How to run an AI Search visibility audit in one day

At the start of every engagement we run the same audit before we touch a single page of content. It takes one day, it needs two people, and the result is regularly not what the client expects. Teams arrive convinced their content is too thin to compete, and the audit shows something else: for most B2B categories, the majority of sources AI engines cite sit on properties the brand does not own, on review platforms, YouTube channels, third-party lists and press coverage.

That finding changes what you do next. If the gap is off your site, another blog post is the wrong first move, and the right one might be a review-platform profile, a YouTube rework or a placement on the lists your buyers already read. If the gap is on your site, no amount of off-site work fixes it. You cannot know which world you are in until you read where the citations in your category actually come from.

This guide covers the one-day audit we run, step by step, with what each step costs in hours and which decision it feeds. You can run it without an agency and without a tool subscription. The prompt set and the reading are the work. The tool, if you use one later, is the easy part.

Key takeaways

• Classify where citations come from before you write anything. In most B2B categories we audit, the majority of cited sources are third-party pages, so the first fix is often a review profile or a YouTube rework.
• The audit is one day, six steps, each ending in a decision. Budget roughly 2 hours for the prompt set, 2 for running the engines, 1.5 for classifying sources and 2.5 for baselines and the fix list.
• Build the prompt set from sales calls, not keyword tools. Buying decisions happen in prompts with zero recorded search volume, so keyword-tool prompts measure a different market.
• Baseline the four attribution signals on the same day. Self-reported attribution, AI-referral sessions, branded-search lift and visibility give next quarter's numbers something to compare against.
• An audit that ends in a report was a day wasted. Every step feeds one decision, and the last step turns the findings into a ranked list where each item names an owner and a first action.

What an AI Search visibility audit is (and what it is not)

An AI Search visibility audit is a snapshot of three things: whether AI engines name you when buyers ask category questions, how they describe you when they do, and where the sources they cite actually come from. You run a fixed set of real buyer prompts across the engines your buyers use, you log what comes back, and you classify every cited URL. The output is a prioritized list of work, plus a baseline you can measure against later.

Three things it is not. It is not a rankings report, because a page can rank third and never get cited. It is not a share-of-voice screenshot, because a visibility number on its own tells you nothing about what to do on Monday. And it is not a tool score, because tools measure what you point them at, and most teams point them at the wrong prompts. Visibility numbers are signals along the way. The goal is conversions that turn into revenue, and every step of this audit exists to connect what you find to that goal.

One day is enough for the audit because the audit and ongoing tracking answer different questions. The audit tells you where to invest. Tracking tells you whether the investment worked. You do the audit once, deeply, then move to ongoing prompt tracking with a smaller standing prompt set. Teams that skip the audit and go straight to a tracking subscription end up with a dashboard whose numbers they cannot act on, because nobody ever classified where the citations land.

Before you start: what you need in the room

Two people are enough. One who talks to buyers, with access to sales-call notes, support tickets and reviews. One who can open Search Console and GA4 without help. A third person with an opinion about the category helps, but the audit does not block on it.

Set the scope honestly before you run anything:

• Engines: the ones your buyers actually use. For most B2B teams in our market that means ChatGPT, Perplexity, Gemini and Google AI Overviews. If your self-reported attribution says buyers arrive from one engine, that engine leads the list.
• Category: the buying category your buyers name on calls, in their words. "Marketing automation for mid-market" beats a generic "software" frame, because engines answer the specific question.
• Competitors: 3 to 5 names you actually lose deals to. Not the biggest brand in the space, the ones your sales team mentions when a deal slips away.

You do not need a paid tool for the audit itself. Manual runs across the engines work for one day, and logging in a spreadsheet is fine. A tracking tool like Peec AI, which we use for ongoing monitoring, pays off once you move from a one-day snapshot to daily reading (we are a Trusted Partner in Peec's directory, and that is a referral link).

Step 1: Build the prompt set from buyer language (~2 hours)

Start with the language your buyers already use. The prompt set decides what the whole audit measures, so this is the step where an extra half hour pays back the most.

What to pull

Collect 25 to 40 real questions from four places:

• Sales-call notes, especially the questions prospects ask in the last two calls before they sign or walk
• Support tickets, where the vocabulary is the customer's, not marketing's
• Reviews of your product, yours and your competitors'
• Non-brand queries in Search Console, the searches where buyers meet your category

You are looking for the questions a buyer asks in the week they are deciding, in the phrasing they actually use.

