AI Visibility Audit: what AI tells your buyers about you

You start with a brand name and a website; a short description is optional. Then you pick the market the questions should come from (Indonesia, Malaysia, Vietnam or Global) and the language they're asked in. The language follows the market by default (Bahasa Indonesia, Bahasa Malaysia, Vietnamese or English) and you can change it, on creation or later.

From your website, Intura drafts a knowledge profile and the target audiences for the brand. Both live in a Brand context tab you can edit. Later steps check AI answers against this profile, so it is worth a few minutes of review before the first run.

The audit is a tree: cluster › topic › keyword › question. A cluster is exactly one target audience, one user intent and one funnel stage, named in plain words such as "first-time buyers comparing prices". Inside each cluster Intura derives the topics that audience asks about at that moment, and under each topic the keywords and names to look for in AI answers.

By default a project gets 6 clusters, up to 5 topics per cluster and 10 keywords per topic. You can change all three when you create the project and edit any cluster afterwards.

• User intents: informational, navigational, commercial and transactional.
• Funnel stages: discovery, consideration and decision.
• Objectives (AI Brand Authority, AI Search Visibility, AI Knowledge Optimization) steer which questions get written and work as filters. They are not scored.

One draft of "what would this audience ask" over-represents whatever came to mind first. So every batch of questions is written three times, independently, from the three angles listed below. The drafts are then grouped by meaning: two phrasings of the same question count as one.

Each group is scored on how many drafts converged on it (45%), how realistic it reads (35%) and how much demand it suggests (20%), with a small bonus when one of its phrasings came from real web content. Agreement gets the largest weight because it is the only signal that doesn't come from a single judgment. The top groups become your questions.

The default is 30 questions, with presets for 10, 50, 100 and 500. You can also add your own, one at a time or in bulk. Questions already in the project are skipped, so regenerating never creates duplicates.

• Chat: how the audience would type the question into an AI assistant.
• Web: real phrasings found online through web search.
• Pain: starting from the problem or trigger that sends someone looking.

Every question runs on every engine selected for the project. The same question can get a different answer on a second try, so you can run each question once or three times per engine for a steadier read.

Runs go through a queue (queued, running, completed) and are paced so the engines aren't flooded. A failed attempt retries automatically, up to three times. You can re-run one question, a selection or everything, and editing a question's text re-runs it.

• Engines: Gemini, DeepSeek, Claude, ChatGPT and Perplexity.
• Gemini answers with live Google Search grounding.
• DeepSeek answers from what the model already knows, without a live web search.

There is no single blended visibility score. Each answer is scored on the metrics that fit its question, because four numbers that each mean one thing are easier to act on than one that means everything.

Questions that don't name your brand ("best sunscreen for oily skin") are scored on mention rate and citation rate: does AI bring you up, and does it cite your site? Questions that do name you ("is Brand X sunscreen safe for kids?") are scored on citation rate, completeness and accuracy: does AI get you right? You can change the metrics for any question.

Completeness and accuracy are checked against a truth set built from your knowledge profile and a short crawl of your website. Any question can also carry its own expected answer.

• Mention rate: the share of answers that name you at least once.
• Citation rate: the share of answers whose cited sources include your own domain.
• Information completeness (0–100): the facts the answer should contain are listed, then scored on how many it covers, with half credit for partial ones.
• Information accuracy (0–100): every claim the answer makes about you is checked against your facts. Correct scores 100, outdated 50, wrong 0.

Every brand, product and person an answer names is extracted and then checked against the answer text itself. If the name isn't literally in the answer, it is dropped: the step that reads answers can invent names, and this guard keeps the list honest. Spellings are merged, so "Acme" and "acme " count once, and each name is ranked by how many answers mentioned it. Your brand is tagged You and the competitors you declared are tagged Competitor.

Cited sources are sorted into 12 categories: your site, competitor sites, social media, marketplace and e-commerce, news and media, forums and communities, reviews and directories, blogs and publishing, reference, docs and code, government and education, and other. You see each category's share and a ranked list of the sites AI cites most.

For engines that search the web before answering, Intura also records the searches the engine ran, shown as a word cloud and as full queries. Every question keeps its full answer, sources and searches per engine and per run.

AI answers vary, so a rate built on 40 answers is an estimate. Each metric shows a ± range at 95% confidence. When a day of results is compared with the previous one, the change is tested for significance and labelled as a real change or within noise. Anything built on fewer than 30 answers is flagged as a small sample.

Once a project has results on two or more days, each metric gets a day-by-day trend line with its confidence band. Every topic gets a verdict (Owned, Winnable, Crowded or Absent) with the rivals competing on it, and a separate list shows names AI talks about that you don't track yet.

Competitors show up through who AI names and the competitor-sites citation category rather than through a separate share-of-voice score.

All answers are combined into a ranked plan. Each action names a topic and its verdict, the page format to publish (a comparison, how-to, best-of shortlist, pricing page, guide and so on), the channel where you need to be cited and who holds those citations today, and the metric it should move. Priorities run from Critical to Low; a project shows up to 12 actions, or 5 per cluster.

You move actions through To do, In progress, Done or Dismissed. When the audit no longer sees a gap, its open action is resolved automatically; if the gap comes back, the action reopens. Anything you set yourself is left alone.