Google Discover’s “Dive Deeper” Test (September 18, 2026): Why Michigan Brands Shouldn’t Call It a Publisher Problem

On Friday, September 18, 2026, Google’s Robby Stein announced on X that Discover is testing a feature called “Dive deeper,” and Search Engine Roundtable published screenshots the same afternoon. Tapping it does not open the publisher’s page. It opens an AI-generated topic overview, labeled “generated with AI” and carrying Google’s own “which can make mistakes” disclaimer, with links to related stories, community reactions, and original reporting placed inside the panel.

Within a day the story had been filed, almost everywhere, under publishing. That filing is the mistake. For plenty of Michigan service companies, dealerships, and catalog sellers, Discover is mobile traffic that never appears in a keyword report because nobody typed a keyword. A layer that answers the question before the tap is a commercial problem wearing a publisher’s coat.

What is Google Discover’s “Dive deeper” feature?

Dive deeper inserts an AI-written topic summary between a Discover card and the source it came from. Stein called it “a new way to make exploring a topic and discovering related web content easier in Discover,” and said that by tapping Dive deeper, “you’ll see a short topic overview with prominent links to related stories, community reactions & original reporting.”

Two details matter more than the description. First, Stein wrote that Google will “test this experience on videos to start, and experiment with multiple designs in the weeks ahead.” Second, this is a test — no rollout date, no help document, no Search Console reporting change. Treat any claim about how widely it is live as unconfirmed. The direction is consistent, though: on September 18 Google was also spotted testing an AI Mode button inside the search results bar, and on September 20 Search Engine Roundtable reported that Web Guide’s “Classic search” button does not actually return users to classic search. Entry points into AI layers keep multiplying; exits keep narrowing.

Why did Google start the test on videos instead of articles?

Google has not said, so this is a reading rather than a fact: video is the surface where Google carries the least publisher-relations risk and already holds the underlying text. A news article belongs to a publisher with a legal department. A video’s transcript is something Google generates itself.

The mechanism tells you what gets summarized. For a video card there is no article body to compress, so the overview is assembled from the transcript, the title and description, and whatever written coverage surrounds the topic. That produces a predictable failure: content whose value lives in the demonstration gets flattened into the words spoken over it.

Take an install video from an aftermarket shop in Warren: a bracket fits a 2019 chassis but fouls on the 2021 revision. The narration says “line it up and torque to spec.” The visual — the half-inch of clearance that disappears — is the entire point. A summary built from that transcript reads as a complete answer and is wrong in the one way that generates a return shipment. Same for a winterization walkthrough out of Harrison Township.

Does “Dive deeper” only affect news publishers?

No. Discover surfaces evergreen how-to content, product explainers, buying guides, and now video — not only dated news. Any page eligible for Discover sits inside the blast radius of a feature that summarizes Discover topics.

The Michigan exposure is seasonal and about to peak. Discover traffic is mobile and mostly evening — browsing behavior, not searching. That is the pattern of a homeowner in Rochester Hills thinking about the furnace in October, a boat owner in Traverse City thinking about shrink-wrap in November, a parts buyer scrolling after a shift ends in Sterling Heights. None of them type a query; they get served a card. If that card opens into a summary instead of your page, the top of your Q4 funnel narrows without a single ranking changing.

How does Google decide which sources appear inside the overview?

Google has documented nothing about selection. But Stein’s description names three slots — related stories, community reactions, original reporting — and three named slots implies three qualification routes rather than one ranked list.

“Related stories” is a topical-coverage slot: it rewards several dated pages on one subject. “Community reactions” is a forum-and-social slot most businesses cannot win and should stop chasing. “Original reporting” is the only slot rewarding something a competitor cannot replicate by writing a better paragraph — a measurement you took, a photo of the actual part, a failure rate you observed across jobs.

Run that forward. A Southfield law firm with one comprehensive page on a practice area has exactly one route into the panel, and it is the crowded one. A Grand Rapids HVAC company with six shorter pages — each dated, each carrying different first-hand evidence — has multiple shots at two of the three slots. The consolidation advice that worked for classic SEO, merging thin pages into one authoritative asset, works against you on a surface that assembles from several sources at once.

Is Google’s framing of Dive deeper accurate?

