so you don't have to.

The questions we put to the models are the questions your audience is putting to them: discovery questions ("which companies do X?"), comparison questions ("how does Acme compare to Beta?"), reputation questions ("what do people say about Acme?"). We give you the full picture across all of them.

Why it matters: an increasing share of people now ask AI before they ask Google. If AI doesn't mention you, gets your facts wrong, or describes you negatively, that's shaping your market — and traditional SEO, PR, and brand monitoring tools won't even see it.

Authority isn't one number any more — it's a network of signals AI cross-references when it decides what to say about you. Your AI Score is a weighted blend of two halves: how AI sees you today (60%) and the real-world signals that feed AI's opinion (40%).

How AI sees you — measured every week by running curated discovery, comparison, and reputation prompts through every major model:

• Visibility — when someone asks AI a question your brand should be the answer to, are you in the answer? How often, and where in the list?
• Sentiment — when AI does mention you, does it frame you positively, neutrally, or critically?
• Recommendation rate — does AI actively recommend you for a relevant need, or just mention you in passing?
• Confidence — are AI's answers consistent across providers and prompt phrasings, or are different models telling different stories?

Real-world authority — measured on demand by auditing the channels AI is actually reading when it forms its opinion of you:

• Website — schema markup, structured pages, AI-crawler hygiene. Discoverability for the bots reading your site today.
• Content — prose quality and citation-shape across crawled pages. What AI quotes back when asked.
• Social — LinkedIn presence, content cadence, thought-leader bench. Third-party authority signals AI picks up.
• Earned media — Wikipedia presence, news mentions, third-party citations on review and analyst sites. The signals AI picks up that you didn't pay for.

Each signal is independently 0–100. Missing data defaults to 10 (absence is penalised, not ignored), so a brand with strong visibility but no LinkedIn presence can't ride the LLM half alone. Improve any signal here and the composite shifts.

A score tells you where you stand. The Source mix and Co-occurrence panels tell you where AI is sourcing that impression and who it groups you with — the two questions that turn a number into a strategy.

Source mix.Every story AI tells about your brand is attributed to a source class — recent news coverage, analyst reports, Wikipedia, your own website, social platforms, or the model's own training data. The panel shows the share of each, plus the tone each source-class leans. If AI's impression of you is 70% news (mostly positive) and 20% training data (mixed), you know fresh PR is moving the dial today and that the baseline shifts only when new mentions feed the next training cycle.

Co-occurrence.Whenever AI mentions you in the same answer as another entity — a competitor, peer, institution, or person — we record it. The Co-occurrence panel shows the names AI keeps returning alongside you, ranked by frequency, with sentiment and a flag for whether you're tracking each name as a competitor or peer already. This is the most actionable lens in the product: it shows you who you belong with in AI's mind, and the adjacent brands you could partner with, court for editorial coverage, or position against.

A scan is a sample, not the whole truth. If we run a hundred questions and forty-one mention you, the real-world rate is around 41% — but it could plausibly be 33% or 49%. We show that uncertainty in plain English on every score: a range you can actually trust, computed using a standard statistical method (Wilson, if you like the technical details — see the Technical tab).

The practical upshot: when you see "78, plausibly 75 to 81" one week and "76, plausibly 73 to 79" the next, the range tells you that's noise, not a story. When the next week drops to 64 and the range moves with it, that's a real shift worth investigating.

A score tells you how strong your AI presence is, but it can't tell you why. That's what Stories are for. We read every model response, pull out the individual claims AI is making about you, and group similar claims into a Story.

A Story has a title (e.g. "Patagonia is positioned as the leading sustainability-focused outdoor brand"), a tone, an exposure number (how often it shows up across the answers your audience would see), and a likely source class — news coverage, analyst or industry reports, Wikipedia, your own website, social platforms, or model training data.

That source attribution is what makes the report actionable. A critical narrative coming from social posts is fixed differently than one coming from analyst reports or your own website. Stories give you the narratives; Recommendations turn them into things to do this week.

Three audits sit alongside the LLM scan. Each one runs on demand from the Brand workspace and grades your owned channels against a structured rubric:

Each audit gives you a per-category grade and a list of specific improvements ranked by likely impact on your AI Score. They are deliberately separate from the LLM scan because they give you the levers to pull, where the scan tells you what AI sees today.

Most reputation tools generate a list of generic best practices. We don't. A recommendation is only useful if it fits your brand, and we've spent real engineering time making sure we never tell an international NGO to "improve your product page" or recommend a SaaS company to "publish a quarterly white paper series".

• Every recommendation is scoped to your sector, your business model, and the kind of entity (brand, product, leader, campaign) it applies to. Wrong-fit suggestions are filtered out before they ever reach you.
• Each recommendation is generated, then validated by a separate pass that checks specificity and plausibility. Failed checks retry once, then suppress — you never see something the validator wouldn't sign off on.
• We measure what actually works. Every recommendation you act on is tracked against the next week's score. The aggregate impact is on your Reports page; patterns that consistently rate 'not helpful' across customers are retired automatically.

We are an LLM-visibility tool. We measure what AI models say in response to a defined set of prompts, with statistical rigour, and we attribute it. We do not measure these things, and any tool that says it does is selling you a number it cannot defend:

• We do not know exactly which prompt a real person typed into ChatGPT yesterday. We sample the space of plausible discovery, comparison, and opinion prompts your audience would ask; nobody outside OpenAI / Anthropic / Google has access to raw query logs.
• We do not measure outcomes downstream of AI search (revenue, donations, member signups, policy citations). We measure what AI is putting in front of your audience. The bot-tracking pixel bridges part of the way for site visits, but no tool can attribute a sale to a ChatGPT recommendation with full certainty.
• We do not predict future model behaviour. We tell you what models say today, with a confident range, and how that has shifted week-over-week. When a new model ships, we pick it up; we don't pretend to forecast what it will say before it ships.
• We do not enforce truth. If AI says something inaccurate about you, we measure it, attribute the likely source, and recommend a corrective action. We don't litigate the claim.

The honest framing matters because the rigorous framing matters. What we do measure, we measure well — and the Technical tab spells out exactly how.