The 15-Minute AI Visibility Audit You Can Run Right Now

Most people check their AI visibility by opening ChatGPT, typing their company name, and reading whatever comes back. The link shows up and they relax. It does not show up and they worry. Either way the sample was one run, on a prompt no customer would ever type, from a system that would answer differently on the next attempt. You can do far better than that in about the same amount of time.

What follows is a timed audit. Set a fifteen-minute timer, work through the six steps in order, and you will finish with a filled-in scorecard that tells you where in the funnel your visibility breaks down. You need a browser, a notes file, and access to the four assistants. No tooling, no API keys, no spend. Keep a single scratch document open and log as you go, because the value is in the record, not in any one answer you happen to see.

Minutes 0 to 3: Write five to ten prompts a real buyer would type

Open your notes file and write the questions your customers actually ask an assistant when they have your problem and do not yet know your name. That last part matters. A prompt with your brand in it tests nothing, because you are handing the model the answer. The prompts that count are the ones where the buyer describes a situation and the assistant has to decide, unaided, whether you belong in the response.

Aim for five if you are moving fast, ten if you want a sturdier read. Spread them across the shapes a buyer uses. A category question ("best tools for tracking brand mentions in AI answers"). A comparison ("X versus Y for a small growth team"). A switching question ("we use a rank tracker, what should we add for AI search"). A problem-first question ("our product never comes up in ChatGPT, how do we find out why"). Pull the wording from support tickets or sales-call notes if you have them, because invented prompts drift toward the language your marketing already uses, and that is not how people talk to a chatbot. This small set is a first sketch of your Prompt-Space Coverage, the real map of questions customers ask, in place of a flat keyword list.

Minutes 3 to 9: Run each prompt across all four assistants and log the result

This is the core of the audit and it takes the longest, so keep it mechanical. Paste each prompt into ChatGPT, Claude, Gemini, and Perplexity in turn. For each answer, record three things: whether your brand appears at all, where it sits if it does (first named, in a list, a passing mention at the end), and which competitors got named instead.

A simple table in your notes keeps this honest:

prompt | engine     | named? | position        | competitors named
-------+------------+--------+-----------------+------------------
p1     | chatgpt    | no     | -               | Acme, Globex
p1     | claude     | yes    | 3rd in list     | Acme, Initech
p1     | gemini     | no     | -               | Acme
p1     | perplexity | yes    | first named     | Acme, Globex, Initech

Use the search-enabled mode where the assistant offers one, since that is the mode that retrieves live pages and shows sources, and it is closer to what a buyer gets. A brand named first in a direct answer is worth more than the same brand buried in a ten-item list, so the position column carries real weight: it separates the recommendation from the also-rans. One name keeps recurring in the competitors column for most teams. Write it down. That is the brand the models reach for when they reach past you.

Minutes 9 to 11: Check whose sources each answer stands on

Now look at what the answers cited. Perplexity shows numbered sources under every response, ChatGPT and Gemini expose links when search runs, and Claude lists the URLs its web search returned. Go back through the answers you just logged and, for each one that retrieved anything, note the domains behind it. You are sorting citations into two buckets: pages you control, and pages you do not.

The split tells you where your visibility actually lives. Third-party pages often account for a meaningful share of the sources behind brand mentions, so seeing competitor and review domains carry the citations is normal, not a failure. Verify the ratio in your own category rather than trusting a general figure. The full set of third-party pages where an assistant can find and cite you is your Citation Surface, and a quick scan of these sources shows how much of it you occupy. If every answer leans on a comparison site, a subreddit, and two review roundups, and you appear on none of them, you have found a concrete gap you can go fix this week.

Minutes 11 to 13: Run two or three prompts a second time to see the variance

Pick two or three of your prompts and run them again on the same engines. The results will move. A brand that appeared the first time may drop out, source lists reshuffle, ordering changes. This is expected, and watching it happen once cures the habit of trusting a single check.

The reason sits in how these systems retrieve. A retrieval-augmented answer is assembled in stages: the model rewrites your question into several sub-queries, pulls a candidate set of documents for each, a reranker trims that pool to a handful of passages, and only then does the model write prose and attach citations. Fan-out and reranking reroll on every call, so your page can land in the candidate set on one run and get cut on the next. When you see a brand flicker in and out across two runs, you are watching it sit on the boundary of the retrieved set. A brand that shows up every single time is genuinely well retrieved for that prompt. One that appears once in three is fragile, and averaging across repeats is the only way to tell those apart. Two or three re-runs are too few for a precise rate, though they will show you which of your wins are solid and which are luck.

