Generative UI in AI Overviews: The Tool Page Playbook

AUG 19

date Google announced generative UI expanding from AI Mode into AI Overviews

68.01%

of US Google searches ended without a click, January to April 2026, per SparkToro

0.34%

share of searches that moved into AI Mode over that same window, per SparkToro

TL;DR · 60 SECONDSGoogle's generative UI builds custom visual layouts, interactive tools and simulations directly inside a search answer. It has been live in AI Mode since November 2025, where almost nobody saw it: SparkToro's clickstream analysis put AI Mode at 0.34% of US searches in early 2026. On August 19, 2026 Google announced it is expanding into AI Overviews, which is a completely different order of exposure. If your site earns traffic and pipeline from calculators, converters, comparison widgets or estimator tools, the surface those pages compete against just got the ability to generate a working version of them on the fly. This is a five-play, 90-day response: inventory and score, deepen past what a generated widget can do, wrap the tool in citable content, instrument displacement early, and move the conversion event off the calculation itself.

Every enterprise site has a handful of pages that outperform their word count by an embarrassing margin. The ROI calculator. The unit converter. The pricing estimator. The compatibility checker. They pull links without being pitched, they rank on queries the blog never touches, and they convert better than anything in the content library because someone using them has already decided they have a problem. On August 19, Google announced that AI Overviews is getting the ability to build those tools itself.

What generative UI in AI Overviews actually changes

Generative UI is Google's term for a search answer that assembles its own interface. Rather than returning text plus links, the system builds custom visual layouts, interactive tools and simulations on the fly, shaped to the specific question asked. Google published the research behind it and shipped it into AI Mode in November 2025 alongside Gemini 3. The examples Google demonstrated are the point: an interactive pH scale a user can manipulate through follow-up questions, a version plotting citrus fruits onto that same scale, and a mortgage calculator that compares different loan terms.

Read that last example again if you work in finance, insurance, real estate or anything with a rate in it. A mortgage calculator is not a hypothetical. It is one of the most linked-to, most-trafficked, most-monetized page types on the commercial web, and Google used it as the demo.

DATE	WHAT HAPPENED	SURFACE
November 2025	Generative UI launches alongside Gemini 3	AI Mode only
May 2026	Google says the capability will become free for everyone over the summer	AI Mode
August 19, 2026	Google announces expansion into AI Overviews, framed as a back-to-school rollout	AI Overviews, English globally
Coming weeks	Full AI Overviews availability, no completion date given by Google	AI Overviews

The reason this announcement matters more than the November one is entirely about denominator. SparkToro's analysis of Similarweb clickstream data across January to April 2026 found that only 0.34% of US searches transitioned into AI Mode. A capability confined to AI Mode was a research preview with a rounding error's worth of audience. AI Overviews is the surface most people actually encounter, and the same SparkToro work found that when AI Overviews appear, click-through rates fall by nearly 60%. Layer a working interactive tool on top of that and the click that was already unlikely becomes unnecessary.

THE DISTINCTION THAT MATTERSThis is not the usual zero-click argument about a snippet answering a question. A snippet quotes your page. A generated tool replaces the function of your page. Those require different responses, and the zero-click playbook most teams already run covers the first one, not the second.

The exposure math: which pages are in range

Not every interactive page is equally exposed, and the difference comes down to one question: could a competent system build a working version of this from public knowledge alone, in the time it takes to render an answer? If yes, you are exposed. If it needs your data, your inventory, your pricing, your customer's account, or a judgment call that is not in any textbook, you are not.

Single-variable converters (units, currency, dates)90%

Textbook-formula calculators (mortgage, BMI, compound interest)80%

Generic comparison tables of public specs65%

Multi-variable estimators with defensible assumptions40%

Tools that run on your proprietary data or inventory15%

Tools behind authentication or tied to an account5%

Displacement risk by tool type (our scoring rubric, not measured traffic)

That rubric is our scoring judgment, not a traffic measurement, and it is deliberately blunt. The top two rows are the uncomfortable ones because they are also where most SEO teams built their link magnets. A mortgage calculator implements a formula that has been public since before the web existed. There is no moat in the arithmetic. The moat, if you have one, is in the assumptions around it: current rates from your own desk, your underwriting thresholds, your regional tax treatment, your closing costs.

