How do you measure whether AI engines are recommending your B2B company?

Here's a ritual I watch founders perform, and I've seen a lot of it this month. They open ChatGPT, type "best [their category] for [their kind of buyer]," read the answer, and either feel great or feel a little sick. Then they close the tab. That's the whole program: a single check, on one engine, in a logged-in account that already knows exactly who they are.

And they'll tell you, straight-faced, that they're "tracking their AI visibility."

You measure whether AI engines recommend your B2B company by running a fixed set of 15 to 20 real, unbranded buyer questions across ChatGPT, Perplexity, and Google's AI answers in a clean session, then logging five numbers every month: named-or-not rate, share of voice, description accuracy, citation sources, and AI referral traffic. Any score that skips that discipline, including the polished ones on AI-visibility dashboards, is probably flattering you.

The instinct behind the ritual is right. If buyers ask ChatGPT, Claude, and Perplexity who to hire before they ever hit your website, you'd want to know whether the machines are naming you. You measure it the way you'd measure anything that matters to the business: a baseline, a fixed set of inputs, and a repeatable cadence.

Why isn't typing your name into ChatGPT actually measuring anything?

Call it the vanity spot-check. You type your own name into one engine, in your own account, once, and read whatever comes back. It feels like data, and four separate flaws sink it.

There's no baseline, so you can't tell better from worse next month. You ran one input, when your buyers ask dozens of different questions. The results are personalized to you, the logged-in founder who's visited your own site a thousand times, so the model is practically reading over your shoulder. And you asked about you instead of your category, which is the only question a real buyer types. Nobody in-market searches your company name. They ask for the best option for their problem and watch which names show up.

That last flaw is the whole game. Your buyer never types your name, which makes your name the one query that tells you almost nothing about whether you'll get recommended.

Why does measuring AI recommendation matter more now than a year ago?

Because the shortlist now forms inside the model, before a human visits a single site. AI brought the cost of content to zero, so every category is flooded with the same averaged-out claims, and buyers cope by asking a machine to pre-filter the field. G2's 2026 research found 51% of B2B software buyers now start their research with an AI chatbot more often than with Google, up from 29% a year earlier. The room where you get shortlisted moved. This is the surface AEO governs, and if the difference between AEO, GEO, and SEO is still fuzzy, start here. If you're not measuring what happens in that room, you're flying blind on the exact surface where deals now begin.

The answer also moves. It's different for every user, and it changes week to week as the engines re-crawl and re-rank their sources. Digitaloft found that 76.4% of the pages ChatGPT cites most were updated within the last 30 days, which means a spot-check you ran in April has already expired. A number you took once and never took again is a souvenir.

This is the same shift behind why AI recommends your competitors and not you. You can't fix the citation gap until you can see it, on a graph, moving or not moving. Measurement comes first.

What should you actually measure? Track these five things.

Forget your name. Build a set of 15 to 20 questions your best-fit buyer would really type, the "best [category] for [buyer]" and "who should I hire to fix [problem]" shapes. Run that same set across the engines, on a schedule, in a clean session. Then track these five numbers.

1. 1Named-or-not rate. Across your full prompt set, in what percentage of answers does your company get named at all? This is the single most important number. It's a clean 0-to-100 you can watch climb or stall. Everything else is detail underneath it.
2. 2Share of voice. When an answer lists vendors, how many of the slots are you versus your named competitors? Getting mentioned once in a list of eight is a long way from being one of the two names the model leads with. Count the slots.
3. 3Description accuracy. When you are named, does the engine describe you correctly? Or does it mislabel what you do? I've seen a model confidently file a positioning firm under "slide design." A wrong description is worse than a missing one, because it actively steers the right buyer away. Log exactly how you're characterized.
4. 4Citation source map. Note which third-party pages the engine pulls from when it builds the answer for your category. Review sites, comparison posts, other people's articles. Those are the sources you have to show up in, because brand is the new backlink: the model recommends the names other people describe clearly and consistently. When you do publish your own answer, structure it so it's actually quotable, which is its own discipline covered in how to write a blog post AI engines will quote. This metric tells you where the work is.
5. 5AI referral traffic. In GA4, segment referrals from chatgpt.com, perplexity.ai, gemini.google.com, and the rest. It's a small number today for almost everyone, but it's the closest thing to a direct-outcome metric, and its trend line is the one that eventually shows up in pipeline.

You can run the first three by hand in an incognito window for free, which is the honest floor every founder should hit before spending a dollar. Above that floor sit the dedicated AI-visibility trackers, Searchable, Profound, Otterly, Peec, and others, which run your prompt set across every engine automatically and chart the movement so you're not copy-pasting into a spreadsheet at midnight. The tool matters less than the discipline. A monthly spreadsheet beats an expensive dashboard nobody opens.

Why are so many AI visibility scores inflated?

A year ago nobody had a number for how often AI recommends them. Now every B2B team has a share-of-voice percentage, a little green arrow, and a board slide that says AI visibility is up and to the right. A lot of those numbers measure the wrong surface, run on the wrong prompts, and get filtered to look better than the truth.

A vanity AI-visibility metric is any "how often does AI mention us" score that looks impressive because of how it was measured, rather than because of what's happening in the answers real buyers see. It inflates through three moves.

1. 1Brand-seeded prompts. The tool asks the model questions that already contain your name. "What makes Acme a good messaging platform?" You'll show up almost every time. Of course you will. You handed the model the answer, which measures recall and tells you nothing about discovery.
2. 2Biased query sets. The prompts skew toward the categories you already win. Skip the ones you're weak on and the average climbs.
3. 3Selective filtering. The denominator shrinks. Prompts where you don't appear get tagged "not relevant" and drop out of the count. The percentage rises because the losses left the room.

