Back to Blogs
Drew Neisser is the founder of CMO Huddles and a globally recognized authority on B2B marketing. He’s an AdAge columnist, LinkedIn TopVoice, leading CMO coach, podcast host & friend of penguins everywhere.
B2B CMOs cannot prove AI-search ROI by counting citations or referral clicks alone. Drawing on four CMO Super Huddle workshops led by Kishan Panpalia of Pepper, this practical guide explains how to audit prompt quality, validate monitoring-platform data, track AI-influenced demand, and run a 90-day topic holdout test.
AI visibility is becoming a familiar dashboard metric. Brands are tracking citations, comparing share of voice, and monitoring whether they appear in answers from ChatGPT, Claude, Gemini, and Perplexity. The numbers may look encouraging, but sooner or later the CFO will ask the question that matters: What did we get for the money?
That question shaped four workshops led by Kishan Panpalia of Pepper at the CMO Super Huddle. Kishan's argument was refreshingly direct: “AI search is now a demand generation problem.” Visibility matters, but CMOs need to connect it to commercial intent, buying behavior, pipeline, and ultimately revenue.
For B2B marketers, that connection is difficult because buyers rarely click directly from an AI answer and immediately request a demo. They discover a brand, share the recommendation with colleagues, search for it later, visit the website directly, and enter a long buying process involving multiple stakeholders. Traditional referral attribution misses most of that journey.
The measurement model needs to catch up with the buyer.
Disclosure: Pepper was a Founding Sponsor of the CMO Super Huddle. This article draws from Kishan's workshop framework and the candid questions raised by participating CMOs. Workshop quotes have been lightly edited for clarity, and participating CMOs are quoted anonymously.
The first problem may be the prompts themselves. Kishan cautioned that many companies celebrate rising visibility across broad informational queries while remaining nearly invisible in the commercial conversations that could produce revenue.
He shared the example of a large healthcare technology company that had risen from essentially zero AI visibility to more than 5%. The company ranked second in its overall category, which looked like a meaningful win. When Pepper examined the company's most important segment, however, its share of voice was only 10%, compared with 45% for a competitor that ranked lower overall.
The aggregate metric obscured the commercial reality.
Kishan recommends organizing AI-search measurement around approximately 10 to 15 core themes tied to products, services, industries, and ideal customer profiles. Those themes should emphasize commercial and transactional intent rather than generic educational questions.
One participating CMO captured the necessary discipline in four words: “Trust your primary research.” AI-search platforms can help monitor results, but the underlying prompt recommendations are still based largely on proxies because the major LLM providers do not offer complete search-volume analytics.
Kishan made a particularly important admission when the conversation turned to platforms such as Profound. Pepper competes with Profound in parts of the market, yet Kishan cautioned CMOs against treating recommendations from any monitoring platform, including Pepper's, as ground truth.
“Don't trust anyone with that,” he said. These providers draw from similar pools of third-party session data and supplement them with proxies such as search volumes and social-listening signals. They then recommend prompts to monitor before reporting how well the brand performs against those same prompts. As Kishan described the circularity, “The heads are mine, the tails are mine, and the coin is mine.”
The platforms may package their findings differently, but none has access to complete query-volume analytics from the major LLMs. Their dashboards can compare performance across the prompts entered into the system; they cannot prove that those prompts reflect what commercially relevant buyers are actually asking at scale.
That distinction matters when companies are spending significant sums on AI-monitoring platforms. A sophisticated dashboard filled with questionable prompts can create the illusion of progress while directing content investment toward conversations that barely exist.
False precision is still false. It just costs more.
CMOs should use these platforms for monitoring, comparison, and experimentation rather than unquestioned prompt discovery. Kishan's advice was to validate every proposed prompt set against primary customer research and then look for confirming signals in server logs. If visibility is rising but the associated pages receive no additional crawler activity, branded search, direct traffic, or commercial engagement, the company may be optimizing for its own dashboard rather than the market.
The audit should also examine whether years of content production have left the brand telling too many stories. More pages can create more opportunities to be indexed, but they can also make it harder for an LLM to determine what the company actually stands for.
Kishan offered Pepper's own experience as proof. The company removed 293 of approximately 400 blog posts so it could concentrate on one central idea. According to Kishan, Pepper's AI citations increased roughly sixfold over the following eight weeks “because we were known for something.”
That result challenges the content-machine mentality that has dominated B2B marketing for years. Publishing more articles about more topics may increase activity, but it can dilute the signals that help search engines, LLMs, buyers, and employees understand the company's expertise.
Content volume is easy to count. Strategic coherence is far more valuable.
CMOs should ask whether each major content asset reinforces a priority commercial theme. Old posts that contradict current positioning, compete for the same topic, or attract irrelevant traffic may be creating more confusion than value. Pruning, consolidating, and redirecting weak content could be as important as producing something new.