The bottom two rows are where the strategy points. A tool that consumes something only you have cannot be generated, because the generator does not have the input. This is the same structural argument that makes first-party data the durable asset in comparison content that wins citations: the format is copyable, the underlying data is not.

Run the scoring before you run anything else. Most teams discover that their exposure is concentrated in two or three URLs, which turns a frightening strategic question into a manageable project. It also occasionally surfaces the opposite finding, that a page everyone assumed was a traffic pillar has been flat for a year and is not worth defending at all.

The five plays

These run in order and they compound. Play 1 tells you how much of a problem you have. Plays 2 and 3 are the actual defense, and they are content and product work, not technical SEO work. Play 4 is the instrumentation that tells you whether the defense held. Play 5 is the structural change that makes the question much less scary next time a surface moves.

A note on who runs this. Every previous version of this problem, featured snippets taking definitional queries, then knowledge panels taking entity queries, then AI Overviews taking explanatory queries, was solvable inside the SEO team because the response was a content response. This one is not. Plays 2 and 5 require someone with actual domain expertise to say what the standard formula gets wrong, and someone with product authority to change what happens at the moment a result appears. If the exposure list from play 1 does not get in front of both of those people in week one, the work stalls at play 3 and you end up with beautifully marked-up pages that a generated widget still makes redundant.

1Week 1Inventory every interactive page and score its displacement risk

THE MOVES

Crawl the site and flag every URL containing a form, a slider, an input field, a calculator, a converter, a quiz or an embedded widget. Include the ones marketing forgot about and product built in 2022.

Pull twelve months of Search Console data for each one: impressions, clicks, average position, and the top twenty queries driving them. Separate branded from non-branded, because branded demand does not get displaced by a generated widget.

Pull referring domains per URL from your backlink tool. Tool pages are usually the top link earners on a commercial site, and a link profile concentrated in a page at risk of displacement is a specific, nameable exposure.

Score each page against the risk rubric: can this be rebuilt from public knowledge alone, or does it require your data, your pricing, your inventory or a defensible judgment call?

Rank the list by traffic times risk score, not by traffic alone. A high-traffic page at low displacement risk is not a project. A moderate-traffic page at high risk with 80 referring domains is the whole project.

DONE WHENYou have a ranked list of at most ten URLs, each with twelve months of query data, a referring domain count and a risk score, and the executive summary fits in three sentences.

2Weeks 2 to 3Go past the point where a generated tool stops being useful

THE MOVES

For each high-risk tool, write down what a generated version would do well: the core arithmetic, the obvious inputs, a clean result. Assume it does all of that competently, because it will.

Then write down every question a real user asks immediately after seeing that result. For a mortgage calculator that is closing costs, PMI thresholds, rate lock windows, what happens with a variable income, and how the number changes in a specific state.

Build the second question into the page. The generated tool answers the first question and stops, because it has no source for the rest. Owning the follow-up is the defensible position, and it is where the buying intent actually lives.

Replace generic defaults with your own current data wherever you legitimately have it: live rates, real inventory, actual average timelines from your book of business, regional variation you can substantiate.

Add the constraint layer that formulas ignore: eligibility rules, minimums, regulatory limits by jurisdiction, the situations where the standard calculation gives a misleading answer. This is expertise, it is not on the open web in structured form, and it is the reason someone stays.

DONE WHENSomeone who already has the number from an AI Overview still has a concrete reason to open your page, and you can state that reason in one sentence without using the word better.

3Weeks 3 to 5Wrap the tool in content the engines have to cite

THE MOVES

Give every tool page real prose above and below the widget: what the calculation assumes, where the inputs come from, when the output is wrong, who should not rely on it. Naked tool pages have almost nothing extractable and get cited accordingly.

Publish your methodology as a named, dated, linkable section. A stated methodology is one of the few things a generated tool structurally cannot produce, and it is what turns your page into a source rather than a competitor.

