In a zero-click answer the citation is the impression. Traffic stops being the measure of visibility, which breaks every reporting model that treats sessions as the leading indicator of awareness.
Pew Research measured the gap on Google: users clicked a search result on 8% of visits where an AI summary appeared, against 15% where one did not.
The awkward part is not that traffic falls. It is that the thing you are trying to measure moved to a place your analytics cannot see, so a flat or falling session count is now compatible with visibility going up.
What does the data actually show?
Pew Research Center’s study of Google users tracked the browsing of 900 US adults through March 2025, covering 68,879 unique Google searches, roughly 12,593 of which produced an AI summary. Three findings matter.
| Behaviour | With an AI summary | Without |
|---|---|---|
| Clicked a traditional search result | 8% of visits | 15% of visits |
| Clicked a link inside the summary | 1% of visits | Not applicable |
| Ended the browsing session entirely | 26% of pages | 16% of pages |
Two caveats before anyone builds a strategy on this. It measures consumer Google behaviour, not B2B buyers inside ChatGPT or Perplexity, and nobody has published equivalent numbers for those surfaces. And it is a snapshot from a period when AI summaries were newer than they are now.
The direction is still the useful part. Roughly half the click-through, and a link inside the summary itself is a rounding error at 1%.
Why does the citation become the impression?
Because the answer is where the reader forms their view, and the citation is the only part of you that appears there. If a buyer reads a paragraph naming three vendors and closes the tab, that paragraph did the work a landing page used to do.
That is a real impression with no session attached. It shaped a shortlist, and the only artefact it leaves is a citation or a mention inside an answer you never see.
A session is evidence that someone arrived. A citation is evidence that someone was told about you. Only one of those still happens reliably.
This is also why a citation is not a backlink in any useful sense. It is closer to a mention on a shortlist that the buyer reads once and acts on, which is the distinction between the three outcomes in mention rate vs citation rate.
What breaks in your reporting?
Three things, and they break quietly rather than obviously.
- The assumption
- Awareness rises, then traffic rises, then pipeline follows.
- What broke
- The first step no longer produces the second, so the chain reports a decline during a period of rising awareness.
- The assumption
- The channel that delivered the session gets the credit.
- What broke
- A buyer who read an answer and typed your name directly is recorded as direct traffic, and the answer gets nothing.
- The assumption
- Ranking for a query produces visits from that query.
- What broke
- The query is answered in place. You can hold the ranking and lose the visit, which reads as a ranking problem and is not.
The diagnostic tell is a rise in branded search or direct traffic alongside flat unbranded performance. That pattern is usually somebody being told about you somewhere you are not measuring, and it is worth checking before concluding that a campaign underperformed.
What should you measure instead?
The answer itself, sampled deliberately, plus the one first-party citation source that exists.
Microsoft publishes actual citation counts. Its AI Performance report in Bing Webmaster Tools shows “the total number of citations that are displayed as sources in AI-generated answers” along with the queries that retrieved you. It covers Microsoft surfaces only, and it is a census rather than a sample, which makes it the most solid number available anywhere in this discipline.
Everywhere else, you sample. Run a fixed prompt set, record the answers, and count how often you are named, cited and recommended. That is a different instrument from analytics and it measures the thing analytics stopped being able to see.
Retrievability still matters underneath all of it. Google’s documentation on AI features states a page “must be indexed and eligible to be shown in Google Search with a snippet,” so the SEO work that used to earn the click now earns the citation instead. The work did not become pointless, the payoff changed shape.
Is this worse or better for a small brand?
It depends entirely on whether you were winning the click before. A brand that ranked third and collected a share of the traffic has lost something real. A brand that ranked eleventh and collected nothing has lost nothing, because it was already invisible.
The second case is the more common one in B2B, and it is where the opportunity sits. An answer names three or four vendors regardless of who ranks first, so the list is shorter than a results page and membership of it is not decided by the same mechanics. A brand that could never reach the top three of a search page can be named in an answer if the sources the engine reads say the right thing.
That cuts both ways. The shortlist being short means the gap between being named and being absent is starker than the gap between ranking third and ranking fifth ever was. There is no long tail of partial credit in an answer.
So the honest summary is that zero-click compresses the field. Fewer winners, a bigger gap between them and everyone else, and a route in that does not require out-ranking incumbents on their own terms.
Does this mean traffic is irrelevant?
No, and overcorrecting here is its own mistake. Traffic still converts, still funds the business, and a citation that never produces a visit or a mention in a sales call did not do much.
What changes is its position in the chain. Traffic moves from leading indicator to lagging one, because awareness now accumulates in a place that produces visits later and indirectly, if at all.
So keep reporting it. Just stop treating a flat line as evidence that nothing happened, and stop treating a rise as proof that the AI work is landing, because neither inference holds any more.
How do you report this to someone who buys sessions?
Show both, and be explicit that they measure different stages. A slide with visibility on one axis and sessions on the other, with the methodology stated, survives scrutiny better than either alone.
State the sample size and the run count alongside the visibility figure, because the first question from anyone numerate is how solid it is. A rate with no stated observation count invites exactly the scepticism you are trying to defuse, which is the threshold discipline in is my AI visibility change significant.
And separate branded from unbranded prompts in what you present. Branded results test whether the engines know you, unbranded results test whether you get discovered, and only the second is a growth number, which is the split in realistic prompts for AI visibility tracking.
If the number moves and you cannot explain it, check the shape of the move before you build a story on it. A change confined to one engine is usually a platform release rather than anything you did, covered in AI visibility dropped suddenly.