Zero-Click Answers and What Happened to Referral Traffic

Every figure you have seen for how much traffic AI search costs publishers is an estimate of a counterfactual — what your traffic would have been otherwise — and nobody can measure that. This page states what is dated and public, explains why the estimates disagree, and gives you the accounting that is genuinely yours.

What actually changed, and when

These are product announcements, which makes them tier-1 facts: they have dates and they were made by the operators themselves.

The measurement nobody can make

The question “how much traffic did AI search cost us” requires knowing what your traffic would have been in a world where the change did not happen. That world is not available. What is available is a before-and-after comparison, and the gap between the two is where every disputed number in this field lives.

Between your before and your after, all of the following also changed: the search operator shipped several ranking updates; your competitors published; your own site changed; seasonality happened; the query mix shifted because people ask differently when they expect an answer rather than a list. A before-and-after difference contains all of that plus the effect you are trying to isolate, and there is no arithmetic that separates them without a control group nobody has.

This is not a counsel of despair. It is the reason to be extremely suspicious of a confident single percentage, and the reason the useful measurements below are the narrow ones.

Why the published figures conflict

Research has been published — the Pew Research Center reported in 2025 on browsing behaviour on results pages with and without AI Overviews, using an opt-in panel of United States adults, and a number of publishers and analytics vendors have released their own analyses. Google has publicly disputed third-party claims of large click declines. Both sides can be arguing in good faith, because the studies measure different things:

• Different populations. An opt-in browsing panel is not a random sample of searchers, and a panel that agrees to have its browsing recorded skews in ways that are hard to correct for.
• Different denominators. Clicks per search, clicks per session, sessions per user and total sessions move independently. A metric that fell can coexist with a metric that rose.
• Different query mixes. A study weighted toward informational queries will find a large effect; one weighted toward navigational or transactional queries will not, because those surfaces did not change.
• Different windows. A comparison spanning a core ranking update attributes that update to AI.
• Different sites. Publisher aggregates are dominated by whichever large publishers are in the sample, and news traffic behaves nothing like documentation traffic.

When you read a figure, the four questions that decide whether it means anything are: who was measured, over what window, with what denominator, and compared to what. A figure missing any of the four is not a finding.

What your search console does and does not separate

A specific and consequential detail, because a lot of analysis is built on a misunderstanding of it. Google’s Search Console reports impressions and clicks for the Search result type, and links appearing inside an AI Overview are counted within that same Search type rather than broken out as their own surface. There is no filter that isolates them.

The practical consequence is that you cannot answer “how many of my clicks came from an AI Overview” from that data, and any tool claiming to do so is inferring it from something else — usually a third-party crawl of results pages, which is a sample of a personalised surface and comes with all the caveats of scraping one.

What Search Console still tells you honestly is the direction of your own impressions and clicks per query over time. Rising impressions with falling clicks on informational queries is the pattern people mean by zero-click, and you can see that pattern in your own data without needing anybody’s aggregate.

The accounting you can run

Assistant referrals are the one part of this that is straightforwardly measurable, because a click from an answer arrives at your server with a referrer like any other click. This is a tier-2 measurement and it is yours.

# Referrals from assistant surfaces, from a combined access log.
awk -F'"' '{print $4}' /var/log/nginx/access.log \
  | grep -Ei 'chatgpt\.com|openai\.com|perplexity\.ai|claude\.ai|copilot\.microsoft\.com|gemini\.google\.com' \
  | sed -E 's#^https?://([^/]+).*#\1#' \
  | sort | uniq -c | sort -rn

# And which of your pages they land on:
awk -F'"' '$4 ~ /perplexity\.ai|chatgpt\.com|claude\.ai/ {print $2}' \
  /var/log/nginx/access.log \
  | awk '{print $2}' | sort | uniq -c | sort -rn | head -30

Two honest limitations, and they both push the number down rather than up. Some assistant clients send no referrer at all, and some strip it via referrer policy, so this undercounts by an unknown amount. And a native desktop or mobile app may open a link in a way that produces no referrer, which means the channel is systematically under-attributed in every analytics package.

You also cannot tag these links. Ordinary campaign tracking works because you control the link; here the link is generated by somebody else from your own URL, so the only identifier available is the referrer. That is a structural limit, not a configuration problem.

The shape of the change, by query type

The mechanism predicts where an effect should be largest, and the prediction is checkable against your own query data rather than against anybody’s aggregate.

Segment your own Search Console queries into these buckets and look at the trend within each. That is a comparison with a shared denominator and a shared window, which is more than most published figures can say.

What to do that is not a guess

1. Instrument the assistant referral channel now, so that in six months you have a series rather than a snapshot.
2. Segment by query intent and stop reporting a single organic traffic number, which mixes categories that are moving in opposite directions.
3. Shift effort toward queries an answer block cannot close. Anything requiring judgement, a procedure, a calculation on the reader’s own inputs, or current data.
4. Make the click worth it. If your page is a summary, a summary of your page is a substitute for it. If your page is a harness, a table of your own working, or a tool, it is not.
5. Measure whether you are cited at all, separately from whether you are clicked, using the sampling protocol in measuring whether AI assistants mention you. Citation without a click still moves demand, and it is the only part of this that shows up nowhere in your analytics.