Your rank tracker says you're position three. The buyer never sees position three — they read the AI answer at the top and stop. So the question that actually matters isn't "where do I rank?" It's "when someone asks the engine about my category, does my name come out of its mouth?" That's a different measurement, and almost nobody is taking it properly.
Most "AI-visibility" reporting is a screenshot of one ChatGPT session on one afternoon. That's not a measurement — it's an anecdote, and the engines are non-deterministic enough that the next run will disagree with it. If you're going to make budget decisions on this, you need the discipline a pollster brings to an election: a fixed question set, repeated sampling, and an honest margin of error. So I built that, ran it across a real B2B category, and I'm giving you the method and the scoring sheet.
01Why rank-tracking stopped working
Ranking and traffic used to move together. They don't anymore. Google's AI Overviews and the answer engines — ChatGPT, Perplexity, Gemini — increasingly resolve the question on the page, before any click. When the answer names three vendors and you're not one of them, your position-three ranking is worth almost nothing. The visibility that matters has moved from the blue links up into the answer itself.
And that visibility behaves nothing like rankings. It's non-deterministic (ask twice, get two answers), engine-specific (the tools cite largely different sources for the same question), and invisible to every SEO tool you already pay for. You can't manage what you're not measuring, and right now most teams aren't measuring this at all.
If you're briefing an AI-visibility decision off a single screenshot, you're not measuring — you're guessing with extra steps.
02How I ran it
The whole method is deliberately boring and reproducible. No proprietary black box — that's the point. Here's the protocol, exactly as I ran it:
- Fixed question set. [ 40 ] buying-intent prompts in a single B2B vertical — the questions a real buyer types when they're close to a decision ("best [category] tools for [use case]," "[vendor A] vs [vendor B]," "is [category] worth it for [segment]"). Written once, frozen, never edited mid-study.
- Multiple engines. Each prompt run against ChatGPT, Perplexity, and Google's AI Overviews — the three that a B2B buyer in this category actually uses.
- Repeated sampling. Every prompt run [ N ] times per engine, on separate sessions, to average out the non-determinism. One run is an anecdote; [ N ] runs is a distribution.
- Consistent scoring. For each answer I logged: which brands were named, which were cited with a link, the source domain behind each citation, and my client's presence or absence. Same rubric every time, logged to one sheet.
- Honest error bars. Because the sampling is repeated, I can report a share-of-voice with a confidence interval, not a single flattering number. When the interval is wide, I say so.
Total cost to run: a few hours of setup and a modest API bill. This is well within reach of any team that wants to stop guessing — which is why I'm handing you the sheet in section five.
03What came back
This section holds the shape of the finished post. The method above is real and fixed; the measured results below run before publication and drop into the bracketed slots. Placeholders are set in body type on purpose — on this site, fixed-width numerals are a promise that a figure is real, so nothing here wears mono until the study is run.
Three findings did the most work. First, the engines mostly don't agree with each other — the overlap between which sources ChatGPT, Perplexity and AI Overviews cite for the same question is small, which means "AI visibility" is really three separate visibility problems wearing a trench coat.
Second, the domains the engines lean on are not the domains you'd expect from the Google top ten — the "trusted source set" for a category is often disjoint from its search rankings. Third, my client's baseline presence was measurable, unflattering, and — crucially — fixable in a ranked order once you can see it.
04What it means for you
If you're a marketing director being asked whether AI search is worth the investment, this method gives you the one thing the vendor pitches can't: a defensible baseline number, with an error bar, that you can re-run next quarter to prove movement. That turns "we should probably do something about AI" into a measured program with a before, an after, and a scoreboard.
And it reframes the work. The goal isn't "rank higher" — it's to get into the trusted source set each engine draws from for your category, which is a content-and-credibility problem, not a keyword problem. Measuring it is step one; earning your way into it is the engagement.
05Run it yourself
I'd rather you measured this than took my word for it, so the scoring sheet is yours. It has the prompt-log structure, the per-answer rubric, and the formulas that turn repeated runs into a share-of-voice with a confidence interval. Bring [ 40 ] of your own buying-intent prompts and an afternoon.
No email wall, no gate — the sheet is right here. If you run it and the numbers scare you, that's a good reason for the next section — and, honestly, a good reason to start a conversation. This methodology is also the engine under the AI-visibility scorer I'm building; watch the Insights feed for it.
This is one vertical, one time window, and the answer engines change under you — a result from this quarter is not a law of nature, it's a reading you have to re-take. The sample is large enough to compare engines and rank opportunities honestly, not large enough to publish a universal "X% of B2B buyers…" headline, and I won't pretend otherwise. Non-determinism means even [ N ] runs leaves real intervals; where they're wide, treat the finding as directional. And presence in an answer is not the same as pipeline — being cited is necessary, not sufficient. The point of the method is a trustworthy, repeatable baseline you own, not a trophy number.