How to Conduct an AI Visibility Audit, Step by Step

Use questions that describe a need or category, just as a customer would ask, without mentioning any brand. Something like "what tools exist for X" or "what is the best option for Y". A prompt that mentions a specific brand generates a response focused only on that brand, which does not help in mapping the complete scenario.

ChatGPT, Gemini, Claude, and Perplexity do not always respond the same way to the same prompt, and each decides differently when to search for updated information on the web instead of responding solely with what it learned during training. Testing just one model shows only a fraction of the scenario. It’s also worth repeating the same prompt multiple times: a study by SparkToro in partnership with Gumshoe.ai, involving 2,961 prompt executions in ChatGPT, Claude, and Google’s AI, found less than a 1 in 100 chance that two executions of the same prompt would yield exactly the same list of brands (SparkToro). A single execution does not reveal the real pattern.

Before making any content adjustments, check if AI bots can access your website. Open "yourdomain.com/robots.txt" in your browser and look for a "User-agent" line with the name of an AI crawler (GPTBot, ClaudeBot, PerplexityBot) just above "Disallow: /". If this appears, that bot is prohibited from reading the entire site, without any visible warning anywhere. It’s also worth checking if there is structured data of the type Organization on the site: it explicitly declares what the company is and helps compatible mechanisms recognize the brand as a specific entity, which is particularly important for generic names or names similar to another company (Google Search Central).

For each prompt where the brand did not appear, or appeared without being cited as a source, look at what is already published on that topic. Does the content answer the question directly, with concrete data, in a way that an AI could use as a reference, or does it remain in generic sales text, with nothing specific to cite? This difference often weighs more than the mere existence of the content.

Run the same prompts for both sides in the same round. A different prompt or round is not a fair comparison, because the difference that appears may come solely from the natural variation between executions, not from a real difference between the brands. Look at frequency (how many executions each brand appears in), position (the order when the response lists options), and whether the cited sources for both sides are the same domains or not.

A single round is just a snapshot. Repeating the same set of prompts periodically shows whether the brand's presence is improving, worsening, or stagnant, and this indicates whether it’s worth investigating before reacting to an isolated result.

These 6 steps describe the manual process. At MencionAI, it happens in an automated and continuous manner: the platform executes the registered prompts in new conversations, repeatedly, across multiple models at the same time, and records mentions, positions, present competitors, and cited sources in each execution.

To summarize all this in a single number, we calculate the ARS (AI Recommendation Score), which crosses mention frequency, position in the response, citation quality, and coverage among the tested models, from 0 to 100. The gaps found feed into the Action Plan, with a specific action for each gap, not just a general diagnosis.

If you want to run this audit for your brand, schedule a demo.

Pedro, CMO of MencionAI

How many prompts do I need for an audit to be worthwhile? There is no fixed number. What matters is covering the real questions a customer would ask about your category, without mentioning a brand, and repeating each one multiple times, as the same question can yield different answers in different executions.

Does a single audit solve everything? It provides a snapshot of the moment, not a trend. The greater value comes from repeating the same set of prompts periodically and comparing with previous results.

Do I need to test all existing AI models? At least the ones your customers are most likely using. ChatGPT, Gemini, and Claude cover most cases, and each can yield different results for the same prompt.

Does this replace continuous monitoring? A manual audit gives an initial snapshot. To track changes over time without repeating the manual process every week, this falls under continuous monitoring, which is what MencionAI automates.