How to run an AI visibility audit, with a real example
Published · Viscovery, Pepper Square
The short answer. An AI visibility audit checks four things: whether AI assistants can read your site, what they say when a buyer asks about your category, who they name instead of you, and what to change first. You can do the first three yourself with a browser. A full audit adds search data, a crawl of your competitors and a ranked list of fixes. Below are the five steps, then a real report taken apart section by section.
The five steps
1. Check that the assistants can read your site
Open yourdomain.com/robots.txt and look for rules against AI crawlers. They come in two kinds. Search crawlers, such as OAI-SearchBot, Claude-SearchBot and PerplexityBot, decide whether an assistant’s search can find and cite your pages; OpenAI, for one, says sites that block its search crawler “will not be shown in ChatGPT search answers”, though they can still appear as navigational links. Training crawlers, such as GPTBot and ClaudeBot, only decide whether future models learn from your pages. Then check for an /llms.txt file, and read your homepage title, first heading and first paragraph: do they say plainly what the business does? The free AI visibility checker runs these checks for you, and our methodology lists every crawler it reads.
2. Choose the questions buyers actually ask
An audit is only as good as its questions. Pick the searches that bring you customers: one or two the buyer types when they already know your category, and the broader ones they ask before they do. Write them the way a person asks an assistant, not as keyword fragments. Keep one branded question as a control, so you can tell “the assistants do not know us” apart from “the assistants know us but do not recommend us”.
3. Ask the assistants, and record exactly what they say
Put each question to ChatGPT, Claude, Gemini and Perplexity. For each answer, record one verdict: named, not named, no answer, or misattributed (named, but described wrongly). Note which pages each one cites, because an assistant can read your page and still credit someone else. Answers vary between runs and between users, so record the date and treat one run as a snapshot, not a ranking.
4. Find who is named instead, and why
The useful question is not “who beat us” but “what do they have that we do not”. Look at the pages the assistants cited and the pages ranking on Google for the same question. In the report below the reasons were structural: pages whose title was the exact question, roundups with a clear “best overall” pick, comparison tables an assistant could lift, and a forum thread with real buyers in it.
5. Turn it into a short, ranked list
An audit that ends in forty recommendations ends in none. Rank at most three actions, each with the demand or outcome that justifies its place, and write down one thing not to do first. Then turn each action into a page brief, put the briefs in order, and decide in advance how you will measure them.
A real audit, section by section
Every example below comes from a real Viscovery report: the Growing Brands audit of lululemon.com, dated August 13, 2026, for five keywords. You can read the whole thing at /sample. Every Viscovery report follows the same fixed section headings; which of them your report includes depends on the plan, as listed after the walkthrough.
Decision Brief: three moves, in order
The report opens with a decision, not a summary: at most three ranked actions, each with the number behind it, and one thing not to do first. For lululemon.com the first move was to turn the leggings category page into a buying guide (“1,900 searches/month; already ranks #6”), the second to build a running-apparel page (“880 searches/month; no result in the top 10”), and the third to give the yoga page quotable content assistants could cite by name.
Current Position: five scores and every keyword
Next come five 0–100 health scores (Search health, AI visibility, Page experience, Security posture, Crawl foundation) and a table with one row per keyword: where the site ranks, the demand, and a decision. The sample scored 71, 54, 55, 84 and 66. Its keyword table showed lululemon.com at #1 for “best workout clothes for women” (6,600 a month) and absent from the top 10 for two of the five terms.
Keyword Portfolio: where the next searches come from
On Growing Brands and Market Leaders, the portfolio puts every keyword’s position beside its demand and the page holding #1. In the sample, the page ranking first for running apparel was a Reddit thread, not a brand.
Evidence: what the assistants actually said
Evidence is a table of real questions and each assistant’s verdict. It made the sample’s central finding plain. Asked “What’s the best apparel to wear for yoga?”, none of the four assistants named Lululemon, although Gemini and Perplexity both cited a Lululemon product page among their sources. Asked “What are the best leggings for working out?”, all four named a Lululemon style. The site was being read; on the broad question it was not being credited.
Competitive Landscape: who wins instead
This section ranks the domains that appear across your keywords. In the sample, Reddit appeared for four of the five, at an average position of 1.75, more often than any competing brand.
What to Build, the roadmap and measurement
The last sections turn the findings into work: one brief per page with its purpose and five or six specific changes, a 90-day roadmap that orders them by dependency and payoff, and the measures to check afterwards. The sample’s first brief listed six changes to the leggings page, among them activity-by-activity picks, a comparison table and an FAQ drawn from real buyer questions.
What each plan’s report contains
- Starter, $49, 1 keyword: Decision Brief, Current Position, Evidence, What to Build, 90-Day Roadmap and Measurement.
- Growing Brands, $99, 5 keywords: all of the above, plus Keyword Portfolio and Competitive Landscape.
- Market Leaders, $149, 7 keywords: all of the above, plus Content Architecture and Execution Plan.
Each report asks ChatGPT, Claude, Gemini and Perplexity about the plan’s top 1 to 3 keywords and ChatGPT alone about the rest. It is a one-time report, paid once; it does not track mentions over time.
Do it yourself, or buy one?
Steps one to three need only a browser, and the free checker covers step one in seconds. Steps four and five are where most of the time goes, because they need search data, a crawl of the pages that beat you and the judgement to rank what you find. If you are deciding who should do that work, we compare the options in agency or self-serve audit, and the tools in the best LLM visibility tools, compared on price.