Decision-stage prompts first

Weight the set toward prompts that carry revenue: "best [category] for [use case]", "[option A] vs [option B]", "is [option] good for [scenario]". Add 2 to 3 fan-out variations per core intent, the sub-questions engines generate on their own when they answer the head question, because the answer your buyer reads is built from those. The full method for selecting and testing prompts is its own piece: how we build prompt sets for AI Search tracking.

The mistake most teams make

They build the set from a keyword tool. Prompts have no reliable search volume, so a keyword tool cannot tell you which prompts matter, and buying decisions happen in prompts with zero recorded volume. A keyword-tool prompt set measures a different market than the one your buyers are in.

The second version of the mistake is auditing on brand prompts. "What is [your company]" feels good to run and you will win it, and you will learn nothing, because nobody asking it is comparing you to anyone. Keep two or three for sentiment reading, no more.

The decision this step feeds

Which questions carry revenue, and therefore where the whole audit looks. Everything downstream inherits this scope, so it is worth getting the set right before you run anything.

Step 2: Run the prompts and log what the engines say (~2 hours)

With the set fixed, run it and read what comes back. This step produces your visibility baseline, and the honest version of it, because you built the prompts from buyer language rather than from whatever a tool had on hand.

What to log

Run every prompt across the engines you scoped, and log four things per answer:

• Whether your brand is named at all
• How it is described when it is named, in the engine's words
• Which URLs the engine cites as sources
• Which competitors appear instead of you

A shared spreadsheet with one row per prompt and engine works fine.

Two words, defined once

A brand mention is the engine naming you in the answer. A citation is the engine crediting a source URL in the answer. They are different events with different fixes: a mention problem is usually a presence problem, a citation problem is usually a source problem. Most of what this audit does with the two happens in the next step, so keep the words separate in your log from the first row.

Run each prompt more than once

Run every prompt 2 to 3 times per engine. Answers vary between runs, and a single run will have you chasing noise: a competitor appearing once in three runs is a different fact than a competitor appearing in all three. Log every run, then read the pattern.

While you log, note the sentiment of how you are described. "A small agency from Berlin" and "a leading provider" are both mentions, and they are not the same mention. Sentiment shapes whether a mention can convert.

The decision this step feeds

Where you stand today, prompt by prompt. This is the baseline every later measurement compares against, so date the sheet and store it.

Step 3: Classify every cited source: owned, earned, or competitor (~1.5 hours)

This is the step most teams skip, and the one that decides where the budget goes. Take every cited URL from Step 2 and put it in one of three buckets.

Three buckets

• Owned: pages on your own domains. Your site, your help center, your YouTube channel if you treat it as yours.
• Earned and third-party: review platforms, other people's YouTube channels, LinkedIn, Reddit, industry lists, press coverage, analyst pages. Sources the brand does not control.
• Competitor: pages owned by the competitors you scoped. Their sites, their profiles, their list entries.

Then count. The distribution across the three buckets is the finding.

What the counts usually show

In most B2B categories we work in, the majority of citations point to sources the brand does not own. We state that as our read from client work, not as a measured industry statistic, because the honest number is yours to measure in this step. When it holds, the implication is uncomfortable and useful: the fix is often a review-platform profile, a YouTube rework or a third-party list placement, and another blog post on your own site would have changed nothing.

The worked example we can put a name on

Heyflow had an existing YouTube library, and the reading showed YouTube was a top source in their category. After we reworked titles, descriptions and chapters of videos that already existed, 19 of them started getting cited regularly, with no new uploads, on a channel under 1,000 subscribers. Cited for exactly what it shows: citations follow structure and metadata, not views or subscriber counts. The full story is in the Heyflow case.

The decision this step feeds

Which surface gets the investment first. That is a finding, never a default: read which source types already win the citations in your category, then put the work where the citations already land. If the cited sources are mostly review platforms, start there. If they are YouTube and you have a library, the off-page playbook for AI visibility covers the levers per surface.

Step 4: Check whether AI crawlers can reach and parse your pages (~1 hour)

Step 3 tells you where to invest. This step checks whether anything technical blocks that investment. It is a blocker check and only a blocker check, because no schema trick fixes a citation gap, and anyone selling one lever as the answer is guessing.

Three checks, in order of how often they find something

1. robots.txt and CDN rules. Look for blocks on AI crawlers like GPTBot, PerplexityBot and Google-Extended. A WAF or CDN rule that quietly returns errors to unknown bots is the most common accidental block we find.
2. Server-side rendering. Open a key page the way a crawler sees it, with JavaScript off. If the answer-relevant content only exists after client-side rendering, some engines never see it.
3. Structured data spot-check. Only on the pages Step 2 showed engines already like. Valid markup helps an engine lift a clean answer; broken markup is noise.

