AI visibility audit: how to measure where your brand stands

In short

An AI visibility audit measures how often and how well AI assistants mention your brand when a buyer asks a question in your category. You build a set of 20 to 40 buying prompts, run them through ChatGPT, Copilot, Gemini and Claude, and record Share of Voice, average position in the answer, whether your own domain or a third party gets cited, and whether the facts are correct.

You then check whether AI crawlers can reach your site at all: according to Otterly.ai’s AI Citations Report 2026, about 73 percent of sites carry technical barriers to AI crawlers (robots.txt blocks, CDN-level restrictions, JavaScript rendering problems). You also check whether the content can be extracted from plain HTML, and how wide your off-site footprint is. The output of the audit is a scorecard and a prioritized gap list, not a ranking promise.

Why you cannot manage what you do not measure

Most companies discover their AI visibility by accident. A sales rep hears on a call “ChatGPT recommended you”, or a competitor shows up in a Copilot answer for a query the company has ranked first on Google for years. Neither event tells you where you actually stand.

AI answers do not behave like search results. OpenAI has no rank-tracking API, Gemini has no Search Console, and the same prompt can return a different brand list tomorrow. According to the AirOps The 2026 State of AI Search report, prepared together with analyst Kevin Indig, only 30 percent of brands stay visible in the next answer to the same query, and after five repeated queries only 20 percent of brands keep their visibility. Without your own measurement, you are guessing, and every conversation about budget for generative engine optimization turns into a clash of opinions.

An audit gives you a dated baseline, a repeatable method and a short list of what is actually broken. Everything that happens next in a GEO program is measured against that point.

Step 1: build the prompt set

The prompt set is the heart of the audit. Get it wrong and every metric downstream is worthless. You need 20 to 40 prompts that mirror how real buyers talk to an assistant, spread across the funnel: problem prompts, where the buyer describes a symptom; category prompts (“who is good at X”); comparisons (“X vs Y”, “alternatives to X”); and a smaller share of brand prompts about you by name, to check factual accuracy and sentiment.

Write prompts in the buyer’s language and geography. Use natural phrasing, with typos and half-finished sentences, because that is how people actually type.

Keep the set in a spreadsheet with columns for stage, intent, market and language. You will re-run the same set every month, so freeze it. Add new prompts as a separate wave: editing old ones breaks month-to-month comparisons.

Step 2: choose the assistants and run the prompts

Test the assistants your buyers actually use:

• ChatGPT with search enabled: the largest audience, and the one where advertising already runs, so organic and paid visibility can be planned together.
• Microsoft Copilot: relies on Bing search data, which ChatGPT Search also uses as one of its sources alongside OpenAI’s own crawler. AI Max, an optional set of settings for Microsoft Advertising search campaigns that Microsoft has rolled out globally to all ad accounts, lets ads appear against more complex conversational queries in Bing and Copilot, though Microsoft itself does not claim advertisers see the full text of those queries.
• Google Gemini and AI Overviews / AI Mode: a different index, different citation habits.
• Claude: a smaller share, but common among technical and professional audiences.
• Perplexity: reference only. It transparently shows which sources it cites, which helps you understand the citation graph, but according to the Financial Times it closed advertising in February 2026, so it is not worth planning a budget around.

Rules for a manual run: a new chat for every prompt, because memory changes the answer; at least two runs per prompt on different days, both recorded, because the spread is itself a result; the full answer text saved, not just a “mentioned” flag; date, assistant, model name and web-search status logged. A set of 30 prompts across four assistants with two runs each produces 240 answers: a day of work for one person, or a tracker subscription.

Step 3: the four metrics

Share of Voice (SoV). The percentage of answers in which your brand appears, per assistant and per funnel stage. From the same answers, calculate SoV for your top three to five competitors: on its own this number means nothing, only relative to them. This is the metric that should sit in a contract instead of traffic, because ChatGPT often names a brand without a link.

Average position in the answer. When you are mentioned, are you the first brand named, somewhere in the middle, or at the end? Score 1 for first place, 2 for second or third, 3 for the rest. Lists in AI answers are short, and the first name gets most of the follow-up questions.

Citation source: own domain or third party. When the assistant cites a source for a claim about you, is it your own site, a review platform, a media article, a directory, or a competitor’s comparison page? According to AirOps, 85 percent of brand mentions in AI answers come from third-party domains, and only 13.2 percent from the brand’s own site, so this column shows which third-party pages work for you and which work against you.

Sentiment and factual accuracy. Is the price correct? The address? Does the assistant recommend a service you have discontinued? Tag each mention as accurate, outdated or wrong, and as positive, neutral or negative. Wrong facts can usually be traced back to a single stale third-party page.

Record all four metrics per assistant, then roll them up into a weighted total. Set the weights in proportion to your audience share, not evenly across assistants.

Step 4: crawler access and technical extractability

Before optimizing anything, check whether AI systems can read the site at all. About 73 percent of sites carry technical barriers to AI crawlers, including robots.txt blocks, CDN or WAF-level restrictions, and JavaScript rendering problems, often from a default security plugin setting or a copied template. It is the fastest fix in the entire audit.

