How to prepare for an AI Visibility Audit: The Six-Step Process

Most teams approach AI visibility the same way: someone runs a few searches in ChatGPT, sees what comes back, and either relaxes or panics. That's not an audit — it's a sample of one. The results are unverifiable, non-repeatable, and impossible to act on with any precision.

A structured AI visibility audit follows a defined sequence — from understanding who your brand is and who your buyers are, to running systematic conversations and turning results into a prioritized build list. This post covers each step at a high level. Subsequent posts will go deeper on each one.

Step 1: Brand Identification

Before any questions are written or conversations run, you need a clear, documented picture of your brand — specifically, everything an AI model might associate with you.

What to do:

Identify the key pages on your website that best represent your brand: homepage, about page, core product or service pages, and any positioning pages that explain what you do and who you serve. These pages are your primary source material.
Feed those pages into an LLM with a structured prompt asking it to extract a brand profile. A complete brand profile should cover: your brand name and any aliases or shorthand your customers use; the problem you solve and who you solve it for; your core offering and how it works; the market context you operate in; your positioning and the tone you use; and any alternate websites or product properties associated with your company.
Document your competitive set. For each competitor, capture their primary name, any aliases they go by, and the domains associated with them. AI models reference competitors by multiple names — if you only track the primary name, you'll miss appearances under variants.
Prepare a list of target audience: This list should be quite exhaustive and contain anybody and everybody who can be a customer/user of your brand. This will be trimmed down in the next step.
Check for brand surface area issues: product names that differ from your company name, former names from a rebrand, abbreviations that show up in third-party content, or related properties that might cause a model to conflate different products.

Why it matters: AI models don't always refer to brands the way those brands refer to themselves. A model might surface you under a product name, a shorthand, or a variant — and without careful surface mapping, those appearances get missed entirely. Getting this step wrong means everything downstream is measuring the wrong thing.

Step 2: Persona Design

AI models respond differently to questions asked from different buyer contexts. The framing, vocabulary, and implicit context of a question shape the AI's response — including which brands it mentions, recommends, or ignores.

What to do:

Start with your existing ICP definition and map out the distinct buyer types who would use an AI assistant to research your category. Aim for 3–5 personas — enough to represent the real diversity of buyer behavior without overcomplicating the test. Use the target audience list from the previous step as the starting point.
For each persona, define: their role and seniority level, the company context they're operating in, the specific problem driving their AI search, their goals, their pain points, and the constraints shaping their decision.
Write out the natural language phrases each persona would actually use when asking an AI assistant about your category. This is not your marketing language — it's how a real buyer would describe their problem to a colleague or type it into a chat interface. The difference is significant.
For each persona, note the vocabulary they would and wouldn't use. A VP Marketing at a 500-person SaaS company uses different language than a content strategist at a 40-person startup, even when they're researching the same problem.
Assign priority to each persona based on buying authority, how actively they use AI assistants, and where you expect the largest visibility gaps.

Why it matters: A question written from the wrong persona frame produces AI responses that don't reflect what real buyers actually see. Persona design is what separates a question set that tests your actual buyer reality from one that tests a hypothetical buyer who already talks like your marketing team.

Step 3: Intent Mapping

A persona doesn't ask a single type of question. The same buyer, at different points in their decision process, asks very differently. Intent mapping defines the full space of questions worth testing before you write a single one.

What to do:

For each persona, list the different intent types that apply to them. The most common in B2B categories are: Diagnostic (figuring out a problem), Educational (learning how something works), Comparison (evaluating options), Transactional (ready to select a vendor), Tactical (looking for how-to guidance), and Brand-Aware (already knows your name and wants to understand your offering).
For each intent type, decide the question context — the frame the buyer brings to their question. Unbranded means they describe the problem with no category or vendor language at all. Category-led means they know the category name. Competitor-led means a specific competitor is already in their frame. Brand-led means they're asking directly about you. The same persona asking a diagnostic question will produce very different AI responses depending on which bias level they're operating at.
Not every intent applies to every persona. A content strategist persona may have strong educational and diagnostic intents but no transactional intent — they're not the one making the purchase decision. Be selective.
For each valid combination, note what the buyer is trying to accomplish and what a relevant AI response would look like. This guidance will directly shape the questions you write in the next step.

