How I Run an AI Visibility Audit

Yes and no, depending on the question

“Are we visible in AI?” It’s usually the first thing a client looking for help with their organic visibility asks me, and the honest answer is that it depends on the question.

AI assistants don’t keep a fixed list of firms. Citations are volatile, and AI is known to build a new answer each time, shaped by how the buyer phrases the request. Ask for “the best marketing agencies” and you get one set of names. Ask for an agency that specializes in what you do, for the kind of client you serve, and you can get a very different set.

A firm can be missing from the first answer and near the top of the second, or the reverse. So “are we visible?” doesn’t have one answer. It has one answer per question.

That’s why, when I run my diagnostic service, I don’t start an AI visibility audit with a tool or a score. I start with the questions my client’s buyers actually ask.

What an AI visibility audit measures

Most people want a rank. A number. Like, where do we show up in ChatGPT? How often? In other words, they want something akin to an SEO metric, and I get why. After all, it’s what they’re used to, and they want something they can measure and bet on.

But AI answers aren’t a ranking page. SparkToro had 600 volunteers run the same prompts 2,961 times, and the odds of getting the same list of brands twice were under 1 in 100 (SparkToro, January 2026). A single position number is a snapshot of a coin toss.

So my audit asks three questions instead. Do you show up at all? Does the engine name you as the answer, and whose pages does it point to? And does any of it reach the pipeline? Those map to the three P’s in my Triple-P Model. Presence, Prominence, and Performance.

Step one: write the buyer’s questions

I write a fixed set of questions in the buyer’s words, and I don’t edit them once the audit starts. If the questions drift, the trend means nothing.

I borrow a modified version of the structure used in Foundation’s Ghost Citation audit. It covers the stages a buyer moves through. Discovery (who should I consider?), comparison (why this firm over that one?), implementation (what does working with them look like?), risk (what could go wrong?), and trust (can I bet my job on this?).

What I do is, I add two twists of my own. I write each question twice, once generic and once naming the firm’s specialty. And I run the geography that matters, because an answer for one country can look nothing like the answer for another — let alone a city, in the case of a local business.

I saw why this matters in a recent audit for a marketing agency. Generic questions buried them. Specialty questions surfaced them. Without both versions, I’d have reported the wrong conclusion either way.

Step two: run them across engines, more than once

I run every question on ChatGPT, Perplexity, and Claude, plus Gemini and Google’s AI Mode where they matter. Web search stays on, because that’s how buyers use them now.

Then I run the set again. One pass tells me almost nothing. SparkToro’s own advice is to ask the same question 60 to 100 times before trusting the pattern. For a client audit, I run the core set more than once and watch what repeats.

Each engine also reads the web differently, and that changes the fix. In this audit, Gemini went looking for the agency by name to verify it. ChatGPT didn’t search for them by name in any of my tests. It leaned on rankings and directories instead.

So one engine rewards a consistent, findable set of facts about you. Another rewards being on the lists it trusts. Which one do you optimize for? You need both.

Step three: separate named from sourced

This is the step most AI visibility reports skip.

Being named in an answer and being cited as a source are two different things. The engine names a firm, then cites sources to justify what it said, and those sources often aren’t yours. Foundation found that n8n showed up in 91 of 96 answers where a brand was recommended, yet its own properties earned only 15.7 percent of the citations behind them. Reddit was cited 95 times. The n8n blog, 33.

Both sit in Prominence, at two depths. Did the engine name you as the answer, and whose pages sit next to your name? If the answer is a forum thread (like Reddit) or a competitor’s list, you’re renting your visibility, not owning it.

One caveat I give every client. Rand Fishkin points out that no study has confirmed AI models build their answers from the citations they show, rather than attaching them afterward (Fishkin, LinkedIn, September 2026). So I read citations as a map of who the machine points to, not proof of what it read. It’s still the best signal we have.

Step four: check the four doors

In my experience, every firm an assistant names got there through at least one of four doors. I check all four, because each one is a different kind of work, and most firms have only opened one or two.

The doors are based on and line up with the Four Fs I use to build brand memory. They just run in the opposite direction, from the outside in.

That order is deliberate. Traditional SEO tends to work from the inside out. You fix the site first, then earn links and mentions to it. AEO, or SEO for AI, flips the priority. Both still matter. But when an assistant builds an answer, what others say about you tends to carry it, and your own site confirms it.

Foundation’s n8n audit points the same way, with most of the citations behind n8n’s recommendations coming from sources n8n didn’t own. Rand Fishkin lands in the same place. His advice is to “invest in and measure mentions, watch incremental lift, ignore the rest.”

It’s a question of priority, not importance. If the SEO foundation is solid, I start outside, because that’s where the assistants seem to look first. As SEO expert Aleyda Solis recently put it, “AI search is an off-site corroboration problem, with an on-site quality floor.”