Minutes 13 to 15: Confirm the assistants can actually reach your pages

The last two minutes check the foundation, because none of the above can work if the crawlers never see your content. Open yourdomain.com/robots.txt in a browser and read it for the AI user-agents. The real ones to look for are GPTBot, OAI-SearchBot, and ChatGPT-User from OpenAI; ClaudeBot and Claude-User from Anthropic; PerplexityBot and Perplexity-User from Perplexity; plus Google-Extended and CCBot. A Disallow rule can affect different products in different ways. OAI-SearchBot and PerplexityBot are search-index controls; training crawlers such as GPTBot and Google-Extended are not equivalent to search inclusion.

Then check whether your key pages render their content server-side. Open the page a buyer would need to land on, view source, and search the raw HTML for a sentence you can see on the screen. If the sentence is there in the delivered markup, a crawler can read it. If the source is a near-empty shell that fills in only after JavaScript runs, many crawlers get the empty version, and your best content is invisible to them regardless of how good it is. These two checks are the base of the AI Visibility Funnel, whose four stages run Retrievability, then Candidacy, then Citation, then Attribution. Fail the first stage and the other three cannot happen.

Tally the scorecard

Add up what you logged. Score each prompt-and-engine cell on a short rubric, then read the totals by stage.

• Retrievability (0 or 1). Crawlers allowed in robots.txt and your key page server-rendered scores 1. A block or a JavaScript-only shell scores 0. This is a per-site score, not per prompt.
• Candidacy (0 to 2). For each prompt, did your domain show up in the cited sources on at least one engine? Never, 0. On some engines, 1. On most or all, 2.
• Citation (0 to 2). When retrieved, were you actually named in the answer text? Rarely, 0. Sometimes, 1. Usually, 2.
• Attribution (0 to 2). When named, was it a strong placement (named first or in the top few) rather than a trailing aside? Weak, 0. Mixed, 1. Strong, 2.

Read the lowest stage first, because that is where your effort belongs. A zero on Retrievability makes every downstream number meaningless until you fix the block, so start there no matter what else the card says. Low Candidacy with the crawlers let in points at a Retrieval Gap: pages exist that could answer these prompts, but the system is not pulling yours into the candidate set, and the fix usually lives on the third-party pages of your Citation Surface rather than in your own title tags. Decent Candidacy with low Citation is a different problem, the one the Princeton GEO study addressed directly. In that research the changes that raised how often a source got cited were adding citations, adding statistics, adding direct quotations, and tightening the writing, with reported gains up to around forty percent on their visibility metric and the largest lifts going to lower-ranked sources. If you are retrieved but not named, that is the layer to work. Strong Citation with weak Attribution means you are in the answer but losing the top spot to a competitor, and the same content moves apply, aimed at earning the lead mention.

What fifteen minutes can and cannot tell you

Be honest about what you are holding. This audit is a snapshot of a handful of prompts on a single afternoon, and these systems change their retrieval behavior week to week without announcing it. Five prompts run twice is enough to locate your weakest funnel stage and to catch an own-goal like a blocked crawler. It is not enough to prove a trend, size a real appearance rate, or tell you whether last month's content push moved anything. For that you need the same measurement run on a wider prompt set, repeated on a schedule, tracked over time.

That is the continuous version of exactly this audit, and it is what shareof.ai↗ runs across ChatGPT, Claude, Gemini, and Perplexity so the numbers stay current instead of decaying the moment you close your notes file. The manual pass stands on its own regardless. Run it now, before you read one more article about AI search, and you will know more about where your brand actually stands than any amount of typing your own name into a chat box will ever tell you.

A note on measurement

The scorecard above is deliberately coarse. Its job is to point, not to grade, and a one-point gap between two stages is inside the noise of a fifteen-minute sample. Treat the ranking of stages as the signal and the exact numbers as rough. Where the audit sends you next, whether onto third-party pages, into your robots.txt, or back into the copy on a page that already gets retrieved, is the real output, and that direction holds up far better than any single score in the table.

Primary sources and evidence

• GEO: Generative Engine Optimization (Aggarwal et al.)↗ — the original paper and benchmark.
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks↗ — the foundational RAG paper.
• Web search in the OpenAI API↗ — current search-tool behavior and citations.
• OpenAI web crawlers↗ — official user-agent controls for search and training.
• OpenAI web crawlers↗ — GPTBot, OAI-SearchBot, and ChatGPT-User controls.
• Perplexity crawler documentation↗ — PerplexityBot and Perplexity-User guidance.
• Google common crawlers↗ — Googlebot and Google-Extended distinctions.

Continue through the system

• The AI Visibility Funnel: A Diagnostic for Where You're Losing Citations
• Reading Your AI Crawler Logs: GPTBot, ClaudeBot, PerplexityBot and Friends
• How to Choose an AI Visibility Tracking Tool (An Honest Buyer's Guide)

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.