AI Visibility for B2B SaaS: How to Measure It, Grow It, and Prove It

Somewhere in your company right now there's a Slack thread with a screenshot of a ChatGPT answer in it. Someone asked the category question, your product either showed up or didn't, and the screenshot became the company's official position on AI visibility for the quarter.

That screenshot is close to worthless, and by the end of this guide you'll know exactly why, plus what to run instead.

Quick scoping note. This is the measurement half of a two-part series. If you want the full mechanics of how AI engines pick brands and the levers that earn citations, that lives in our AEO guide for B2B SaaS. This piece assumes you know what AEO is and answers the harder operator questions. What does "AI visibility" even mean as a metric, how do you measure it without fooling yourself, and how do you prove it to a CFO?

What "AI visibility" actually means

Three different things get called AI visibility, and teams that don't separate them end up optimizing the wrong one.

A mention is your brand named in the answer text a buyer reads. "For mid-market teams, look at Acme, Beta, and Gamma." That's the commercial event. The buyer saw your name.

A citation is your page linked as a source under or beside the answer. It means the engine read you and used you. It doesn't mean the buyer ever saw your name.

Those two come apart far more often than you'd expect. Kevin Indig's ghost-citation research with Semrush, run across 3,981 domains and four AI engines, found that 62% of AI citations never name the cited brand in the answer at all. Your comparison page can feed an answer that recommends your competitor. Indig's own verdict from that study is blunt. "AI mentions are far more important than citations."

Answer Share is the metric that makes the first two useful. It's the share of AI answers, across a defined set of buying prompts for your category, that mention or recommend your brand. Not one answer. The rate across many, tracked as a trend. Answer Share is what you put on a slide, because it's the only one of the three that behaves like a metric instead of an anecdote.

The mention outweighs the link because the answer is where the decision happens. Google ranks pages. AI recommends brands. The buyer takes the shortlist from the answer text, and Indig's broader research on shortlist behavior finds buyers pick the first brand named about 75% of the time. Position inside the answer is the new position one.

So the definition worth writing down for your team is this. AI visibility for B2B SaaS is your measured rate of being named and recommended across the prompts your buyers actually use, on the engines they actually use. Everything below is how to get that number without lying to yourself.

Why your single ChatGPT check is worthless

Two properties of these engines break every intuition you carry over from rank tracking.

First, answers are personalized. ChatGPT keeps a memory of who you are. Liam Dunne of Discovered Labs, whose team has done some of the only traffic-interception research on how ChatGPT builds answers, describes it plainly. "They build up this memory graph of you and they know, oh, okay, Liam is an agency founder doing this amount of revenue. He has these problems. This is where he's geographically based and that really influences the answer."

Google's version is documented at the patent level. Mike King's teardown of AI Mode surfaces a Google patent describing a persistent user embedding that shapes retrieval and synthesis for each person. In his words, "Two users asking the same query may see different citations or receive different answers, not because of ambiguity in the query, but because of who they are."

Now think about who runs the screenshot test at your company. A marketer whose account history is saturated with your brand, your category, your competitors. You're the single least representative test rig on earth for what a cold buyer sees.

Second, answers are non-deterministic. Ask the same engine the same prompt ten times and you'll get materially different answers, different shortlists, different sources. Dunne's rule for handling this is the operative one for any measurement program. "You want to log the distribution." One reading is noise. A hundred readings form a distribution, and the distribution has a trend you can trust.

There's a third problem hiding under both of these. Even a perfectly clean answer only shows you the visible layer of what influenced it. When Discovered Labs intercepted ChatGPT's actual retrieval traffic and analyzed 144,000 citations, Reddit made up 27% of what ChatGPT read for buying-type answers but appeared in just 0.35% of the citations shown to users. The sources shaping the recommendation and the sources displayed under it are different lists. A screenshot can't see the first list at all.

Put those together and the measurement rules write themselves: run many prompts from clean sessions, log them over time, and trust the trend instead of any single answer.

The measurement stack

So how do you actually instrument this channel? Four layers, ordered from cheapest to most involved. Ship them in this order.

Layer 1: survey attribution

The highest-return move in AI measurement is a form field. Add "How did you hear about us?" with an explicit AI option to every signup and demo form, and pipe the answer into your CRM as a property you can report on.