Door 1: the sources others publish

This is Footprint. It’s what the rest of the web says about you, and it’s where assistants go for third-party validation. It’s about brand mention and repetition.

Directories and “best of” lists are the obvious ones. But it also covers podcasts, YouTube videos, Reddit threads, Substacks, guest articles, speaking at events, and contributions to industry or peer-reviewed publications. Anything you don’t own that carries your name.

In this audit, the agency was on one directory and none of the lists that kept showing up in the answers. And a lot of its public story lived on LinkedIn, which blocks the major AI crawlers in its robots.txt, including GPTBot, ClaudeBot, and PerplexityBot. Well known to people, invisible to assistants.

Door 2: the sources you publish

This is Frameworks. It’s the first-party material only you can produce, and it gives a model something specific to cite. Is your content unique, helpful, and high quality? Then chances are it will be citable, too.

Named methods, thought leadership, original research, unique data, studies you’ve run, even your own curated lists. Some firms publish their own “best of” list, include themselves, and get cited from it. It works. If you do it, be fair to the competitors you list, because buyers read it too.

What I look for is proof the firm owns. The agency’s directors told me clients value their reporting and share-of-voice analysis. None of it was on the site.

Door 3: the signals that confirm the sources

This is Fidelity. This also uses my SOME technique applied to the page. It’s less about content and more about structure, metadata, and the signals a machine uses to check who you are.

The sources here are the legitimacy of those mentioned in Doors 1 and 2. That is, are the facts about the firm, the people behind it, the brand, and the data correct, corroborated, and consistent? When those facts hold up everywhere, a model can trust what Doors 1 and 2 say about you.

Consistent name, address, and phone across every profile. Schema that states the organization and the people behind it. Author bios with real credentials. Clear E-E-A-T signals. Content organized so a model can lift a section without the rest of the page. The traditional SEO basics still sit here too.

A lot of the search traffic to a firm’s site can be machines checking basic facts. Who runs the company, when it was founded, how many people work there. Those are the facts a model checks before it describes you, and many sites don’t answer them anywhere.

Fidelity also means old versions stop contradicting new ones. I found that on my own site. A retired page with my framework’s old pillar names was still being quoted by two of four engines after I’d redirected it.

Door 4: the topic you cover completely

This is Fan-out. When a buyer asks one question, the assistant quietly runs several searches behind it. The firm that answers all of them, fully and logically, is the one it cites.

This door opens with the rule of one. One message, one market, one move. A page built for one kind of buyer with one kind of problem, then covered so completely that the follow-up questions are already answered. And it asks the reader to make one move, with one clear outcome for making it.

If Door 2 is about content and Door 3 is about structure, Door 4 is about the topic. Comprehensiveness and quality carry the weight here.

The agency had one page like this, built around its specialty. It was carrying almost all of their category answers. The problem is that it also had almost no internal links pointing to it. The assistants found it anyway because it was listed in its sitemaps, but buyers browsing the site couldn’t — and when they landed, they didn’t know where to go next.

That surprised me most. The page doing the selling was invisible on the site itself.

Step five: report the direction

I don’t hand a client a rank. I give them a baseline and a direction.

The baseline has every number dated and sourced, so the next quarter has something solid to compare against. For the agency (the client mentioned throughout this article), that meant naming the numbers that looked better than they were, and leaving them out of the scorecard.

Then I report three columns. Presence asks whether they showed up at all, and for which claims. Prominence asks whose pages the engine cited, and whether it named them as the recommendation. Performance asks what reached the pipeline, which is the hardest column to fill honestly, because buyers touch a firm many times before they buy, between 6 and 60 touchpoints by Rand Fishkin’s count, and AI is often one in the middle.

What matters is movement. Direction, not a decimal. The trend, whether big or small, upwards or down, is more important than a number. Named in three of four engines this month and four of four next month is progress. A number that wobbles from run to run is the machine, not the marketing.

What comes out of it

The audit ends with a diagnosis and a prioritized roadmap. For the agency in question, it came down to one sentence. They were getting found for a position they had never stated, and not found for the one they stated or wanted.

The fixes followed from the doors. Fix the tracking so the next quarter can be measured. Connect the pages that already carry the position. Rewrite the core pages so the category and proof sit up front, using my SOME technique so a model can lift them. Put the proof where people and assistants look. Then feed it with content built for both.

That’s the AI side of The Gauge, my diagnostic engagement. Sometimes it’s the whole engagement. More often it sits inside a wider look at organic visibility across search, AI, and social, because buyers don’t use one channel and neither do the machines.

If you’ve asked an assistant about your own firm lately, try it both ways. Name your specialty, then leave it out. The gap between those two answers is usually where the work is.

---

Frequently Asked Questions