Partly. Google says the links inside the overview are “prominent,” and in the published screenshots they are. The framing is still misleading, because prominence inside a panel is not equivalent to the tap that used to be the whole card.

Before this test a Discover card had one destination: you. Now it opens onto a set of destinations, where your link competes with the summary above it. Barry Schwartz put his judgment on the record the day it was announced — the overview “gives the full context to the user – so why would a user click?” The part that will hurt small teams most: Search Console’s Discover report cannot separate a user who read the summary and left from one who never saw your card. Impressions may hold while clicks fall, with no reason given. Teams without a baseline will spend November debugging a content problem that is actually a surface problem.

How should a small Michigan team prepare for Dive deeper?

Eight steps a three-person team can finish inside a week, none requiring an enterprise platform.

1. Export your Discover baseline today. Pull impressions, clicks, and CTR for September 1–21, 2026 from the Discover performance report. Once Google cycles “multiple designs in the weeks ahead,” establishing what normal looked like is no longer possible.
2. Verify max-image-preview:large in the robots meta tag on every template. Without it Discover serves a small thumbnail instead of the full-width card. Check the blog, service-page, and video landing templates separately — they often diverge.
3. Publish the transcript on the same page as the video. Timestamped, in real HTML, not in an accordion loaded by script. If the overview is built from text, the transcript is the document.
4. Write the visual caveat into the words. Wherever the demonstration carries what the narration does not — clearance, color match, torque sequence — say it out loud. A summary can only carry what the text contains.
5. Add one piece of first-hand evidence per cluster. A photo you took, a measurement you made, a count from your own job records. This is the only route into the “original reporting” slot.
6. Keep dateModified honest in your schema. Accurate, not cosmetically refreshed. An overview assembling “related stories” has to order them somehow, and dates are the cheapest signal available.
7. Move the conversion above the explanation. If the summary answers the “what,” your page has to answer “who do I call and how fast can you get here.” Phone number, service area, and next-available window above the fold.
8. Tag Discover landing pages as their own GA4 cohort. Segment those URLs now so an October drop is attributable to a surface rather than blamed on the content. That plumbing is the least glamorous, most load-bearing part of the technical SEO work we do for Detroit-area brands, and it takes an afternoon.

One step is missing on purpose: nosnippet and max-snippet. Blocking snippets to starve the overview also removes you from the card preview that earns the tap. For commercial content that is bad arithmetic. Publishers with subscription revenue may calculate it differently.

What dates should Michigan marketers have on the calendar?

• September 18, 2026 — Dive deeper announced; test begins on video surfaces.
• September 22 – October 3, 2026 — the clean baseline window. Export Discover data weekly before design variants start cycling.
• October 2026 — Google’s stated “weeks ahead” of multiple designs. Expect inconsistent screenshots; resist concluding anything from one.
• October 1 – November 15, 2026 — Michigan heating-season and winterization demand climbs, and evening mobile browsing peaks for exactly the content Discover distributes.
• Mid-January 2027 — the earliest date a year-over-year Discover comparison means anything.

Frequently asked questions

Is Dive deeper live for everyone right now?

No. As of September 21, 2026 it is an experiment, announced September 18 and confirmed by Google only as a test on video surfaces with multiple designs planned. Sightings outside that scope are unconfirmed. There is no help documentation, no rollout schedule, and no change to Search Console reporting yet.

Does this change anything for Google Ads campaigns?

Nothing has been announced about ads inside Dive deeper panels. The indirect effect is on remarketing: if organic Discover sessions decline, audience lists shrink, and Performance Max and Demand Gen campaigns leaning on site-visitor signals get thinner input. Watch list sizes in October rather than waiting for cost-per-acquisition to drift.

How will I know if Dive deeper is what cost me clicks?

Compare impressions against clicks in the Discover report. Impressions holding flat while CTR falls points at a surface change intercepting the tap. Impressions and clicks falling together points at eligibility or ranking instead. Without the September baseline, neither pattern is distinguishable from ordinary seasonality.

The practical read

Tests die, and this one may. The instinct survives it: Google keeps building surfaces that satisfy the topic before delivering the visitor, and what survives is content carrying something the summary cannot reproduce. For a Michigan company that is usually physical and local — the part in your hand, the job you finished in Dearborn last Tuesday, the measurement nobody else took. Auditing which pages actually qualify is the conversation to start with our Detroit team before October demand arrives.