Direct referral traffic is an incomplete measure of AI-search performance, especially for companies with complex, high-value sales. Kishan described three revenue categories that can help CMOs establish a more realistic model:
For many B2B companies, AI-influenced revenue will be the largest and hardest category to measure. One CMO described a high-dollar, multimonth sale involving channel partners, CFOs, and security executives. Direct attribution would capture very little of that journey, even if AI introduced the company or influenced several members of the committee.
Kishan suggested watching three signals together: branded website traffic, branded search volume relative to competitors, and server logs showing visits from AI crawlers and agents. If AI visibility rises while those signals remain flat, the company may be optimizing for prompts that few commercially relevant buyers are using.
Server logs are particularly useful because they reveal which AI systems are crawling which pages. Kishan advised CMOs to request view-only access from their technical teams and correlate bot activity with the pages mapped to priority themes. As another Huddler noted, server-side data can remain available even when tighter cookie-consent policies reduce the visibility provided by traditional web analytics.
None of these signals provides perfect attribution. Together, however, they can reveal whether AI visibility is producing meaningful demand patterns rather than attractive dashboard numbers.
Most AI-search programs focus on questions asked by end users. B2B buying decisions involve many more people, each approaching the decision with different concerns.
Kishan recommended mapping the prompts used by procurement, finance, IT, security, legal, executive sponsors, and channel partners. Pepper's research found that common procurement queries include category shortlists, direct vendor comparisons, and searches for less expensive alternatives.
That means a company may need visibility for prompts such as:
A company that dominates end-user questions but disappears when procurement begins comparing vendors has not won AI search. It has simply entered the consideration set.
Once the audit is complete, Kishan recommends a controlled experiment that can give the CFO a stronger answer. Select two business themes with reasonably similar demand potential. Concentrate the company's AI-search effort on one theme for 90 days while leaving the other relatively unchanged.
During the test, compare:
Kishan said Pepper used this approach with six customers and generated enough evidence to support continued investment. The test does not create perfect attribution, but it replaces vague correlation with a structured comparison.
That is a significant improvement over reporting citation counts and hoping no one asks what they are worth.
CMOs should establish the baseline before the experiment begins, document which content and technical changes are being made, and avoid altering the control theme unnecessarily. A credible holdout test requires discipline from content, web, analytics, demand generation, and sales operations.
Measurement occupied most of the workshops, but the discussion repeatedly moved toward the next challenge: websites designed for AI agents as well as humans. Personal agents will increasingly help buyers research categories, compare providers, investigate implementation requirements, and assemble shortlists.
Kishan urged CMOs to examine whether agents can access technical documentation, pricing information, migration guidance, security details, and other late-stage buying content. Critical information is often hidden behind JavaScript, forms, firewalls, or gated PDFs, making it difficult for agents to retrieve.
This does not mean publishing confidential information or disclosing a fixed price for a complex enterprise sale. It means providing enough structured, machine-readable context for an agent to understand the offer. If the company will not define its likely pricing range, migration requirements, or distinguishing capabilities, third-party sources may define them instead.
One CMO raised the tension between accessibility and legal, intellectual-property, and security concerns. Kishan suggested creating public summary pages containing the information that can be shared, with deeper materials still gated when necessary. He also recommended making older assets accessible when they no longer generate meaningful leads and ensuring FAQs carry appropriate schema.
YouTube emerged as another opportunity. Kishan argued that LLMs can learn from transcripts, descriptions, and metadata even when a video generates little conventional engagement. A useful video with modest viewership may still influence AI answers if it clearly addresses an important question.
Before expanding the budget, ask the team to answer these questions:
The immediate priority is not to manufacture a definitive attribution number from incomplete data. It is to build a credible body of evidence connecting AI visibility with commercially relevant behavior.
Audit first. Establish the baseline. Then run the holdout test.
AI search may be changing how B2B buyers discover and evaluate vendors, but CMOs still have to prove that marketing investment creates business value. The tools are new. The obligation is not.
Combine AI visibility with branded website traffic, branded search volume relative to competitors, server-log activity from AI crawlers, and downstream pipeline or revenue signals. Direct referral traffic alone will miss much of AI's influence on a complex B2B purchase.
They can recommend and compare prompts, but the major LLMs do not provide complete query-volume analytics. Treat platform recommendations as hypotheses and validate them against customer research, sales conversations, commercial intent, and server-log activity.
A topic holdout test compares two business themes with similar demand potential. A company concentrates its AI-search activity on one theme for approximately 90 days while leaving the other relatively unchanged, then compares visibility, crawler activity, branded demand, pipeline, and revenue signals.
Content should be removed only after an audit. Consolidating, redirecting, or retiring outdated, contradictory, duplicative, and strategically irrelevant posts can clarify what the company is known for and strengthen the signals available to search engines and LLMs.
Make approved pricing context, technical documentation, security information, migration guidance, FAQs, and product details accessible in structured, machine-readable formats. Important information hidden behind JavaScript, forms, firewalls, or gated PDFs may be difficult for buyer agents to retrieve.