How we measure what AI says about your brand

AuthorityOn AI reports what the major language models (ChatGPT, Claude, Gemini, Mistral, Grok) say about your brand when someone asks about it. Every week we run a selected set of prompts, store every response word for word and analyse the responses.

We ask the models the kinds of question your audience asks them: discovery questions ("which companies do X?"), comparison questions ("how does Acme compare to Beta?") and reputation questions ("what do people say about Acme?"). We report on all three types. Discovery questions get equal weight with the rest, because appearing in an answer when your name was not in the question is how a brand reaches new audiences.

A growing share of people now ask an AI model before they search on Google. If AI does not mention you, gets your facts wrong or describes you negatively, that affects how your market sees you, and traditional SEO, PR and brand monitoring tools do not record it.

AI models draw on many signals when they decide what to say about a brand. Your AI Score combines two weighted parts: how AI models describe you today (60%) and the real-world signals that inform what they say (40%).

How AI models describe you. We measure this every week by running selected discovery, comparison and reputation prompts through every major model:

• Visibility. This measures how often your brand appears, and how high in the list, when someone asks AI an open question that does not name you but that your brand should answer. These discovery questions are weighted equally with how often you're recognised when named, so the headline figure reflects organic reach without penalising a brand AI already knows well.
• Sentiment. This measures whether AI describes you positively, neutrally or critically when it mentions you.
• Recommendation rate. This measures how often AI recommends you for a relevant need, as opposed to mentioning you in passing.
• Consistency. This measures whether answers agree across providers and prompt phrasings, or whether different models describe you differently.

Real-world authority. We measure this on demand by auditing the channels AI models read when they form a view of you:

• Website. We check schema markup, page structure and whether AI crawlers can read your pages.
• Content. We assess the quality of the writing on crawled pages and whether it is in a form AI models can quote.
• Social. We assess your LinkedIn presence, how often you post and the range of your thought leaders. These are third-party authority signals that AI models pick up.
• Earned footprint. We check Wikipedia and Wikidata presence, web-traffic rank and verified social reach. AI models cross-check these public signals.

Each signal is scored from 0 to 100 on its own. Missing data counts as 10, so an absent signal lowers the score rather than being ignored. A brand with strong visibility but no LinkedIn presence therefore cannot reach a high score on the LLM half alone. An improvement in any of these signals raises the overall score.

The score shows where you stand. The Source mix and Co-occurrence panels show where AI models get their information about you and which other names they group you with.

Source mix.We attribute every story AI models tell about your brand 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 class and the usual tone of each. If 70% of what AI says about you traces to news (mostly positive) and 20% to training data (mixed), recent PR is changing the answers now, and the training-data share changes only when new mentions enter the next training cycle.

Co-occurrence. We record every answer in which AI mentions you together with another entity (a competitor, peer, institution or person). The Co-occurrence panel lists the names that most often appear with yours, ranked by frequency, with their sentiment and a flag showing whether you already track each one as a competitor or peer. It shows which organisations AI models group you with, and which nearby brands you could partner with, approach for editorial coverage or position yourself against.

A scan is a sample of possible answers. If we ask a hundred questions and forty-one answers mention you, the true rate is around 41%, but it could plausibly be anywhere from 33% to 49%. Where a figure is a rate over countable answers, we show that uncertainty as a range calculated with a standard statistical method (the Wilson interval, described in the Technical tab). Where a figure rests on only a handful of answers, we show the count and the sample size instead of a percentage, because a count stays accurate at any sample size.

In practice, if you see "78, plausibly 75 to 81" one week and "76, plausibly 73 to 79" the next, the overlapping ranges show that the change is noise. If the following week drops to 64 and the range moves with it, the shift is real and worth investigating.

A score shows how strong your presence in AI answers is, but it does not explain why. Stories do that. We read every model response, extract the individual claims it makes 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.

The source class decides what you can do about a Story. A critical claim that comes from social posts needs a different response from one that comes from analyst reports or your own website. Stories record the claims, and Recommendations turn them into actions for the coming week.

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

Each audit gives you a grade per category and a list of specific improvements ranked by their likely effect on your AI Score. The audits are kept separate from the LLM scan because they show what you can change, while the scan shows what AI models say today.

A recommendation is only useful if it fits your brand. The system is built so that it does not tell an international NGO to "improve your product page" or advise 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. Suggestions that do not fit are filtered out before you see them.
• Each recommendation is generated and then checked by a separate pass for specificity and plausibility. A recommendation that fails is regenerated once, and if it fails again it is withheld, so you only see recommendations that passed the check.
• Every recommendation you act on is tracked against the next week's score. The combined effect appears on your Reports page. Types of recommendation that customers consistently rate 'not helpful' are retired automatically.

AuthorityOn AI measures what AI models say in response to a defined set of prompts, applies standard statistical methods to the counts and attributes each answer to a likely source. It does not measure the following, and no tool can measure them reliably:

• We do not know which prompt a real person typed into ChatGPT yesterday. Nobody outside OpenAI, Anthropic and Google has access to raw query logs. We sample the discovery, comparison and opinion prompts your audience is likely to ask, weighted toward discovery questions, which carry most of the visibility signal.
• We do not measure what happens after an AI search (revenue, donations, member sign-ups, policy citations). We measure what AI shows your audience. The bot-tracking pixel covers part of this 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 report what models say today, with a confident range, and how that has changed from week to week. When a new model is released we add it, but we do not forecast what it will say before release.
• We do not enforce accuracy. If AI says something inaccurate about you, we measure it, attribute the likely source and recommend a corrective action. We do not contest the claim ourselves.

The Technical tab sets out how each measurement is made.