Founders keep asking us a simple question: “When a buyer asks ChatGPT for a tool like ours, do we come up?” Most have tried it once, seen a mixed answer, and been left unsure whether that one result means anything. It usually doesn’t. AI answers vary between runs, between assistants and between users, so a single screenshot tells you very little.
An AI visibility audit replaces that guesswork with a repeatable test. This guide shows you how to run one this week with a spreadsheet: how to design the prompt set, how to test without personalisation skewing the results, what to log, how to read the patterns, and where the measurement honestly runs out. If you’ve read our 30-day plan for getting recommended by ChatGPT, think of this as the detailed version of its first week.
What is an AI visibility audit?
An AI visibility audit is a structured check of how AI assistants mention, describe and cite your brand when buyers ask questions in your category. It covers ChatGPT, Perplexity, Gemini, and Google’s AI Overviews and AI Mode, and compares you against the competitors who show up instead.
It answers five questions:
- Are we mentioned for the questions that matter?
- Where do we appear in the answer, and how are we framed?
- Is what the assistant says about us accurate?
- Which sources are the answers built on?
- Which competitors are winning, and why?
An SEO audit looks at your pages and rankings. An AI visibility audit looks at the answer, which is assembled from many sources, most of which you don’t own. We explain that difference in more detail in GEO vs SEO.
What do you need before you start?
- A spreadsheet (Google Sheets or Excel)
- A list of three to five direct competitors
- Access to ChatGPT, Perplexity, Gemini and Google Search
- A clean browser profile used only for testing
- Two to four hours for the first run, less for repeats
Step 1: Design a prompt set that reflects real buying
The prompt set is the most important part of the audit. If it doesn’t reflect how buyers actually ask, the results will look precise and mean nothing.
Start from real language, not your tagline
The best prompts come from your buyers’ own words. Pull phrasing from sales call notes, demo request forms, support tickets, Reddit threads in your category, and the longer, question-style queries in Google Search Console. A buyer rarely asks for an “AI-powered revenue intelligence platform”. They ask how to stop deals slipping in the last week of the quarter.
Cover the whole buying journey
Aim for 30 to 50 prompts, grouped by stage. Tag each one so you can analyse groups later.
| Stage | Example prompt (for an onboarding tool) | What it tells you |
|---|---|---|
| Problem | “Why do trial users drop off after the first day?” | Whether you’re part of the conversation before buyers know solutions exist |
| Category | “Best user onboarding tools for B2B SaaS” | Whether you make the shortlist |
| Use case | “How can a 5-person team build product tours without developers?” | Whether you’re matched to specific jobs |
| Constraint | “Onboarding tool under $100 a month that integrates with HubSpot” | Whether your pricing and integrations are understood |
| Comparison | “[Competitor A] vs [Competitor B] for startups” | Whether you’re added to head-to-head decisions |
| Alternatives | “Alternatives to [Competitor A]” | Whether you’re seen as a substitute |
| Brand check | “What is [your brand] and who is it for?” | Whether the description is accurate |
Keep brand-check prompts in their own group. They measure accuracy, not discovery, so mixing them in inflates your mention rate.
Weight and freeze the set
Mark each prompt as high, medium or low priority based on how close it is to a buying decision. Then freeze the set and give it a version number. You can add prompts later, but compare like with like: if you change half the prompts every month, you can’t tell whether visibility changed or your test did.
Step 2: Run the tests without personalisation skew
Assistants increasingly personalise answers. That’s good for users and bad for measurement: you want to know what a typical buyer sees, not what an assistant has learned about you.
Here’s how to reduce the skew in each tool, based on their own documentation:
- ChatGPT. OpenAI’s Memory help page explains that memory can personalise responses and can even be used when rewriting web search queries. It also says you can start a temporary chat and choose Unpersonalized so existing memories and custom instructions aren’t used. ChatGPT’s search documentation adds that it may use an approximate location based on your IP address, so test from the market you sell to.
- Perplexity. Its Memory help page says memory and previous searches are always off in incognito mode.
- Gemini. Google’s Gemini Apps help says temporary chats don’t give personalised responses or save information for future personalisation.
- Google AI Overviews and AI Mode. Use a logged-out or private browser window and note the country you’re searching from. AI Overviews don’t appear for every query, so record when one doesn’t trigger.
