The CMO guide to AI Search Intelligence

On July 23, 2025, Google said AI Overviews had passed 2 billion monthly users. AI Mode had also passed 100 million monthly active users in the United States and India. Both figures came from Alphabet's second-quarter update reported by TechCrunch.

Those adoption figures made AI-mediated discovery a distribution question at executive scale. They did not create a clean revenue-attribution model for any individual brand.

That distinction should shape the CMO program. AI Search Intelligence can measure how a brand appears, which sources support the answer, where competitors lead, and which owned- or earned-media work deserves attention. It cannot claim that an observed visibility change caused pipeline or revenue.

Key takeaways

• Google reported more than 2 billion monthly AI Overview users in July 2025. It also reported more than 100 million monthly AI Mode users, but neither figure proves commercial impact for one brand.
• The operating cycle keeps six inputs fixed: buyer questions, prompt version, engine scope, geography, sampling, and counting rules.
• The executive view reports four separate signals: mention rate, share of voice, average position, and citation share.
• Weeks 5 through 10 belong to the teams that write, publish, maintain the website, and earn coverage. Measurement does not replace execution.
• The board memo ends with limitations and the next measurement date, not an invented AI-attributed pipeline figure.

AI Search Intelligence turns a metric into an operating decision

AI Visibility describes how prominently a brand appears across tracked AI engines. Zumi is an AI Search Intelligence Platform. Its metric dictionary defines mention rate, share of voice, average position, and citation share as four separate descriptive signals.

AI Search Intelligence is the wider operating discipline. It defines the questions worth observing, preserves the answers and sources, separates the four signals, assigns the resulting work, and measures the same scope again.

The distinction matters because a dashboard can report movement without creating accountability. A CMO needs to know which decision follows, which team owns it, and which claim the evidence can support. That is the role of the 90-day cycle.

The first 4 weeks fix scope and baseline

Weeks 1 and 2 fix the measurement object

The first two weeks prevent a changed ruler from becoming a success story. The program owner defines the buyer-question universe before any baseline is reviewed.

The scope record contains:

• Market, geography, language, and target audience
• Unbranded category, use-case, comparison, and problem questions
• Stable prompt IDs and a versioned prompt panel
• Named engines or answer products selected for the audience
• Declared competitors and canonical entity names
• Session, sampling, exclusion, and counting rules
• Decision owner and executive review date

Engine scope follows the audience and engagement. Zumi can cover up to nine AI engines, with the tracked set scoped to the requirement. A report should state the exact set measured rather than convert platform capacity into a claim of universal coverage.

The prompt panel also stays unbranded where the purpose is discovery. A question containing the brand name tests description or validation. It does not test whether the brand enters an unprompted category answer.

Weeks 3 and 4 establish a four-signal baseline

The baseline captures the unedited answers and citations before the team begins work. Screenshots can make the result legible, but the structured answer record remains the evidence.

Each engine and question pair produces the same fields:

• Mention rate: Whether the brand appears across the eligible tracked answers.
• Share of voice: How the brand's observed presence compares with the declared competitor set.
• Average position: Where the brand lands when it appears with other brands.
• Citation share: How often the brand's own pages support the answer.
• Engine disagreement: Where the same question produces different brand or source patterns.
• Source risk: Which owned or third-party pages carry outdated, weak, or conflicting information.

The four signals remain separate. A composite score can make a board slide look tidy while hiding the reason for movement. The baseline should preserve the diagnosis that the next phase needs.

The baseline is also dated. One collection period describes what the selected engines returned under the stated conditions. It does not establish a permanent rank or an estimate of what every user saw.

Weeks 5 through 12 assign work and measure again

Weeks 5 through 10 assign owned and earned work

The middle of the cycle turns observed gaps into a finite work queue. Zumi's Recommendations feature groups actions into Owned Media and Earned Media, then prioritizes them per engine. The brand or agency team performs the execution.

Owned-media work can include a missing category explanation, an outdated product page, inconsistent entity information, or a page that fails to answer the question an engine is resolving. Content, product marketing, and web teams own those changes.

