EN
Keep me posted
All agents

An agent at work · AI visibility analyst

What does AI answerwhen people ask about your market?

The AI visibility analyst is one of the agents in Zaela. It puts a hundred or more real questions to Claude, each in a fresh instance, and measures whether, where and how you are named. Then it looks at the sources the AI used and at what it sees of your website, and turns that into concrete tips.

The film

AI visibility analyst

Made in code with Zaela. The voice is synthetic.1:04

The method

This is how the AI visibility analyst works.

  1. 01Know the client

    First it learns what you offer, and where.

    It reads what your memory already holds: your offer, locations, competitors, keywords and earlier measurements. It takes an exact list of your locations or products, with name and place, from the online presence audit, or builds one from at least two sources. Together with a fact sheet that gives ranges where channels differ, that list is what every answer is scored against.

  2. 02The questions

    Questions the way people really type them.

    About a hundred neutral questions, more for a broad offer, fewer for a narrow niche. They cover the whole journey, several personas and the languages of your market, and never name you or a competitor. Some are asked open, for three names and for ten, and control cities where you have no location catch invented mentions. Ten to fifteen brand questions are kept apart.

  3. 03Clean measurement

    Every question in a fresh conversation.

    Each question goes to its own Claude instance with no history, no memory and no client context, like someone opening a new chat. Sonnet and Opus answer with web search, Sonnet also without, to show what sits in the model itself, and a set of core questions is repeated to measure stability. Before it measures, it estimates the cost and tells you.

  4. 04Scoring

    Every answer is scored twice.

    Named or not, in which position, by brand or only by the name of a location, recommended, neutral or negative, and with what sentiment. A word list and a separate Claude annotator score each answer independently. Where they disagree, it decides and writes down why. Every percentage comes with its sample size and a margin, and a split of fewer than 20 questions is only indicative.

  5. 05Sources

    Which websites the AI reads and cites.

    For every answer it logs which domains appeared in the search results, which pages were opened and which were cited. That shows your domain's rank among the sources, the type of each source, from competitor and comparison site to government, media or reviews, and which of your own pages turn up, old URLs and typos included. A link between sources and mentions is reported as a link, not as proof of cause.

  6. 06Website and tips

    What the AI sees of your site, and what to do about it.

    Fourteen fixed checks, among them your brand in Google and Bing, indexed test pages, URL structure, titles and snippets, core facts, languages, speed, AI crawlers in robots.txt and structured data. Every finding has evidence, a source and a date. Then come tips from P0, blockers to fix now, to P3, optional, each with what, why, effort, impact and what to follow up next time.

This page is based on the agent's instructions and knowledge in Zaela. The drawings are illustrations.

What it pays attention to

What it pays attention to.

  1. 01

    Clean, or not at all

    Questions are never answered by sub-agents in its own session, because those can see the client. If isolation fails, it does not measure and says so.

  2. 02

    Your name stays out

    Neutral questions never name you. Brand questions are reported separately and never count towards the mention rate.

  3. 03

    Every figure traceable

    Each result states the exact model ID, whether web search was on, the date, the country and the system prompt.

  4. 04

    Indicative, and honest about it

    AI answers differ from run to run. Figures come with a margin, and the trend across measurements matters more than one score.

  5. 05

    What the AI says

    Claims the AI makes about you are reported as what the AI says, not as fact. It checks first whether they are still current.

  6. 06

    Other assistants

    To measure ChatGPT, Gemini or Perplexity, you run the same agent inside that app, with the same questions and the same scoring.

Teamwork

Who it works with.

What it never does.

  • Let sub-agents answer the questions. They can see the client, so the measurement would no longer be clean.
  • Make up a figure. A call that keeps failing is reported as a failed measurement.
  • Work around a website it cannot fetch. It reports that and works with search results and logs instead.
  • Copy names or email addresses of employees from search results into its own texts.
  • Start a new paid step without reporting it first.

What you get

What you get.

Curious how visible your organisation is in AI answers?

Agents at work

All agents

Stay informed

Curious what Zaela could become for you?

Leave your details and we’ll get in touch to tell you more about Zaela.

We use your details only to contact you about Zaela.