Google's own AI features guidance states that no special markup, AI file or schema is required for AI features to consider your content. That is the primary-source version of our point: this step removes blockers, it does not create visibility.

The decision this step feeds

Whether anything is technically blocking the content work you are about to prioritize. Usually the answer is no, and that is a fine result. The step exists so you never spend a quarter building content an engine cannot read.

Step 5: Baseline the four attribution signals (~1 hour)

The audit is a snapshot, and a snapshot only pays off if the next measurement has something to compare against. Spend the hour capturing today's values for the four signals we read together.

The four signals to capture

• Self-reported attribution: does a "How did you hear about us?" field exist, is it mandatory, is it free text, and what have the last months of answers said? If it does not exist, that is the first fix to leave the audit with.
• AI-referral sessions: current GA4 sessions from chatgpt.com, perplexity.ai and similar sources. A floor, never the total.
• Branded-search lift: current clicks and impressions for your branded queries in Search Console. This demand signal tends to move a few weeks after AI visibility does.
• AI Search visibility: the prompt-by-prompt baseline from Step 2, dated and stored.

Why all four

Last-click attribution lies about AI discovery. A buyer who gets a recommendation in ChatGPT, hears your name, then types it into Google shows up as Direct or Organic. Read the four signals together and you get a causal chain you can defend to a board. The full setup, including the CRM field sales fills in after every call, is in our AI Search attribution guide.

The decision this step feeds

The measurement contract for the next quarter. Without these four numbers, the next audit cannot show movement, and the quarter in between cannot be defended.

Step 6: Turn the findings into a prioritized fix list (~1 hour)

Close the day with one list, ranked by expected impact, built from everything above. This list is what the audit exists to produce.

What goes on the list

For each prompt you lose: who wins it, which source types those winners are cited on, and the one or two moves that would put you where the citations already land. Then the prompts you win on sentiment and want to defend. Then the baseline fixes from Step 5, starting with the self-reported attribution field if it does not exist yet.

Every item names an owner and a first action

"Improve our review presence" is a finding. "Marketing lead: claim and fill the G2 profile, answer the six open reviews, by end of month" is a decision. An audit that ends in a report instead of a list of work was a day wasted, and the report will be correct and useless in equal measure.

Where judgment enters

The data says where the gaps are and which surfaces win the citations. What to do first, given your capacity, your budget and your sales cycle, is a senior read. It is fine for the fix list to carry a note on confidence: items straight from the citation data, items that are a judgment call, items that are experiments. We flag what is still an experiment in our own roadmaps, and the habit is worth copying.

The decision this step feeds

The next quarter's roadmap. Hand the list to whoever owns the quarter, and the audit is done.

What the audit usually finds (patterns from our client work)

Run this audit across enough B2B accounts and the same findings recur. Each of these is our read from client work, stated as a pattern, not as an industry statistic:

• The gap sits on third-party surfaces. Review platforms, YouTube and third-party lists carry the citations, and the brand has no deliberate presence on them.
• Brand prompts are fine, decision prompts are lost. Engines describe the company accurately when asked about it, and never mention it when asked to compare options.
• Whole product lines have no pages. Thin content is real. It just usually is not the first finding.
• Tracking cannot show whether past work paid off. No self-reported attribution field, no AI-referral visibility, so any future win would be invisible too.

The whole day in one view:

What following the list produces, on a real account: deeploi's audit pointed the work at the surfaces their buyers actually read, and AI Search visibility went from 0% to roughly 20% in four months, reaching the number two spot in their competitor set, while organic leads grew 3x over eight months. The details are in the deeploi case. Planeco Building is the same shape at a different scale: citation rate climbed from 55% to over 110% and lead growth reached 5x in about ten months, in that case.

If you take one thing from this: classify where the citations come from before you write anything. Every other step of the day serves that one read, and it is the read that decides whether the next quarter's budget goes into your site or somewhere you do not own.

Where Radyant comes in

Radyant runs this audit at the start of every engagement, and the free audit on radyant.io/audit is a compressed version of Steps 2 and 3: it checks whether the engines name you and where the cited sources come from. The methodology above is open on purpose, and if your audit shows the gap is on your own site and your team has the hours, you may not need us.

The case where teams do call us is the one this guide keeps pointing at: the gap sits across review platforms, YouTube and third-party lists, nobody internally owns those surfaces, and the fix list needs senior hands for a year, not an afternoon. That is the work we run as a Growth Partnership, from €7,000 per month on a 12-month commitment with an opt-out after six months, senior team, in your Slack, owning the result. If that is the shape of your list, talk to us.

FAQ