The reason this beats your analytics is measured. Graphite's attribution study at n8n compared GA4 against a post-signup survey. GA4 credited AI with 1% of conversions. The survey said 9%. And 90% of AI-driven conversions never clicked a citation, because the buyer read the answer, typed the brand into the address bar, and landed as direct traffic. The study also found organic and paid proportions matched between GA4 and the survey, which is what makes the survey's AI number credible rather than convenient.

Last-touch analytics doesn't undercount this channel a little. It files most of the channel under a different name. The survey is how you get the real denominator, and it costs one field.

Layer 2: prompt-set tracking across engines

This is the core discipline, and it's the direct answer to the screenshot problem.

Build a prompt set from first-party buyer language. Sales-call transcripts, onboarding calls, and win-loss notes are the source material. Prompts run longer and more contextual than keywords, and your buyers hand you their exact phrasing on every call. Dunne's practice, and ours, is a universe of 100 to 300 prompts grouped by topic cluster, because ten prompts is a vibe with a spreadsheet attached.

Then run the set across engines on a schedule. ChatGPT, Google's AI surfaces, Perplexity, and Gemini at minimum, from clean sessions with no account memory. A clean session means a fresh browser profile, no login, memory features off. It's an imperfect instrument, since IP and rough geography still personalize results, which is why serious tools run prompt sets from rotating clean environments instead of somebody's laptop. Log three things per run: mention rate, citation rate, and Answer Share against named competitors. Watch them weekly as trends with an honest error margin, and refuse to react to any single week.

This is the layer where your position in answers becomes a chart that moves, which is what makes the rest of the program manageable.

Layer 3: AI referral segmentation

In your analytics, break AI referrers out into their own channel group. Traffic from chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and friends should never sit inside generic referral traffic where nobody looks.

Treat this layer as a floor, not a total. The n8n finding above means clicks capture only the visible sliver of the channel. But the sliver is still useful, because its trend tracks the health of the whole, and AI-referred visitors are worth watching separately for conversion behavior. Pair it with branded search volume, since buyers who meet you in an answer often show up next as a branded Google search.

Layer 4: the new Search Console reports

As of June 2026, Google gives you ground truth for its own engines. Search Console's generative AI performance reports show how often your URLs appear in AI Overviews and AI Mode, broken down by page, country, device, and date.

Two caveats worth knowing before you present it. The report covers impressions but not clicks, CTR, or queries yet, so it shows visibility without traffic value. And Google confirmed these impressions were always inside your overall performance totals, so nothing about your existing charts changes. What you're getting is the split you never had. It's free, it's first-party, and it makes Google the easiest engine in your stack to measure. Baseline it now so you have history when the query data arrives.

A fair read on the tracking-tool category

Layer 2 is where most teams reach for a tool, and the category is real. Here's what it is and isn't, without the affiliate-listicle gloss.

What these products do at their core: run your prompt set across engines on a schedule from clean environments, log mentions, citations, sentiment, and sources, and benchmark you against competitors over time. That's genuinely valuable. It's the "log the distribution" discipline, productized, and it beats a spreadsheet and an intern at any real prompt volume.

A few names you'll evaluate, described from what each publicly claims. Profound positions at the enterprise end, with answer-engine insights across ChatGPT, Gemini, Claude, Perplexity, Copilot, and Google's AI surfaces, prompt volume data on what people ask AI, and agent analytics that track how AI crawlers read your site. Peec focuses on prompt-based tracking across ChatGPT, Perplexity, and Gemini, with competitor benchmarking, source analysis showing which sites AI answers pull from, and exports for your own reporting. Ahrefs Brand Radar brings AI mention tracking into the Ahrefs suite, covering AI Overviews, AI Mode, ChatGPT, Copilot, Gemini, Perplexity, and Grok against a large database of search-backed prompts, alongside YouTube and Reddit visibility.

All three are credible, and we're not ranking them. The right pick depends on your engines, your prompt volume, and whether you want a point tool or a suite. Capabilities in this category also change monthly, so check current pricing pages directly; some publish plans, some are sales-led quotes.