None of this needs a shady vendor. It usually happens by accident, because whoever builds the prompt list wants to see the brand do well. The result is a green number that doesn't survive contact with a real buyer's question, the old vanity impression dressed up in AI clothes.

One score, two surfaces

Blending them is like a bank statement that adds your checking balance to your thirty-year mortgage payoff and hands you one figure. Technically math. Completely useless. Track them apart, or you'll credit a blog post for a shift it could never have caused.

Seven questions to pressure-test any AI visibility number

You don't need a tool to check a tool. Run your dashboard through these before it goes on a slide.

1. 1Which surface is this measuring, grounded or parametric? If the tool can't tell you, the number is noise.
2. 2Who wrote the prompts, you or the buyer? Prompts containing your brand name measure memory. You need to know whether the model chooses you.
3. 3Are branded and unbranded queries separated? "Who's the best messaging firm" is the one that turns into a sale.
4. 4Is the query set how your buyer actually asks? Pull ten questions from real sales calls and run those. If your score craters, your dashboard was curated.
5. 5What's the denominator, and is it stable? If prompts you lose get dropped as "not relevant," the percentage is inflated by subtraction.
6. 6Does the number move when you changed nothing? Ask the same model the same question three days running. If the answer swings, some of your "trend" is just model variance.
7. 7Can you read the receipts? If you can't see the actual answers and sources, you're trusting a score you can't audit. Read five real responses by hand.

What do the numbers look like across most B2B companies right now?

Grim, and for a reason most founders miss. The Walker Sands B2B AI Search Visibility Benchmark studied 828 enterprise B2B companies across more than 45 million queries and found the median company gets cited in just 3% of AI answers, even on the keywords where it already ranks well on Google. Ranking number one and getting named by the model are two different achievements, and most companies have exactly one of them.

That's the gap the spot-check hides. Your homepage can rank fine and your traffic can look healthy while you're nearly invisible in the answers your buyers read. Your name query is the one place you probably do show up. The blind spot is the category question, and it's where the deals are. This is also why ranking on Google doesn't get you cited in AI search: the two systems reward different things.

Across the 100+ B2B companies we've seen track this, a second gap shows up almost every time. Re-run a company's number with clean, unbranded, buyer-real prompts and the honest score comes in lower than the dashboard, sometimes by half.

The other common miss is surface confusion. A team ships a strong owned-content push, watches grounded visibility climb, and reports "we're winning AI," while parametric visibility hasn't moved a point. They're winning the surface that moves in weeks and losing the one that decides the hard recommendations.

What does this look like for a real company?

A $31M cybersecurity company came to us convinced they had an AI problem. They'd done the ritual, typed their name into ChatGPT, gotten a decent paragraph back, and couldn't square that with a pipeline that had gone quiet. The paragraph was fine. The measurement was the issue.

We built a 20-question prompt set from their real buyer language, ran it across three engines in clean sessions, and logged the five metrics. The baseline was brutal and clarifying: named in 3 of 20 category answers. Share of voice near zero when they did appear. And in two answers, described as a compliance-reporting tool, which is not what they sell. Meanwhile the engines were pulling their category answers from a handful of third-party roundups that named two competitors and never mentioned them.

Every bit of that traced back to entity consistency, the same problem I break down in why does AI recommend our competitors and not us. We locked their narrative to one clear, repeated description of who they're for and what they fix, then seeded that same answer across the surfaces the model was already reading. We re-ran the identical prompt set every month. About 90 days later they were named in 12 of 20, described correctly in all of them, and starting to appear in the roundups. They could see the jump because they'd measured the floor first.

Our own scoreboard, including the ugly number

PitchKitchen tracks its own AI visibility on both surfaces and reports both out loud. Over about six weeks, our grounded visibility moved from roughly 12 percent to about 19 percent, driven by owned content that finally tells one consistent story. Our parametric visibility is still close to zero, and we say that in the same breath as the good number.

Why volunteer the ugly half? A scoreboard that can't show you losing is a mirror you've decided to like. The honest read told us exactly where the next quarter of work goes: getting independent sources to describe us the same way our own pages already do.

What should you do this week?

Stop checking your name and start measuring your category. You don't need budget or a tool to begin, just a fixed set of questions asked the same way on a schedule. Here's the whole starter kit.

1. 1Write your prompt set. List 15 to 20 questions your best-fit buyer would really type, in their own words. "Best [category] for [kind of company]," "who fixes [the problem you solve]," "[competitor] alternatives." Keep your brand name out of every one. This list is the instrument. Save it, because you'll reuse it forever.
2. 2Take the baseline today. Open an incognito window, run every question across ChatGPT, Perplexity, and Google's AI answers, and log four things per question: are you named, who else is named, how you're described, and which sources it cited. That grid is your before picture.
3. 3Split the report by surface. Track grounded and parametric visibility separately and refuse to look at a blended number. They move on different clocks and need different work.
4. 4Set a cadence and commit to it. Monthly at a minimum, weekly if you're actively pushing on AEO. Re-run the exact same set and read the answers by hand, because week-to-week swings on unchanged content are noise. If the copy-paste gets old, a tracker like Searchable earns its keep.

Measurement tells you whether the consensus is forming. A clear narrative, documented once in a Magnetic Messaging Framework so it repeats everywhere without drifting, is what forms it, and it's how you get your B2B brand to show up in ChatGPT and Claude recommendations in the first place. Fix the story, and the numbers stop needing to be gamed.