Two more rules make results comparable. First, run each prompt as a fresh conversation, never as a follow-up, because earlier messages change the answer. Second, run important prompts more than once. Generative answers vary by design; the researchers behind the GEO paper sampled five responses per query in their experiments to reduce statistical deviation. For a founder audit, three runs per high-priority prompt is a practical compromise.
Record the assistant, the model or mode you used, and whether web search was on. A ChatGPT answer with search and one without can name different brands.
Step 3: Log what matters
Use one row per prompt, per assistant, per run. These columns are enough:
| Column | What to record |
|---|---|
| Date, assistant, mode | e.g. “ChatGPT, search on, temporary chat” |
| Prompt ID and stage | Links back to your frozen prompt set |
| Run number | 1, 2 or 3 |
| Mentioned? | Yes / No |
| Position | 1st, 2nd, 3rd… among brands named, or blank |
| Framing | How you’re described: “best for startups”, “budget option”, “enterprise only” |
| Accuracy | 2 = correct, 1 = partly wrong or vague, 0 = wrong |
| Errors | The exact wrong claim (old pricing, missing feature, wrong category) |
| Competitors named | In order |
| Cited sources | URLs, where the assistant shows them |
| Source type | Your site, competitor site, review platform, roundup, Reddit/forum, video, docs |
| Notes | Anything odd: refusals, hedging, AI Overview didn’t trigger |
Screenshot anything surprising. When someone asks “why did we drop?”, last month’s screenshot beats anyone’s memory.
Step 4: Turn the log into a few clear metrics
Raw rows are hard to act on. Summarise them into five numbers per assistant and per prompt stage:
- Mention rate: runs where you’re named ÷ total runs (excluding brand-check prompts)
- Share of voice: your mentions ÷ all brand mentions across you and your tracked competitors
- Average position: across runs where you’re named
- Accuracy score: average of your accuracy ratings on brand-check and mention runs
- Citation share: runs where one of your own pages is cited ÷ runs that show citations
Here’s an illustrative example, not client data. Say you ran 40 discovery prompts three times in ChatGPT, so 120 runs, and you were named in 30 of them. Your mention rate is 25%. If your four competitors were named 210 times between them, your share of voice is 30 ÷ 240, or 12.5%. Those two numbers, tracked monthly, tell you more than any single answer.
Step 5: Read the patterns
This is where the audit becomes useful. Look for patterns across groups of prompts, not individual answers.
| Pattern | What it usually means | Where to look first |
|---|---|---|
| Absent everywhere, including brand checks | Assistants don’t recognise you as an entity yet | Positioning, consistent descriptions across profiles, third-party mentions |
| Named in brand checks but not category prompts | Known, but not associated with the category | Category language on your site, inclusion in roundups and comparisons |
| Named, but described wrongly | Outdated or conflicting sources | The cited sources: directories, old reviews, stale threads |
| Present in ChatGPT, absent in AI Overviews (or the reverse) | Different assistants lean on different sources | Compare cited domains per assistant |
| Competitors cited from the same few roundups | Those pages shape the shortlist | Outreach to authors with accurate information |
| Reddit threads cited for your prompts | Community discussion is part of the answer | The threads themselves; see our Reddit marketing guide |
| Your page cited, but your brand not named | Your content is useful but not attributed | Make the brand and product explicit in the cited passage |
The differences between assistants are real. In Ahrefs’ analysis, only 14% of the top 50 mentioned sources were shared across ChatGPT, Perplexity and Google’s AI, which is why you log source types per assistant rather than assuming one fix works everywhere. A separate Ahrefs study of 75,000 brands found that branded web mentions correlate strongly with AI visibility, while the number of pages on a site barely correlates at all. As the authors note, correlation isn’t causation, but it’s a good reason to look outward at who is talking about you, not just at how much you publish.
If many problems trace back to a vague description of what you do, start with product positioning before anything else.
Step 6: Rule out technical blockers
Before you blame content, confirm the assistants can actually reach your pages.
- Google. Google’s guide to AI features says that to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to show with a snippet. There are no extra requirements or special schema.
- ChatGPT. OpenAI’s crawler documentation says sites that opt out of OAI-SearchBot won’t be shown in ChatGPT search answers. Check robots.txt and your CDN or firewall rules.
