Google tests AI 'Dive deeper' overviews in Discover as AI answer citations diverge

Google is testing an AI layer inside Discover

Google is running an experiment in Discover called "Dive deeper" that inserts an AI-generated topic overview between a card and the publisher page a user would normally open. Google Search VP Robby Stein announced the test on September 18, 2026, according to a report on dev.to. The experiment begins with video cards: tapping the option opens a short AI-written overview of a topic, followed by related coverage, community reactions, and links to original reporting.

Google intends to try multiple designs over the coming weeks. It has not said which countries or languages are covered, has released no referral or click-through data, and has offered no opt-out. PPC Land, cited in the same report, noted that independent researchers spotted code flags hinting the feature could eventually reach article cards, but Google has not confirmed any expansion beyond video.

The sequence changes even if links stay

Publisher links remain in the experience; what changes is the order of steps. A Google-hosted summary becomes the first stop, and it may answer a user's question before they ever reach an external article. Discover is also structurally different from Search: it is a recommendation feed rather than a query-driven results page, so referral patterns and optimization tactics from Search cannot be assumed to carry over.

Because no traffic data exists yet, neither a benign nor a damaging outcome can be quantified. The dev.to report recommends publishers record baselines now: Discover clicks and impressions in Search Console, engagement on landing pages, video performance, and how concentrated referral traffic is. Those baselines make it possible to distinguish a genuine shift from Discover's normal volatility once any change arrives.

Semrush data shows mentions and citations diverging

Separately, Semrush's expanded 2026 AI Visibility Index, built on 126 million AI search prompts across platforms including ChatGPT, Google AI Mode, and Gemini, found that the brands AI systems name and the websites they cite frequently do not match.

Across ChatGPT and Google AI Mode, 67% of mentioned brands overlap, but only about 30% of cited sources do. In other words, two systems can identify largely the same companies while building their answers on substantially different evidence. Source diversity also varies widely: Semrush measured roughly 15 cited sources per ChatGPT response against around three for Gemini, a gap the company notes makes raw citation counts non-comparable between platforms.

The index also identifies a small set of brands, the "Universal 36," that hold visibility across four major AI platforms, while indicating that new brands can still break into top visibility results.

Why it matters

Both developments describe the same underlying shift: attention and influence are consolidating inside AI surfaces, increasingly before, or instead of, a click. Discover's Dive deeper adds one more point where a user can remain within Google's own presentation of a topic. Semrush's data shows the effect is already real for brands, since an AI answer can shape a purchase decision even when the company's own site is never cited, with third-party reviews, publishers, and retailers doing the describing.

Neither development proves traffic collapse. Google has published no impact data for Dive deeper, and Semrush itself cautions that citation frequency alone says little about trustworthiness. But together they sharpen two practical tasks. Publishers need baselines and close attention to how prominently links appear as Google iterates on Dive deeper designs, since placement and click paths will determine whether referral value survives. Brands need to track high-value prompts across AI platforms, recording who gets mentioned and who gets cited, rather than relying on a single domain-level authority score. The open question is whether Google's iterations keep publisher links prominent enough to preserve meaningful referral traffic.