Google's Dive deeper test inserts an AI overview ahead of the publisher's page

Count the screens in the flow the reporting describes. A tap on "Dive deeper" opens an AI-generated summary, and behind that sits a second screen of related-source cards and an "Explore more stories" area ^[4]. So two Google-hosted screens sit in front of the publisher page that a card tap used to open on its own ^[16].

The publisher links survive the change ^[9]. A link can hold its position and still take a smaller share of clicks, and the dev.to write-up leaves both readings open: a concise overview may satisfy some users before they visit an external article, or it may give the links that follow enough context to make them more useful ^[9]. Google has released no referral or click-through data, so neither outcome can be quantified ^[8].

Google has not said which countries or languages the test runs in, what it did to traffic, when it might roll out, or how a publisher would opt out, and it calls this a live experiment ^[7].

Discover is a recommendation feed, and on the dev.to piece's reading publishers cannot assume Search behaviour, referral patterns, or optimisation tactics transfer to it ^[10]. The Search Console Discover report is the only shared yardstick, and the useful version of it is segmented by content type ^[14].

The same piece argues for measurement and distribution resilience, and for establishing a baseline for Discover performance before any change reaches an audience, so a real shift in referrals can be told apart from normal Discover volatility ^[13]. Its list of signals runs to landing-page engagement from Discover referrals, including returning users, pages per visit, subscription starts, leads and ad-view depth, plus video performance, topic-level patterns and referral concentration ^[14]. Those are all things a publisher can measure without Google's cooperation.

For this test to move a given publisher's numbers, that publisher's Discover traffic has to arrive through video cards. The experiment begins there ^[5], and Google's public statement covers video cards only ^[12]. PPC Land reported that independent researchers observed code flags that may point to broader formats, including article cards, and those flags are not a confirmed expansion ^[11]. Google also plans to test multiple designs in the coming weeks, so the prominence of the publisher links in the version people eventually see is not fixed by the one that was demonstrated ^[6].

Stein's own description of the feature puts the links inside the overview: a short topic overview with prominent links to related stories, community reactions, and original reporting ^[3]. Whether "prominent" survives design iteration is the thing to measure. From the outside, only the sites on the receiving end can measure it ^[15].