Mark the page up properly. SoftwareApplication, HowTo, FAQPage and Dataset where each genuinely applies, plus author and organization markup, following the same rules as any other [schema an AI engine reads](/blog/schema-ai-engines-read).

Add a results interpretation section with worked examples at three or four realistic input sets, each with a short explanation of what the output means in practice. This is highly extractable, highly citable, and it is the part users screenshot.

Link the tool into the surrounding topic cluster in both directions, so the page stops being an orphaned utility and starts being the anchor of a subject you demonstrably own.

DONE WHENThe page would still be worth citing with the interactive element removed, and you can point to at least three self-contained passages an engine could lift and attribute.

4Weeks 4 to 6Instrument displacement before it reaches the revenue number

THE MOVES

Baseline each tool page now: impressions, clicks, click-through rate by query, and assisted conversions. Do it before the AI Overviews rollout completes, because a baseline taken afterward is worthless.

Split the reporting between impressions and clicks rather than watching sessions. The displacement signature is stable or rising impressions with falling click-through, and a sessions chart hides it completely.

Track the specific query set most at risk, the generic formula queries, separately from branded and long-tail queries. Blending them delays detection by about a quarter.

Set an alert threshold in advance and write down what you will do if it trips, so the response is a decision made calmly in August rather than an argument held in November.

Check monthly whether AI Overviews with an interactive element are appearing for your priority queries at all. The rollout has no announced completion date, so presence is a variable to watch, not an assumption to make.

DONE WHENYou can answer the question have we lost tool traffic to generated interfaces with a chart and a date, instead of a theory, and the alert has a named owner.

5Weeks 6 to 12Move the conversion event off the calculation itself

THE MOVES

Audit what actually converts on each tool page today. In most funnels the calculation is not the conversion, it is the qualifier, and the conversion is the next step someone takes once the number scares or reassures them.

Make that next step available at the moment of result, not at the bottom of the page: the saved scenario, the shareable output, the personalized breakdown, the specialist conversation.

Offer something the answer surface cannot: save this, compare it against my last three, send it to my team, apply my actual account data. Persistence and personalization are outside what a generated widget can do.

Where the tool is genuinely commodity arithmetic, stop defending it as a traffic asset and start treating it as a conversion utility for people already on the site. Keep it, maintain it, and stop investing in ranking it.

Reallocate the effort saved into the tools that consume your proprietary data, since those are both harder to displace and better qualified when they convert.

DONE WHENA visitor who gets the number without visiting your site still has a reason to visit your site, and your team can name what that reason is per tool.

Sequencing the plays across a quarter

The sequence matters more than the speed. Teams that start with play 3, the content wrap, because it feels like familiar work, end up doing it on pages that were never at risk. Teams that start with play 5 rebuild a funnel before knowing whether it was under threat. Inventory first, always.

WEEKS	PLAY	OWNER	OUTPUT
1	Inventory and score	SEO lead	Ranked exposure list, at most ten URLs
2 to 3	Deepen past the generated version	Content plus subject expert	Follow-up layer live on the top two pages
3 to 5	Wrap in citable content	Content plus engineering	Methodology, schema, worked examples shipped
4 to 6	Instrument displacement	Analytics	Baseline captured, alert with a named owner
6 to 12	Move the conversion event	Product plus demand gen	Save, share or personalize live at result

Twelve weeks is the honest estimate for a mid-market site with three or four exposed tools, and it assumes the subject expert in play 2 gives you real time rather than a review pass. Enterprise sites with a dozen tool pages across several business units should not try to run all of them at once. Take the top two by traffic times risk, run the full sequence, measure the result, then use that result to fund the rest. A completed defense on two pages with a number attached buys more internal budget than a half-finished program across twelve.

Plays 3 and 4 overlap deliberately. The instrumentation baseline has to be captured while the rollout is still incomplete, and the content work takes longer than a week, so running them in parallel is the only way to get a clean before-and-after. If you have to drop one play this quarter, drop play 5 and carry it into the next one. It is the highest-value play and the one that survives being late.

