AI visibility audit
An AI visibility audit is a structured check of how AI assistants such as ChatGPT, Claude and Perplexity answer the prompts buyers actually ask, testing whether a brand appears in those answers and whether the description of it is accurate.
The core of an audit is a prompt set that mirrors an actual buying journey. That means category prompts such as best accounting software for freelancers, comparison prompts that name competitors, and direct brand prompts such as is this product legitimate or what does it cost. Each prompt is run against the engines that matter for the market, and every answer is scored on two questions: does the brand appear at all, and is what the engine says about it factually correct.
A thorough audit separates answers from memory and answers from live retrieval, because they fail differently. An assistant answering from training data repeats whatever the web said months or years ago; one using live search depends on which pages it retrieves and cites. In our July 2026 study, 27 of 30 brands were materially misdescribed by an AI assistant answering from memory, and 13 of 30 were still misdescribed once live web search was turned on. Both modes need testing.
A good audit produces a source map alongside any score. For every prompt where the brand is missing or misdescribed, it records which pages the engine cited instead: often Reddit threads, G2 or Capterra listings, Wikipedia, or a competitor's comparison page. That list becomes the work queue. If Perplexity keeps citing a two-year-old Reddit thread with outdated pricing, the fix is a corrected, citable pricing page and a reply in that thread with the current numbers; the engine will keep retrieving that thread regardless of what the homepage says.
Audit results have a shelf life because engines retrain and retrieval indexes update, so plan to re-run the same prompt set monthly, or quarterly at minimum, rather than treating one pass as settled. A manual audit is workable at small scale: pick twenty buyer prompts, run each in every engine you care about, and log presence, accuracy, and citations in a spreadsheet. Free tooling exists as well: Discoverable's snapshot check needs no login, with Claude checks live today and other engines rolling out.