AI Mode Stopped Showing Search Results. It Started Building Apps.
Gemini 3 in AI Mode generates interactive tools, calculators, and simulations on the fly. Here's what that does to brand visibility, and what to do.
AI Mode Stopped Showing Search Results. It Started Building Apps.
Ask Google AI Mode whether to refinance your mortgage and it doesn’t return a list of articles anymore. It writes a loan calculator. Live. In the response. With your numbers pre-filled and a sliding rate input you can drag to see what happens.
That’s not a feature tweak. That’s a different product.
Gemini 3 launched in AI Mode on November 18, 2025 and got a broader rollout at Google I/O 2026. The headline upgrade isn’t reasoning. It’s that AI Mode now decides, query by query, whether to answer with text, a table, a chart, a calculator, a physics simulation, or a custom interactive tool it codes on the spot. AI Mode crossed one billion monthly users about a year after its debut, and queries keep doubling every quarter.
The thing most teams haven’t grasped yet: when the answer is a calculator, there is no paragraph to cite. The whole citation game shifts. Your brand isn’t competing for a quoted sentence anymore. It’s competing to be a default value, a comparison row, an option in a dropdown that Google generated in 800 milliseconds.
What “generative UI” actually means here
Google’s been careful with the term. The internal name is generative UI. The public-facing pitch is “dynamic visual layouts with interactive tools and simulations.” Both mean the same thing: Gemini 3 decides what surface the answer needs, then renders it.
Two examples Google led with:
- A user learning about the three-body problem gets an interactive physics simulation where they can drag mass values and watch gravitational orbits play out.
- A user comparing mortgage options gets a custom-built loan calculator inside the response with side-by-side scenarios.
A third class is rolling out through 2026: persistent “mini apps” generated inside Search using live web data. Wedding planners that pull from your calendar. Fitness trackers that mix weather data, local gym hours, and your routine. Google says these are assembled in real time by Gemini 3.5 Flash from live web sources.
Notice what’s common across all of those examples. The user doesn’t leave Google. They don’t even scroll past the result. They interact with the result itself. That’s the part that should make every brand strategy lead pause.
The citation surface is shrinking
For two years the GEO playbook has been “earn the cited paragraph.” Write the clean H2. Structure the FAQ schema. Get the quote pulled into the answer. We’ve written about this approach more than once — see our piece on how AI Mode pulls content at the chunk level.
That playbook still works for one specific kind of query: informational, single-intent, text-shaped. “What is generative UI?” still produces a text answer with cited sources.
But text answers are no longer the default for every query type. When the model decides a comparison table works better, the citation surface collapses to whatever it can fit in row labels. When it decides a calculator works better, citations might not appear at all — they get pushed into a “sources” tray most users never open.
That changes what visibility means. Two outcomes can coexist on the same query:
- Your page is the best-cited source for the text portion of the answer.
- Your brand doesn’t appear anywhere in the generated calculator that 80% of users actually interact with.
You can win the citation and still lose the query.
Different queries, different visibility surfaces
This is the part most GEO frameworks haven’t been updated for. The optimization target depends on what Google decides to render. Here’s how we’ve started bucketing it internally at RivalHound:
| User intent | What AI Mode renders | Where your brand can appear | Optimization target |
|---|---|---|---|
| ”What is X” | Text + inline citations | Cited paragraph, sources tray | Chunk-level content clarity, schema |
| ”X vs Y” | Comparison table | Row label, spec column, footnote citation | Structured comparison data, product schema |
| ”How much for X” | Interactive calculator | Default value, option in dropdown, embedded brand row | Pricing schema, product feeds, structured specs |
| ”How does X work” | Simulation or diagram | Entity in the model, labeled component | Wikipedia-class authority, named-entity recognition |
| ”Plan X for me” | Mini app or dashboard | Data source feeding the app, suggested option | API-accessible data, structured listings |
| ”Should I buy X” | Ranked list with filters | Filtered-in row, badge, review citation | Reviews, third-party authority, comparable specs |
The right-hand column is what most teams haven’t built for. Schema markup that helped you win a featured snippet doesn’t necessarily make you a default in a generated tool. Different surface, different requirements.
What this breaks in your measurement stack
Citation rate, the metric most AI visibility tools lead with, starts to undercount your real exposure once generative UI scales. A user who drags the slider on an AI Mode mortgage calculator may have seen your brand as the default rate or a comparison row, but no citation event fired. Your dashboard says “not visible.” Reality says you were the answer.
Share of voice gets stranger across response types. Your brand can hold 40% SOV on text answers in a category and 5% SOV on the comparison tables Google renders for the same intent. Roll those into one number and the headline metric hides the actual exposure.
The bigger shift is in how you classify queries. It used to be enough to know “is my brand cited for this query?” Now you also need to know what surface AI Mode rendered. The same query string can produce a text answer for one user and a calculator for another, depending on Gemini’s read of intent and what context it has on the session. We touched on a related problem in our analysis of citation divergence across AI platforms, but generative UI pushes the variance one level deeper. It’s no longer just across platforms. It’s across response types on the same platform for the same query.
What to actually do this quarter
Stop optimizing for the cited paragraph as if it’s the only outcome. Start treating your product as a structured entity that needs to be machine-readable in five or six different surface contexts.
A short list of what’s worth doing this quarter:
- Audit your top 50 queries for response-type vulnerability. Run each through AI Mode logged out, then logged in. Note which render text, which render tables, which render calculators. Anything that renders something other than text is a query where your existing citation strategy is incomplete.
- Make your data feedable. Pricing in structured schema. Product specs in tables. Comparison fields in Product schema with comparable units. If Gemini wanted to build a calculator for your category, could it pull your numbers without guessing? If no, fix that first.
- Watch the third-party data sources Gemini reaches for. If a mortgage calculator pulls average rates from a specific aggregator, your visibility inside that aggregator now matters as much as your own homepage. The handful of domains AI Mode actually trusts for category data is a separate investment from your own content.
- Add a response-surface dimension to your visibility tracking. Not just cited or not cited. Track which surface your brand appeared on. RivalHound users have started tagging queries by response type so trend lines stay meaningful as more queries flip to generative UI.
- Don’t abandon text optimization. It still drives informational intent and feeds the sources tray inside generative UI. Most generated tools still cite something. Being in the citation pool the model draws from is how you get into the rendered surface in the first place.
The uncomfortable part
The honest read on this: Google is moving search up a layer of abstraction faster than most GEO tooling can keep up with. Citation tracking was the first metric that mattered in AI search. It will not be the last, and it probably isn’t sufficient by itself for another twelve months.
The brands that get this right won’t be the ones with the cleanest content. They’ll be the ones whose product data is structured well enough that an AI assistant can use it without asking. That’s a different kind of investment than blog posts. It’s product data hygiene, schema discipline, and being authoritative in the third-party sources Gemini reaches for when it builds a tool.
Search has gone from a list of pages to a paragraph of synthesis to a thing — a calculator, a planner, a comparison grid — that users tap and discard. Your brand has to live as a fact inside that thing, not as a page hoping to be cited next to it.
RivalHound tracks your brand’s visibility across ChatGPT, Google AI, Perplexity, and more. Start monitoring to see where you stand.