AI search visibility for B2B: measuring whether buyers find you

You can see every Google query that brought someone to your site. You cannot see a single conversation where an AI assistant recommended a competitor instead of you, and for a growing share of B2B buyers that conversation is where the shortlist gets written.

What is AI search visibility?

AI search visibility is the degree to which large language model answer engines mention, cite or recommend your company when someone asks a question your business could answer. It has three separate parts, and teams that blur them end up measuring the wrong one.

The first part is retrieval: does the engine fetch and read your pages at all. The second is selection: given everything it retrieved, does it name you in the answer. The third is action: does the buyer click, remember the name, or write it down. Retrieval is a technical problem. Selection is an editorial one. Action is mostly out of your hands. Ranking, in the classic sense of position one through ten, does not exist here at all.

Why Search Console cannot measure ChatGPT

Google Search Console reports on Google Search. It does not report on ChatGPT, Perplexity, Claude or Copilot, and Google folds AI Overview appearances into ordinary search performance data rather than breaking them out as their own line. So the dashboard most marketing teams open first is structurally blind to the thing they now want to know.

Server logs help a little. They show visits from AI crawlers, which confirms your pages are reachable and being read. They tell you nothing about whether any of that reading turned into a mention. A crawler hit is an input, not an outcome, and treating crawler volume as a visibility metric is the AI-era version of counting impressions.

The uncomfortable conclusion is that nobody hands you this data. If you want to know what the engines say about you, you have to ask them, on a schedule, and write down what they said.

The four layers worth measuring

No single number covers this. Four layers stacked together give a usable picture, and each one catches what the others miss.

• Prompt panel. A fixed set of buyer questions you run in every engine, every month, with the wording unchanged.
• Mention share. The percentage of those runs where your company appeared, tracked next to the competitors that appeared with you.
• AI referral traffic. Sessions in analytics whose referrer is an answer engine, judged by trend and behaviour rather than volume.
• Self-reported attribution. A free-text source field on your booking form, which catches the buyers who never clicked anything.

How do you run a prompt panel?

Write 20 to 40 prompts a real buyer would actually type. Spread them across five kinds: category questions ("what does a B2B lead generation agency do"), shortlist questions ("best outbound agencies in Europe"), problem-first questions ("our cold email stopped getting replies, who can fix it"), geographic questions ("lead generation agency Lithuania"), and brand questions ("is Ripe Leads any good"). The shortlist and problem-first prompts matter most, because that is where a name either enters the buyer's consideration set or does not.

Run each prompt in a fresh session, signed out where the engine allows it, so personalisation and chat history do not colour the result. For each run, record five fields: were you named, were you cited with a link, roughly where in the answer you appeared, which competitors appeared, and what the engine claimed about you. That last field is the one people skip and later wish they had kept, because a confidently wrong claim about your pricing or your service is a content bug you can go and fix.

Answer engines are non-deterministic. Two identical prompts, minutes apart, can produce different lists. Run each prompt at least twice and treat a single reading as noise. You need three or four monthly cycles before a movement means anything, which is exactly why monthly is the right cadence: weekly testing produces variance you will misread as progress.

Mention share: the one number to report

Roll the panel up into a single figure. If you ran 30 prompts across three engines and were named in 24 of those 90 runs, your mention share is 27%. That number is crude, and it is still the most honest summary available, because it moves when your citability improves and it does not move when you publish for the sake of publishing.

Track competitor mentions in the same sheet. Within two cycles you will have something more valuable than your own score: the actual consideration set the engines are building in your category, including firms your sales team has never heard of. That list is a better map of your competitive position than any analyst report, and it is free.

Tracking referral traffic from AI sources

Build one analytics segment covering the hosts that send AI referrals: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai. ChatGPT also tags many outbound links with a source parameter, so a filter on that parameter catches sessions the referrer alone would miss.

The volume will look disappointing next to organic search. That is expected and not a failure, because most answer sessions end without a click. What matters is the shape of the visit. Buyers arriving from an answer engine have usually had the basics explained to them already, so they land and move to pricing, service or contact pages far faster than a cold organic visitor. Compare pages-per-session and conversion rate for that segment against your search traffic before you judge the channel by its size.

Two caveats keep the numbers honest. Some assistants and in-app browsers strip referrer data, so those sessions land in direct. And traffic from a chat window is lumpy by nature, since one widely repeated prompt can send a small cluster of visits in a week and none the next.

Why zero-click discovery breaks classic attribution

Here is the sequence that quietly wrecks reporting. A buyer asks an assistant who solves their problem. The assistant names four companies including yours, with a sentence on each. The buyer reads it, closes the tab, and clicks nothing. Three weeks later they type your company name into Google, land on your site, and book a call.

Your analytics record that as branded organic search, or as direct. The conversation that created the demand receives no credit at all, and the channel that merely harvested it takes the win. Do this for a year and you will defund the thing that was working while congratulating yourself on the thing that was not.

There is no clean technical fix. There are two partial ones. Watch branded search volume as a proxy: if people are increasingly searching for your name without having clicked any of your ads or pages first, something upstream is introducing them. And ask buyers directly, which is the fourth layer.

What self-reported attribution actually catches

Put an open-text "How did you hear about us?" field on your booking form and leave it optional. Open text beats a dropdown here, because the answers you are hunting for are the ones you would never have thought to list. Then ask the same question on the call, where people give a longer version.

Watch for two patterns. "I asked ChatGPT and you came up" is direct evidence of AI-influenced discovery. "Someone recommended you" often turns out, on the call, to mean an assistant did. Self-reported data is messy, biased toward whatever the buyer remembers last, and worth collecting anyway, because it is the only signal that crosses the zero-click gap at all. Treat it as a directional vote, never as a precise percentage.

What you cannot control, and how to hold it

Be clear about the limits before you build a dashboard around this. Engines change retrieval behaviour without notice, and a model update can reshuffle who gets named overnight for reasons no vendor will explain. A citation this month guarantees nothing next month. Much of what feeds the answers sits on properties you do not own, including directories, comparison articles and forum threads, which is why getting cited by ChatGPT depends partly on third-party mentions you can influence but not edit.

So treat AI search visibility as a leading indicator you nudge rather than a channel you operate. The practical work sits in generative engine optimization: clear definitional answers, published numbers, consistent facts across your site, and FAQ markup that makes a passage easy to lift. The measurement work described here just tells you whether any of it landed. This shift is one of the few genuine changes in B2B marketing this year, and it is the one with the worst instrumentation.

The monthly review that takes an hour

Block sixty minutes at the start of each month. Run the prompt panel and log the results. Update mention share for you and your three closest competitors. Pull the AI referral segment and compare its conversion behaviour to organic. Read every self-reported source answer from the month. Then pick one thing: the weakest prompt category, the wrongest claim an engine made about you, or the page that keeps getting cited so you can strengthen it.

Do that for four months and you will have a trend line nobody else in your category has. Do it while also running a channel you fully control, because AI visibility compounds on its own timetable and pipeline does not wait. That is the honest case for pairing GEO work with cold email outreach: one builds discoverability you cannot schedule, the other books conversations on a calendar. Our pricing is published for the same reason we recommend publishing yours - engines quote pages that contain real numbers.