- Microsoft Copilot and Bing. Verify your site in Bing Webmaster Tools, which matters for the next section.
How should you measure AI visibility over time?
Run the full prompt set once a month, under the same conditions, and compare against your baseline. Monthly is frequent enough to see movement and infrequent enough to ignore day-to-day noise. In our consulting work, results typically start to show in around eight weeks, which is why we track from the very first week rather than waiting.
Each quarter, review the prompt set itself. Add prompts from new sales conversations, retire ones nobody asks, and bump the version number. Each month, pick the two or three biggest gaps and assign an owner: a page to fix, a listing to update, a roundup author to contact, a thread to answer.
Reporting works best when AI mentions sit alongside Reddit engagement, Google Analytics referrals from assistants, and Search Console performance.
What can’t an AI visibility audit tell you?
- There’s no Search Console for ChatGPT. You can’t see how many people asked a given question or saw your brand. Even Ahrefs notes in its AI assistant analysis that there is no reliable search volume for AI assistants.
- Google mixes AI features into Web data. AI Overviews and AI Mode traffic is reported inside the regular “Web” search type in Search Console. Google’s newer generative AI performance report shows impressions from AI Overviews and AI Mode by page, country, device and date, but it’s a filtered view of that same Web data, not a full picture of what was said.
- Position inside an AI Overview isn’t tracked. According to Search Console’s methodology, an AI Overview occupies a single position and every link in it gets that same position.
- Bing shows citations, not answers. Bing Webmaster Tools’ AI Performance report shows how often your pages are cited in Copilot and Bing’s AI summaries, plus a sample of grounding queries, but not placement or ranking.
- Referral traffic undercounts influence. Many buyers read an answer, then search your brand or type your URL directly. Add an “AI assistant” option to your “How did you hear about us?” field.
- Your sample is a sample. Fifty prompts can’t represent every buyer. Treat the audit as a consistent instrument for trends, not an exact count.
AI visibility audit checklist
- 30–50 prompts written from real buyer language, tagged by stage and priority
- Brand-check prompts kept separate
- Prompt set frozen and versioned
- Clean testing setup: temporary or incognito chats, logged-out Google, correct market
- Three runs per high-priority prompt, each in a fresh conversation
- Log completed: mention, position, framing, accuracy, competitors, cited sources
- Five metrics calculated per assistant and stage
- Patterns mapped to likely causes
- Technical access checked for Google, OpenAI and Bing
- Next audit date in the calendar
Frequently asked questions
How often should we run an AI visibility audit?
Monthly for the core prompt set, with a quarterly review of the prompts themselves. Running it weekly mostly measures noise.
How many prompts do we need?
For most startups, 30 to 50 is enough to see patterns without making the audit a chore. Smaller niches can start with 20; broader products may need more, grouped by use case.
Do we need a paid AI visibility tool?
Not to start. A spreadsheet works for the first few months and forces you to read the answers, which is where most insights come from. Paid tools help at hundreds of prompts or several markets.
Why do we get different answers every time we ask?
Generative answers vary between runs, and they can be personalised by memory, location and earlier messages. That’s why you control the setup, use fresh conversations and repeat important prompts.
How do we improve brand visibility in AI answer engines once the audit is done?
Fix the sources behind the patterns: clearer positioning on your own site, accurate listings, fair comparison pages, inclusion in the roundups assistants already cite, and honest participation in communities like Reddit. Then re-run the audit to see what moved.
Can we run the audit ourselves?
Yes, and we encourage it. Outside help is most useful when designing the prompt set, interpreting conflicting results, and deciding which fixes to do first.
Want an expert read on your AI visibility?
Running this audit yourself will teach you a lot about how assistants see your brand. If you’d like a second pair of eyes, our AI visibility consulting starts with exactly this kind of audit across ChatGPT, Perplexity, Gemini and Google, followed by a prioritised roadmap and monthly monitoring. A one-hour session starts at $349, including 30 minutes of preparation where we review your site and how AI assistants currently describe you, plus a written summary of next steps. If Reddit threads keep showing up in your results, our Reddit marketing consulting covers that side. We work fully remotely with mostly US-based SaaS teams, and our core service is advice and oversight; hands-on execution is available on request through partner agencies. Get in touch when you’re ready.