Pew: Google AI Overviews roughly halve traditional clicks; new tool tracks AI staleness

AI summaries cut traditional search clicks nearly in half

A Pew Research Center study has put numbers on what many publishers suspected: users clicked a traditional search result in 8% of Google visits that included an AI Overview, versus 15% of visits without one — a relative drop of almost half. The study, summarized in a dev.to write-up, covered 68,879 Google searches performed by 900 U.S. adults in March 2025.

Summaries were not rare in the sample. They appeared on roughly 18% of all searches, and their frequency tracked query length: present on only 8% of one- or two-word queries, but on 53% of searches containing ten or more words. About 58% of participants ran at least one search during the study window that produced a summary.

Zero-click sessions grew as well

The zero-click gap was even wider. Sessions ended without a visit to any website in 26% of cases where a summary appeared, compared with 16% where none did. As the write-up stresses, this is correlation rather than proof — Pew cannot show that the summary alone prevented clicks, only that journeys containing one more often stopped at the results page.

The summaries themselves were barely a source of outbound traffic: users clicked a link inside an AI Overview in just 1% of all visits. Citations were broad, though, with 88% of summaries referencing three or more sources. Wikipedia, YouTube and Reddit together made up about 15% of cited sources, and government sites appeared more often among AI-summary links than among standard results, 6% versus 2%. News sites accounted for 5% of links in both formats.

What it means for publishers and SEO

The practical problem is attribution. Falling visits to a question-answering page can now reflect a redesigned results page rather than lost rankings or weaker demand, and because long informational queries trigger summaries most often, quick-explainer content is the most exposed. The dev.to analysis argues the durable response is material a concise summary cannot substitute for:

• original evidence such as research, tested data or documented methodology
• interactive resources like calculators, templates or product information
• implementation detail grounded in real, firsthand work
• a clear next step so arriving visitors can evaluate, buy or subscribe without friction

It also recommends widening measurement beyond rankings and aggregate sessions — impressions, landing-page conversions and the paths visitors take after arrival — and comparing summary-prone query groups against high-intent pages to separate shrinking click opportunity from genuine page problems. The write-up cautions that Pew measured behavior, not ranking factors, so this is not an optimization playbook endorsed by Google.

A separate tool tracks how stale AI models are

On the answering side of this shift, a single-page tracker that reached the Hacker News front page charts release dates and training cutoffs for 20 current models from 8 labs, with live counters ticking upward from each date. The data on "How Stale Is Your AI?", hosted at stale.jock.pl, was checked on 2026-09-16, and each model entry links to the lab document its date came from.

Only 10 of the 20 models carry a cutoff their lab actually publishes, and just 5 of the 8 labs — Anthropic, Google DeepMind, Meta, OpenAI and xAI — disclose one for at least one model; for the rest, the page's checked vendor sources did not establish a date. The gaps can be large: Gemini 3.1 Pro shipped on Feb 19, 2026 with training data stopping in Jan 2025, while GPT-6 Astra launched on Sep 3, 2026 having last read on Apr 30, 2026. The page also argues that web-search features mask rather than close the gap, since retrieved pages inform a single answer and are then forgotten rather than learned.

Why it matters

The Pew figures quantify a structural change in the web's traffic economics: when a summary appears, roughly half the clicks that once flowed to ordinary results do not, and visibility becomes decoupled from visits. Standard analytics will increasingly understate reach on precisely the informational queries publishers used to win. The staleness tracker exposes the mirror-image problem: the systems absorbing that attention carry knowledge horizons months behind their launch dates, and half the models surveyed publish no cutoff at all. The search layer answers faster while knowing less than it appears to.