Why it matters: Without intent mapping, question sets tend to cluster around the middle of the buyer journey — category-led and comparison questions — and miss the unbranded diagnostic layer entirely. That layer is where the most significant visibility gaps typically live, because buyers haven't been handed a vendor frame yet and AI models are answering from pure content authority.

Step 4: Question Development

With your brand surface, personas, and intent map in place, you're ready to write the actual questions. This step takes longer than most people expect, and the quality of your questions directly determines the quality of your findings.

What to do:

For each intent and bias level combination you mapped, write a core question in the buyer's language. For unbranded scenarios, do not name your brand, your category, or any competitor in the question itself. The test is whether the AI surfaces you without being pointed toward you.
Write 2–3 variants of each core question — slightly different phrasings that a real buyer might use. This reduces the chance that a single phrasing produces an outlier response that skews your findings.
For each scenario, write a brief note on user context (what the buyer believes or has experienced coming into this question), the goal of the conversation (what they're trying to learn), and what a genuinely useful AI response would look like. This becomes your evaluation frame.
Define a stop condition for each scenario: at what point has the conversation achieved its purpose? This keeps multi-turn conversations focused and prevents them from drifting into unrelated territory.
Decide your brand injection policy per scenario: in most scenarios, you should not introduce your brand name into the conversation unless the AI mentions it first. Introducing it artificially inflates your prompted visibility and distorts the results.

Why it matters: The instinct when writing test questions is to phrase them in a way likely to surface your brand. That instinct produces results that look good and mean nothing. A question set written in genuine buyer language — the way someone would type a problem into a chat interface before they know your category exists — is harder to write but produces findings you can actually act on.

Step 5: Test

With your scenarios ready, you run them as conversations — not single queries — across multiple AI providers. Volume and consistency matter here more than most people anticipate.

What to do:

Run each scenario as a multi-turn conversation. Start with the core question, then follow up naturally based on the conversation goal you defined. A single query doesn't reflect how buyers actually use AI assistants; conversations do.
Test each provider separately: ChatGPT, Claude, and Gemini. Don't assume results are consistent across providers — the differences are real and diagnostically important.
For each conversation, record: whether your brand appeared, whether the appearance was organic (the AI surfaced you without prompting) or prompted (you asked about your brand directly), which competitors were mentioned and in what context, which sources were cited, and whether anything the AI said about your brand was inaccurate or misleading.
Run enough conversations to see patterns rather than outliers. A single conversation per scenario is a data point; twenty conversations across a scenario type is a pattern.

Why it matters: A brand might appear consistently in ChatGPT responses and be completely absent from Gemini — a different problem requiring a different fix. Running across providers and recording results systematically is what separates findings you can act on from impressions you can't.

Step 6: Analyze

Raw conversation data becomes useful only when it's structured into findings and connected to actions. Analysis is where the audit earns its value.

What to do:

Compute your core metrics: overall appearance rate across all conversations, the split between organic and prompted appearances, competitor displacement rate (how often a competitor appears when you don't), and brand citation share (what percentage of cited sources point to your domain).
Break these metrics down by persona and by provider. A metric at the aggregate level is interesting; a metric at the persona-and-provider level is actionable.
Map competitor displacement patterns: which competitors are appearing when you don't, and in which question types? This tells you where your content authority gaps are and which competitors have built it.
Identify the sources being cited in your category. These sources shape what AI models say about your space — understanding them tells you where third-party citation work would have the most impact.
Translate your findings into a prioritized action list. Each finding should map to a specific action: a content page to build, a positioning element to make more explicit, a third-party citation target to pursue. The analysis is only complete when it produces a clear answer to: what do we build next, and why?

Why it matters: The analysis step is where the audit becomes a build list rather than a report. Without it, you have a set of conversations that are interesting but directionless. With it, you have a precise picture of where you stand, who is winning the conversations you're losing, and what to do about it in the next 60–90 days.

What Comes Next

Each of these steps has real depth behind it. In subsequent posts, we'll go step by step: how to build a brand profile that captures everything an AI model needs to identify you accurately, what makes a persona test-ready versus too generic, how bias levels change the questions you write and the results you get, and what the analysis looks like when the data comes back.

If you'd rather see what your brand's starting point looks like before working through the full process yourself, request a diagnostic run.

AI Visibility Audit

Brand Tracking

By Gaurav

Published June 10, 2026