The Impression Is Dying. What Replaces It When the Audience Is an Agent?

For most of digital advertising’s history, value has started with exposure. An ad is served, a person has the opportunity to see it, and an impression is counted. From there come clicks, conversions, viewability and, eventually, sales.

That logic is so embedded in advertising that it is easy to mistake it for a law of nature. It isn’t.

In September, OpenAI began testing Sponsored Agents, a format that lets users move from an ad in ChatGPT into a conversation with an AI representative of the advertiser. The format itself is still experimental. But it points towards a bigger change: AI is bringing advertising closer to the decisions that happen before a person clicks, visits or sometimes even sees anything.

An AI system can retrieve publisher content, compare products, weigh competing claims and narrow a category down to a few recommendations. If that process changes what someone eventually buys, commercial influence has occurred, even if no traditional ad impression was generated along the way.

That creates a difficult question for an industry built around observable exposure:

What do you measure when the thing being influenced is the decision process itself?

From human exposure to machine influence

The impression was designed for a visible web. The commercial relationship was relatively straightforward: publisher or platform → advertisement → human.

Agentic discovery adds another participant:

publisher → AI system → recommendation → human.

And eventually, some journeys may go further:

publisher → agent → agent → transaction.

Consider someone asking an AI assistant which running shoes to buy for their first marathon. The system might use a specialist publisher’s comparison, combine it with manufacturer specifications and customer reviews, rule out most of the category and recommend two models.

The publisher’s work influenced the answer. The recommendation may influence the purchase. Yet the publisher might record no pageview and the advertiser no impression.

The economic influence happened. The familiar measurement events did not.

There are already signs of this gap. Research from Scrunch found that after AI platforms recommended unfamiliar brands, users were more likely to subsequently search for those brands, visit their websites or view their products on retailer sites. That does not prove that every later action was caused by the AI recommendation, but it illustrates the attribution problem: influence that begins inside an AI system may reappear elsewhere as branded search, direct traffic or retail activity.

For publishers, the same problem runs in the opposite direction. Their reporting, comparisons and reviews can shape an AI answer even when the person asking the question never visits the original page.

Traffic can fall while influence persists.

What replaces the impression?

AI referral traffic is one useful signal, but it only measures the occasions when an AI sends a person somewhere. It tells us very little about what happened before that referral.

Which information did the system encounter? Which sources actually shaped the answer? Which brands were evaluated but rejected? Which piece of publisher content changed the recommendation?

New initiatives are beginning to make those distinctions more visible. The SPUR Coalition’s Content Telemetry Standard, for example, separates stages such as retrieval, grounding, citation, presentation and engagement.

The underlying idea is more important than any individual standard: retrieving information is not the same as using it, and using information is not necessarily the same as influencing an outcome.

Advertising may eventually need a similar way of describing machine-mediated influence.

That could mean thinking about metrics such as machine reach, how often relevant AI systems encounter a brand or piece of content, or consideration share, measuring how often a brand makes it into the set of options an agent evaluates. For publishers, grounding shares could describe how often their information materially supports an answer. Further down the chain, recommendation share might capture how often a brand actually emerges as an option presented to the user.

These are not established industry standards, and some may prove difficult to measure. But they expose a growing mismatch: advertising has sophisticated tools for determining whether something appeared on a screen and far weaker ones for understanding whether information affected what an AI decided.

Publishers and advertisers face the same shift

For publishers, the traditional bargain has been relatively simple: create useful content, attract an audience and monetise its attention.

AI loosens the connection between the first two steps.

Imagine two product reviews that both receive 50,000 human visits. One is also regularly used by AI systems helping hundreds of thousands of people research purchases. The other is almost never used in AI-generated answers.

Their traffic dashboards could look identical.

Their broader influence would not.

That creates the possibility of a second kind of reach: not generic bot traffic, but machine reach, AI systems acting on behalf of people and using publisher information to make decisions.

This audience behaves differently. It does not consume content in the conventional sense. It uses it as input.

For advertisers, the same shift appears from the other side.

Imagine a traveller asking an AI assistant to find a hotel in Copenhagen with strong reviews, late checkout and a room under €300. Before the traveller sees any options, the system might evaluate dozens of properties using price, availability, publisher reviews, location and cancellation policies.

By the time three hotels appear on screen, much of the competition is already over.

The hotel that never entered the shortlist cannot recover with a brilliant ad placed underneath the answer.

That means advertisers increasingly need to think about more than reaching consumers. They also need to understand the systems determining which options consumers get to consider.

And this is where simple measures of “AI visibility” start to become insufficient.

Being mentioned is not the same as being considered. Being considered is not the same as being recommended. Being recommended is not the same as changing what someone buys.

The commercially useful question is not simply: Did our brand appear?

It is: Did our brand influence the decision?

The risk of turning consideration into another advertising market

Once brands recognise that AI systems shape consideration, they will naturally want to influence that layer. Platforms will want to monetise it.

That creates an important boundary.

There is a clear difference between transparent commercial participation and paying to quietly manipulate what appears to be an independent recommendation. If an AI assistant is useful because consumers trust it to narrow down a market intelligently, selling hidden preferences inside that process could undermine the product itself.

OpenAI’s Sponsored Agents offer one relatively clear model: the commercial relationship is explicit, and the user chooses to enter a branded conversation.

The harder questions emerge further upstream.

If a publisher’s review helps an agent choose one product over another, what exactly created the value? Was it the content being retrieved? Being used as grounding? Affecting the recommendation? Or helping produce the eventual transaction?

And who should capture that value, the AI platform, the publisher, the advertiser, or some combination of all three?

These sound like measurement questions. In practice, they are questions about who owns the economic layer between information and action.

The impression will survive. It just won’t tell the whole story.

The impression is not disappearing tomorrow. Neither are clicks, conversions or traditional attribution.

But they were designed for a media economy in which the commercially important events mostly happened in front of people.

Agentic discovery creates another layer upstream. AI systems retrieve information, compare alternatives, eliminate options and construct recommendations before the human sees the result. By the time an answer appears, part of the commercial competition may already have happened.

For publishers, that means understanding not only how much traffic content receives, but where that content influences machine-mediated decisions.

For advertisers, it means knowing not only where a brand appears, but whether it enters consideration and changes the outcome.

The web spent two decades learning how to measure opportunities to be seen.

The Agentic Web may force it to measure something harder:

the opportunity to matter before anyone looks.