How to measure displacement before revenue notices

Displacement has a specific signature and it is easy to miss because it does not look like a ranking problem. Your position holds. Your impressions hold or rise. Only the click-through falls. Anyone watching a sessions chart or a rankings dashboard will see nothing for two months and then see a revenue conversation they cannot explain.

1Impressions flat, clicks fallingThe core displacement signature. If position is stable and impressions are stable while click-through drops, something in the result is satisfying the query before the click. That is the number to alert on.

2Formula queries decoupling from branded queriesWatch generic queries like how much can I borrow separately from your brand plus calculator. Generic falling while branded holds is displacement. Both falling is a different problem, usually a ranking one.

3Time on page rising while sessions fallCounterintuitive, and a good sign. The casual users who only wanted the number stopped coming, and the ones still arriving are further down the funnel. Check conversion rate per session before treating the traffic loss as damage.

4Referring domain growth stallingTool pages earn links passively because people cite them as a resource. If a generated interface answers the question in the result, the writer who would have linked to you never lands on you. This shows up months later and is the most expensive part to rebuild, so track it as part of how citations get earned rather than as a link building metric.

One caution on all four signals. The AI Overviews rollout has no published completion date and Google has not said which query categories will get interactive elements first. Any change you see in the next eight weeks could be the rollout, could be the ordinary volatility that has followed the August spam update, and could be seasonality, since a back-to-school rollout lands in the same weeks that half of these tools see their annual demand shift anyway. Hold the interpretation loosely and require a couple of months of data before you rebuild a funnel on the strength of one chart.

What not to do about generative UI in AI Overviews

Do not block the crawlers. It is the reflexive response to any displacement story and it is almost always wrong. Blocking removes you from the answer without removing the answer, so the generated tool still appears, built from other public sources, and now your page is not in the citation list beside it. You have given up the visibility and kept the displacement.

Do not gate a tool that was earning links purely to force a click. It converts a passive link-earning asset into a friction point, and the demand simply routes to a competitor whose version is open. Gating works when what is behind the gate is genuinely worth an email address, which is a much higher bar than a calculator most people can approximate in their head.

Do not rush a rebuild of a page you have not scored. We have watched teams spend a quarter rebuilding a converter that was never at risk, because it was the page everyone had an opinion about, while an estimator carrying 90 referring domains sat untouched. And do not treat this as a reason to stop building tools. Tools that consume your own data are among the strongest assets you can own precisely because they cannot be generated, which is the same argument for investing in a conversion rate optimization program on the pages you keep rather than chasing replacement traffic elsewhere.

Where to start on Monday

Open Search Console, filter for your calculator or tool URLs, set the comparison to the last 28 days against the previous 28, and look only at the impressions and click-through columns. That takes ten minutes and it tells you whether this is a live problem or a planning exercise. Then run the crawl for play 1 and get the ranked exposure list in front of whoever owns the number those pages produce.

The strategic read underneath all five plays is simple enough to say in one line. Search surfaces keep absorbing the commodity layer of everything, and every time they do, the value moves to whatever sits underneath that layer: your data, your judgment, your constraints, your customer's context. Generative UI is that pattern arriving at interactive content, which happened to be the last content category that felt safe from it.

If the exposure list comes back ugly, that is a generative engine optimization and site build problem in roughly equal measure, and it is particularly acute in fintech, where the demo Google chose to show off is somebody's highest-converting page. Score first. Defend what has a moat. Stop paying rent on what does not.

Sources● LIVE

Google Research - Generative UI: a rich, custom, visual interactive user experience for any prompt
https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/
Search Engine Journal - Google expands generative UI beyond AI Mode into AI Overviews, August 2026
SparkToro with Similarweb clickstream data - US zero-click search analysis, January to April 2026

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

Schedule a free consultation

Tyler TruffiMANAGING PARTNER, SOMETHING INC.

Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

The generative UI playbook: defending tool pages in AI Overviews

What generative UI in AI Overviews actually changes

The exposure math: which pages are in range

The five plays

Sequencing the plays across a quarter

How to measure displacement before revenue notices

What not to do about generative UI in AI Overviews

Where to start on Monday

See where you are cited today

Keep reading

Let us be the last SEO agency you ever work with