Open yourdomain.com/robots.txt and look for these user agents:

If any of these sits under Disallow: /, or there is a blanket User-agent: * Disallow: / with no exceptions, the site is invisible to that assistant regardless of content quality. Also test the CDN or firewall layer with a request using the bot’s user agent: some providers block these bots at the network edge even when robots.txt allows them. Next, confirm the site is in the Bing index (site:yourdomain.com in Bing) and that Bing Webmaster Tools is set up with IndexNow enabled: this speeds up page discovery in Bing, one of the sources ChatGPT Search relies on alongside OpenAI’s own crawler.

Access opens the door; extractability decides whether the bot leaves with anything useful. On the pages that matter for buying prompts (home, services or products, pricing, about, top ten articles), check:

• HTML vs. JavaScript. View the source without JavaScript. If prices, the service list or the article text appear only after a script runs, most AI crawlers will not see them.
• Freshness. A visible, machine-readable, recent date. A 2023 date on a pricing page is a reason to skip it.
• Author and entity clarity. Who wrote it, what company this is, and the same spelling of the name everywhere.
• Structure. A direct answer near the top, tables for comparisons, definitions, numbers with a source. Structure and freshness weigh more than length.

Treat schema and llms.txt as hygiene only: according to Ahrefs (May 2026), which tracked 1,885 pages with added JSON-LD schema against 4,000 control pages, schema produced no statistically significant lift in citations on ChatGPT, Google AI Mode or Google AI Overviews.

Step 5: off-site footprint inventory

Since about 85 percent of AI visibility comes from other sites, the audit has to inventory them. Expand the citation-source column from Step 3:

• Directories and review platforms in your category: are you listed, is the profile complete, are the reviews recent, is the name and address consistent?
• Media: which publications mentioned you in the last 24 months, and do those articles still load?
• Comparison and “best of” round-ups: are you in them, and are the facts correct?
• YouTube: videos that name you, with the brand spelled correctly in titles and descriptions.
• Wikipedia and Wikidata: is there an entry, and is it accurate? Do not create one yourself; log it as a gap.
• Forums and communities, including Reddit: according to Peec AI, it is the most cited domain in AI search answers. Honest mentions only: Reddit actively removes spam and fake accounts, so astroturfing does not work.

For each source, record whether an assistant actually cited it in your prompt runs. Cited sources get fixed first. The logic behind that prioritization is covered in how ChatGPT chooses sources to cite.

Tools and the scorecard

The entire audit can run in a spreadsheet with a monthly manual pass. That is the right choice for one market, one language and up to 40 prompts. Columns: prompt, stage, assistant, date, mentioned, position, cited source, sentiment, accuracy, competitor mentions, notes. Once you cross two markets or need weekly data, a tracker pays for itself:

The prompt set and scoring rules must stay identical between the manual baseline and the tracker, otherwise the first tracked month will not compare to the audit.

The scorecard itself should fit on one page. Bracketed values below show the format, not benchmarks.

Below the table: the top five fixes in order of effort versus impact, each with an owner and a date.

What a $1,500–7,500 audit should include

Market prices for a one-off AI visibility audit run from $1,500 to $7,500. The spread depends on the number of markets, languages, assistants and the depth of the off-site inventory. Whatever the price, insist on these deliverables:

1. A frozen prompt set (20–40) tagged by stage and market, in an editable file.
2. Raw answer logs from at least four assistants, two runs each, dated.
3. Share of Voice for you and three to five competitors, by assistant and by stage.
4. Position, citation source, sentiment and accuracy for each mention.
5. A crawler access report per bot, including the CDN layer, and Bing index status.
6. An extractability check on key pages with specific fixes.
7. An off-site footprint inventory flagged with “cited by an assistant”.
8. A scorecard and a prioritized 90-day gap list.

What should not be there: a promised position, schema markup sold as a growth lever, or a traffic forecast from AI referrals.

At MOVE we run this audit from $1,500 for one market and one language, with the same prompt set and scoring rules that we later hand off to the tracker, so month one of ongoing tracking compares directly with the audit. We also open the client’s robots.txt and CDN settings on the first call, because that fix alone changes the picture for a meaningful share of sites.

Six mistakes that make an audit worthless

1. Testing only branded prompts. Asking “tell me about Brand A” will always make you look great. The money is in category prompts and comparisons.
2. One run per prompt. With persistence around 30 percent from one answer to the next, a single run is a coin toss.
3. Ignoring the CDN. robots.txt says allow, the firewall says block. Test with the bot’s user agent.
4. Counting links instead of mentions. ChatGPT names brands without links; referral traffic misses most of your visibility.
5. Skipping the off-site inventory. Fixing your own site touches about 15 percent of the problem.
6. Changing the prompt set every month. Freeze the baseline; add new waves separately.

Next step

If you need a dated baseline instead of anecdotes from calls, book an AI visibility audit with MOVE. You get the prompt set, the raw logs, the scorecard and a prioritized 90-day gap list, and the same measurement then carries into AI advertising and GEO work. Start on the contact page and tell us your market, language and top three competitors.