AI Search and the End of the Session Metric

Marketing has measured itself in sessions for twenty years because sessions were a reasonable proxy for attention. Someone wanted to know something, found your page, and arrived. Counting arrivals counted interest.

The proxy is degrading, and the underlying behaviour shift is documented rather than anecdotal: Pew Research’s 2026 survey of Americans and AI tracks how widely chatbots have been adopted for everyday questions. When an assistant answers one of those by synthesising from several sources — naming you, describing your product, recommending you — the user got what they needed and never arrived. Your influence was real and your session count is unchanged.

This is uncomfortable in a specific way. It is not that the numbers are getting worse. It is that they are getting less related to the thing you care about, which is much harder to explain to anyone.

Three distinct things called “zero-click”

The term is doing too much work, and the three cases have different implications.

Answered and satisfied. The question was informational and small. “What’s the refund window for X.” The user got the fact. They were never going to convert on that visit, and the click you lost had little value. This is the largest bucket and the cheapest loss.

Answered and influenced. The question was evaluative. “Best CRM for a small agency.” The user did not click, and they now hold an opinion about which vendors are credible. This is the case that matters most, and it produces exactly zero data on your side. Whether you were named, and how, decided something — and your analytics is silent about it either way.

Answered and redirected. The user did click, later, from somewhere else — searched your brand directly, typed your domain, came back a week later. The visit exists in your data and is attributed to direct or branded search. The AI answer that caused it is invisible.

Only the first is a genuine loss of a valuable session. The second is influence without measurement. The third is measurement without attribution. Reporting all three as “traffic decline” describes one of them correctly.

What replaces the session

Not a single metric. A small set, each answering a different question, and the honest framing is that you are trading one number of decreasing relevance for three of partial coverage.

Presence. Of the questions your buyers actually ask, in what share are you named at all? This is the closest available analogue to “did they hear about us,” and it is directly measurable by asking the engines. It does not depend on anyone clicking.

Position quality. Named as the recommendation, or named seventh in a list? These are different commercial outcomes with the same presence score, and the distinction carries most of the value — mentioned is not recommended covers the taxonomy.

Share against competitors. The most defensible number in the set, because it is relative. If you appear in 40% of relevant answers and your main competitor appears in 70%, that gap is meaningful regardless of what either absolute figure is worth, and it moves in response to your work. Share of voice is the formalisation.

Referral traffic, as a floor. AI-sourced visits are real and countable, and they materially undercount AI influence for the reasons above. Report them as a floor, never as a total, and never as a denominator in an ROI calculation. Tracking AI referral traffic covers what can be concluded.

Branded search volume. An old proxy that becomes newly useful. If AI answers are introducing you to people, some fraction of them later search your name. A rise in branded search with flat spend is one of the few external signals that discovery is happening somewhere you cannot see.

The part most articles skip

Here is where this topic usually overreaches, and the overreach is worth naming because it is tempting.

You cannot close the attribution gap. Not with better tooling, not with a clever model. A person who read an answer, formed an impression, and bought three weeks later through a direct visit has left no evidence connecting the two. Nothing recovers that link. Any vendor claiming to attribute revenue to AI mentions is modelling, and the model rests on assumptions you cannot validate.

Correlation between presence and revenue is not established at the industry level. It is intuitive that being recommended by AI helps. It has not been demonstrated with the rigour that, say, search ranking’s relationship to traffic has. Individual companies can build internal evidence over time — presence up, branded search up, pipeline up, nothing else changed — and that is genuinely useful. It is not the same as a published, generalisable relationship, and treating it as one invites a challenge you cannot answer.

Some of the traffic decline is not AI. Search behaviour, seasonality, competitors, your own content decisions and Google’s own changes all move traffic. Attributing a decline to AI without controlling for those is exactly the reasoning error this article is warning about, pointed in the other direction.

So the honest position: presence is measurable, influence is inferable, revenue attribution is not available, and the traffic decline has multiple causes. That is a weaker set of claims than most content on this subject makes, and it is what the evidence supports.

Reporting this without losing the room

The practical problem is organisational. Someone is accountable for a traffic number that is going to keep declining for reasons partly outside their control, and “we’re influential in ways I can’t show you” is not a survivable position.

Three things help.

Change the metric before it becomes a crisis. Introducing AI presence reporting while traffic is stable is a strategic addition. Introducing it the quarter traffic falls is an excuse, and it will be received as one regardless of being correct.

Report the composition of traffic, not just the total. If organic sessions fell 15% but conversion rate rose and branded search grew, the visits you lost were low-intent informational ones. That is a story about traffic quality with numbers behind it, and it is frequently true.

Show the competitive frame. “Our AI presence went from 30% to 45% while our closest competitor stayed at 60%” is a sentence an executive can act on. It contains a target, a trend and a gap. Competitor benchmarking is what makes it available; a single absolute score with no comparison is much harder to interpret and much easier to dismiss.

The counter-argument

The strongest objection: this is an elaborate justification for not being accountable to revenue. Marketing has a long history of adopting new metrics precisely when the old ones stop flattering it, and a discipline whose headline metric cannot be tied to money has a credibility problem it deserves.

That is a fair hit, and it should constrain how far this goes. Two responses.

The first is that presence is not a soft metric. It is directly observed — you ask the engines, you read what they say, you count. It is more directly measured than impressions, which nobody objected to for twenty years. The unmeasurable part is the link between presence and revenue, and that link was never actually measured for brand advertising or PR either; it was assumed.

The second is that the alternative is worse. If you measure only what leaves a session, you will systematically underinvest in the channel where your buyers increasingly form opinions, and you will find out it mattered when your pipeline reflects it. Measuring something imperfectly beats measuring nothing confidently.

But the objection earns a constraint: do not let presence become a vanity metric disconnected from commercial reality. Track it for the questions that precede purchases, not for every question in your category, and be candid that it is a leading indicator rather than a result. Measuring GEO/AEO ROI covers the boundary properly.

The summary

The session was a proxy, and proxies degrade when the underlying behaviour changes. It is degrading now.

What replaces it is not one number but a set — presence, position, share, and a referral floor — none of which closes the attribution gap, and all of which describe reality better than a session count that is quietly measuring less each quarter. The awkward truth is that you are trading a precise measurement of a decreasingly relevant thing for an imprecise measurement of the relevant one. That is still the right trade.