Why AI assistants make advertising harder to measure

AI advertising creates a measurement problem for marketers. Familiar buying models preserve familiar advertising transactions. But an AI assistant can take part in questions, comparisons, and recommendations before a referral occurs, so a click may reveal only one part of the customer’s path to a decision.

The distinction changes the management problem. A paid-search campaign can expose signals such as the query, impression, click, and subsequent conversion on properties the advertiser can measure. An assistant-mediated journey can contain other interactions before the customer reaches the advertiser. Marketing teams need to separate measurable events from influence that may occur earlier in the journey.

AI advertising can make the click less informative

CPC measures a clear event: somebody clicked. Consider a customer who asks an assistant several questions, narrows a shortlist, compares products, receives a recommendation, and then visits an advertiser once. The advertiser could record one referral even though several earlier interactions helped shape the customer’s requirements. Click data alone cannot describe every interaction in that journey.

The management issue is the gap between what marketers can observe and the customer activity that may precede an observable event. CPC still measures the economics of clicks when clicks occur. Marketing teams need other evidence to assess influence that happens earlier in the journey.

From query and click to a partially visible journey

Paid search gives marketers operational signals that can include keywords, queries, impressions, clicks, and conversions. AI assistants can add conversational information. A person can describe a problem, add requirements, reject an option, ask for a comparison, and eventually request a recommendation. Each step can reveal more about the customer’s intent, while the advertiser may see only a later referral.

OpenAI says a ChatGPT ad “considers multiple signals, including the context and intent of the current conversation, the ad’s landing page, title, copy, advertiser-provided context hints, and, when ads personalisation is enabled, select signals from a user’s broader ChatGPT experience.” OpenAI has a commercial interest in adoption of ChatGPT advertising, so this is its description of its own ad-selection system.

The strategic issue is how much context can inform relevance and what that context contains. A search query provides one explicit expression of demand. A conversation can reveal needs, constraints, prior questions, use cases, and options the customer has rejected. Marketing strategy must represent those needs in the information and messages supplied to an AI system.

This changes where influence can occur. In an assistant-mediated journey, category explanation and product comparison can happen before a customer reaches an advertiser’s site. By the time a referral occurs, the customer may already have formed judgments about available choices. The advertiser then has direct visibility into only part of the process that led to the visit.

Attribution has to separate transactions from influence

Attribution is the process of assigning credit for a customer outcome to earlier marketing interactions. The journey described above creates a specific problem: one measured referral may follow several interactions that helped shape the choice. A click report records the referral event. Measuring the contribution of earlier assistant interactions requires evidence about those interactions.

This separates transaction metrics from influence metrics. CPC answers a defined question about the cost of acquiring a click. Influence measurement asks how earlier interactions contributed to consideration and choice. Keeping those questions separate prevents one observable event from standing in for the entire decision process.

The distinction can affect budget allocation. A channel that contributes to comparison or recommendation can receive little credit when a measurement system assigns value mainly to a later observable interaction. Management can then direct investment according to the events its systems capture rather than the full path that shaped the choice. Executives need to test where attribution methods can connect assistant-mediated interactions with eventual outcomes.

Product data becomes part of advertising strategy

Assistant-led product discovery can make catalogue data relevant earlier in customer consideration. Pricing, specifications, availability, and other structured product information can help a machine determine whether a product meets stated requirements. When an assistant uses that information to compare choices or formulate recommendations, catalogue quality can affect how it represents a product.

Amazon describes Rufus as “an expert shopping assistant trained on Amazon’s product catalogue and information from across the web to answer customer questions on shopping needs, products, and comparisons, make recommendations based on this context, and facilitate product discovery in the same Amazon shopping experience customers use regularly.” Amazon has a commercial interest in increasing shopping activity within its own marketplace, so executives should read this as Amazon’s characterization of Rufus.

For marketing leaders, that model connects catalogue quality with demand generation. A stale specification or poorly structured attribute could affect a machine’s assessment of whether an item satisfies a customer’s requirements. Machine-readable product information can support product representation during discovery as well as the accuracy of an e-commerce listing. Marketing, e-commerce, and data teams therefore share a dependency on the same information.

The broader management question applies across assistant-led product discovery. If assistants consume product data directly, weak information can affect the inputs used for comparison and recommendation. Marketing performance can then depend partly on the quality, structure, and freshness of data maintained outside the media-buying function. Executives need clear ownership of those inputs when they become part of customer-facing machine decisions.

Cheaper creative increases the value of selection

When asset generation requires less production effort, teams can create more variants within the same workflow. More of the management burden then shifts to choosing concepts, deciding which variations deserve tests, and determining which results merit wider deployment. Teams also need clear brand direction and review standards. Greater production capacity is useful only when the organisation can select and govern what it produces.

Conversational advertising raises the stakes because richer context creates more situations for tailored messages. Generative systems can expand the set of messages a team can produce for those situations. Disciplined selection can then become the scarce capability: deciding which customer needs matter, what the brand should say about them, and how much variation the organisation can review reliably.

More relevance creates a harder control problem

Conversational context can support more tailored outputs while increasing the number of factors involved in a recommendation. When an AI system interprets customer requirements, combines information, and presents a product inside a generated response, advertisers may need ways to assess how it represented their products. Governance then extends to both the information supplied to the system and the records needed to reconstruct consequential outputs.

Agent systems raise a related audit question when several automated participants contribute to an action. If one agent acts for an advertiser while another assists a customer, management may need records showing what information affected a recommendation, what policy governed the action, and which party was responsible for each step. Auditability means being able to inspect and reconstruct those actions from reliable records.

The framework in “Auditable Agents” turns a broad governance problem into operational tests. Management can ask whether teams can reconstruct an agent’s action, inspect relevant activity across its lifecycle, check behavior against policy, identify responsible parties, and trust preserved evidence. Those capabilities become material when automated systems exchange recommendations or take actions that affect how customers encounter a brand.

Key highlights

• Measure beyond the click: AI assistants can shape customer requirements and choices before a referral occurs. Marketing leaders should distinguish click economics from evidence of earlier influence.
• Adapt attribution to partially visible journeys: Assistant interactions can contribute to consideration without appearing in conventional campaign data. Executives should test whether attribution methods can connect these interactions with eventual outcomes.
• Treat product data as a marketing input: Assistants can use pricing, specifications, availability, and other structured data to compare products. Marketing, e-commerce, and data teams should establish clear ownership for its quality and freshness.
• Make creative selection the priority: Generative AI can reduce the effort required to produce more creative variants. Leaders should strengthen testing, brand standards, and review processes so greater production capacity results in useful advertising.
• Build controls for AI-mediated decisions: More contextual recommendations and agent interactions can make it harder to reconstruct how products were represented or actions were taken. Leaders should require reliable records, policy checks, and clear accountability for consequential automated activity.