Earned-media work begins with the third-party sources already cited for the category. Public relations, communications, partnerships, or an agency decides whether the missing coverage is relevant and attainable. Zumi identifies where coverage may matter; it does not conduct outreach on a public self-serve engagement.

The action register keeps observation and interpretation apart:

• Observation: The brand was absent from one priority question cluster in the selected engines.
• Source evidence: The answers cited category pages that did not mention the brand.
• Interpretation: The category may lack third-party support for the brand's stated fit.
• Action: The communications team evaluates the cited publications and corrects the owned category page.
• Owner: The register names the function and accountable leader.
• Review: The register stores the publication date, evidence link, and remeasurement date.

The register does not label every absence a content problem. Some gaps reflect weak relevance, a narrow product fit, inaccessible source material, or a prompt that never belonged in the panel. Removing a weak question requires a new panel version rather than a silent edit to the baseline.

Weeks 11 and 12 measure the same object again

The final phase reuses the unchanged panel, scope, competitors, and counting rules. New questions can enter a separate series, but they should not pad the original comparison.

The review asks four questions:

• Which of the four signals changed within the original scope?
• Which questions, engines, and sources account for the movement?
• Which owned- or earned-media actions were completed before the second measurement?
• Which explanations remain plausible but unproven?

A before-and-after comparison is observed evidence, not a controlled experiment. Engines, retrieval systems, source indexes, and competing content can change during the same period. The report can state that a movement followed a shipped action and appears in the affected question cluster. It should not label the action as the sole cause.

Ownership keeps the program out of the SEO silo

AI discovery crosses several operating functions. The CMO owns the business question and the decision standard, while specialist teams own the work their function can actually perform.

• CMO or marketing leader: Scope, budget boundary, decision standard, and executive narrative.
• Product marketing: Entity definition, category language, product facts, and comparison accuracy.
• Content: Owned-page gaps and evidence quality.
• Public relations or communications: Relevant third-party source gaps and earned coverage.
• Web or technical search: Crawl access, page structure, canonical information, and publishing.
• Analytics: Referral and commercial data kept in a separate attribution layer.
• Legal or compliance: Claims, regulated statements, source permissions, and risk review.
• Agency lead: Cross-client method, reporting consistency, and delivery ownership where contracted.

No single team can repair every source behind an AI-generated answer. The operating model works because the diagnosis routes to the team with authority to act.

The board receives one page and one decision

The board memo should fit on one page. Supporting answer excerpts, source tables, and methodology notes can sit behind it.

The page contains:

• Decision required: Continue, change, expand, narrow, or stop the program.
• Scope measured: Buyer-question panel, market, engines, dates, and sampling.
• Four-signal movement: Each signal shown separately against the baseline.
• Engine disagreements: The material differences hidden by an average.
• Source risks: The pages shaping weak, outdated, or conflicting answers.
• Actions completed: Owned- and earned-media work that shipped during the cycle.
• Limitations: What the observation cannot establish.
• Next cycle: Accountable owner, budget, and measurement date.

The memo should include a typical answer excerpt, not the most flattering screenshot available. The excerpt makes the data concrete without replacing the panel.

The investment case closes a blind spot

The defensible budget case is operational. Marketing already manages how a brand appears in search results, media coverage, review sites, and owned content. AI-generated answers create another discovery layer that can be observed and governed.

Initial investment funds a bounded baseline, source diagnosis, responsible action, and repeated measurement. The decision after the first cycle can be to expand the scope, continue at the same level, reduce it, or stop.

A vendor evaluation should reject:

• A composite score without its formula and underlying signals
• A result with no prompt panel, engine scope, date, or raw answer evidence
• Engine-capacity language presented as identical coverage on every engagement
• Screenshots selected without a stable comparison set
• Causal revenue claims inferred from visibility movement
• Recommendations with no owner, evidence path, or remeasurement date

Commercial attribution remains outside this guide. Referral analysis, self-reported discovery, customer-relationship data, and incrementality work belong to a separate analytics design with their own evidence rules. The manual AI visibility baseline is the concrete starting point for a team that has not yet established the discovery measure.