What to actually demand in an evaluation:

• Clean, unpersonalized sessions, with the vendor able to explain how

• Your prompts from your buyers' language, not only a synthetic library

• Distribution over time with competitor benchmarks, not single-answer snapshots

• Source-level data, so you can see which pages feed the answers you're losing

• Exports or an API, so the data reaches your reporting instead of living in one more dashboard

And one caution from inside the discipline. Dunne, who builds this kind of tooling himself, warns that "a lot of companies are just being fooled by randomness" by dashboards that report rate movements without uncertainty bounds. A 4-point Answer Share wiggle on 40 prompts is noise wearing a trend line. Whatever tool you buy, keep the statistical humility in-house.

The deeper limit is shared by every tool in the category, ours included. A tracker sees answers to prompts it asks. It can't see your buyers' actual conversations, their personalized answers, or the hidden retrieval layer. Tools give you the best available proxy for the distribution. Proxy is the honest word.

What moves the number once you can see it

Measurement without a lever is just a prettier way to watch yourself lose, so what actually moves Answer Share?

The full playbook, with effect sizes per lever, is in the AEO guide, and we won't rerun it here. The compressed version is three moves. Build citable pages, meaning direct answers early with real quotes, real numbers, and named sources. Those modifications are benchmarked at 25 to 28% more citation visibility each. Get talked about off your own site, because in Ahrefs' study of 75,000 brands web mentions predicted AI visibility about 3x better than backlinks did. And make the web agree about you. Engines cross-check your claims against third-party sources before naming you, and contradictions get you quietly dropped from answers.

The measurement stack tells you which of those to pull first. Low citation rate means a content problem. Cited but not mentioned means a positioning problem, since comparative and evaluative content is what gets brands named. Mentioned but described wrong means a consistency problem across your third-party surfaces. The diagnosis is the point of the dashboard.

The honest limits

A few things this whole discipline can't yet do, so you can present your numbers with the caveats attached instead of getting caught without them.

Tools test prompts. Buyers hold conversations. Real buying happens across multi-turn sessions with follow-ups, and sourcing behavior shifts as conversations deepen. Profound's analysis of ~730,000 ChatGPT conversations found the first turn is about 2.5x more likely to trigger citations than the tenth. Single-turn tracking oversamples the moment engines cite most and undersamples the long conversational middle where positioning gets tested.

Attribution stays fuzzy at the edges. Survey data is self-reported, buyers misremember, and some AI influence launders itself through branded search and dark social before it ever touches a form. You'll get a defensible estimate, never a clean number. Say so on the slide. It builds more trust than false precision.

The engines move under you. Models get updated, sourcing mixes shift, and a quarter of stable methodology isn't guaranteed. That's a reason for continuous tracking over quarterly audits. It's also a reason to report trends against competitors, which survive engine changes better than absolute rates do.

Answer Share is a leading indicator, and leading indicators don't pay salaries. The chain you're proving is visibility, then survey-attributed signups, then pipeline. If your reporting stops at the first link, you've built a nicer vanity metric. Every claim in the report should trace to a measurement someone can inspect. That standard is what separates a program from a subscription.

What should you do Monday morning?

Six moves, measurement only. The growth moves live in the other guide.

1. Add the attribution field. "How did you hear about us?" with an AI option, on every signup and demo form, synced to a CRM property. Two weeks from now you'll have your first real channel data.

2. Draft a 50-prompt starter set. Pull the language from your last 20 sales calls and win-loss notes, group prompts by topic cluster, and include competitor-comparison phrasings. Expand toward 100 to 300 over the quarter.

3. Segment AI referrals. Create the channel group for chatgpt.com, perplexity.ai, gemini.google.com, and copilot.microsoft.com today, and annotate the date so the trend has a start line.

4. Baseline Search Console. Open the generative AI performance report, export current impressions by page, and file it. Future you needs this number.

5. Start logging the distribution. Run the prompt set weekly across engines, by tool or by script, from clean sessions. Record mention rate, citation rate, and Answer Share per topic cluster. Commit in advance to ignoring any single week.

6. Build the one slide. Answer Share trend on the left, survey-attributed signups and pipeline on the right, caveats printed at the bottom. That slide is the program's contract with your CFO, and it's the artifact that keeps the budget alive.

Want the roadmap for your category? One call, and you leave with the map. The prompts worth winning, the pages to ship first, and how it all turns into pipeline. Get your roadmap.