Google builds the tool instead of linking to it
On 19 August 2026 Google wrote that generative UI "have started rolling out as well in AI Overviews". A Google Research paper from February says a generated interface was preferred over the top Google result in 90% of comparisons — and lost to a human-built page, 35.3% against 50%. Meanwhile our own free speed tool has been shown 271 times since 8 July and clicked zero times. Both facts matter, and they point in opposite directions.
What changed. Generative UI — Google building a small interactive thing inside the answer instead of listing pages — left AI Mode and started appearing in AI Overviews. Google's own wording, published 19 August 2026: it "have started rolling out as well in AI Overviews". Globally, in English, with no completion date given.
Why it isn't everyone's problem. The 90% figure everyone is quoting was measured against the top search result. In the same table, pages built by human experts beat the generated interface 50% to 35.3%. Ranking first is what is exposed. Being good is what is not.
What we can add. We run a free page-speed tool in three languages. Google has shown it 271 times since 8 July 2026 and it has never been clicked once — not in English, not in Italian, not in Romanian. That is the other half of this story, and nobody publishes it about themselves.
- The one sentence Google actually published, and the four words in it that are doing the work.
- Where the 90% comes from, how many comparisons it rests on, and the number sitting next to it that nobody quotes.
- How to tell in ten minutes whether your own tool page is exposed or simply invisible — two problems that look identical in a traffic chart.
- Why the Italian version of our tool is, for now, in a different position from the English one.
There is a specific kind of page that has quietly paid for a lot of small websites: the one that does one thing. A mortgage calculator. A unit converter. A speed test. A VAT estimator. You build it once, it ranks for a hundred long-tail phrasings of the same need, and it keeps working while you sleep. Search Engine Journal put the distinction better than we would have on 21 August 2026: "Summaries compete with the pages people read, while generated interfaces compete with the pages people use."
That sentence is the whole reason the 19 August announcement is worth twenty minutes of your attention rather than a bookmark. Two years of argument about AI in search have been about reading — the answer above the links, the citation, the click that never happened. Generated interfaces are about using. Different pages, different owners, different defence.
And then there is the part we can only say because it is embarrassing. We own exactly the kind of page this is about. We can tell you precisely what it has earned us in organic search, because the number is zero, and it has been zero since the day it went live.
What a generated interface actually is
A generated interface is a small, working piece of software that Google's model writes on the spot, inside the answer, for one specific question — a calculator, a slider, a simulation — instead of sending you to a page that already contains one. That is the definition, and it is worth having in one sentence because the marketing language around it ("generative UI", "dynamic view", "custom experiences") hides how concrete the thing is.
Google's own example is not ours and is not hypothetical. From the Gemini 3 announcement of 18 November 2025, read in full for this article: "Or perhaps you're researching mortgage loans: Gemini 3 in AI Mode can make you a custom-built interactive loan calculator directly in the response so you can compare two different options and see which offers the most long-term savings."
Read that with the eyes of somebody who owns a mortgage calculator. The example Google chose to illustrate the feature is, verbatim, a category of website that exists in every market, in every language, and that has been a reliable source of traffic for twenty years.
A page that explains a calculation competes with a summary. A page that performs the calculation competes with a generated interface. If your page does both, it has two different exposures and they move at different speeds.
The sentence Google published on 19 August
The announcement did not arrive as an announcement. It arrived inside a post about back-to-school study features, published on the Google blog on 19 August 2026, among practice quizzes and Lens homework help. The load-bearing sentence: "Generative UI capabilities have launched globally in English in AI Mode and have started rolling out as well in AI Overviews."
Four things follow from those words, and it is worth being pedantic about each, because the difference between them is the difference between planning and panicking.
"Globally in English." Not globally. English. The same post is explicit about the contrast elsewhere — Notebooks in AI Mode are described as rolling out "across more than 180 countries in English, with local language support to follow in the weeks ahead". Google knows how to say "and then other languages" when it means it. For generative UI in AI Overviews, it did not say it.
"Have started rolling out." Present progressive, no completion date. We looked for one and did not find it in that post; the reporting on 19 and 21 August did not have one either. So an operator cannot plan against a date, only against a direction.
AI Mode versus AI Overviews. This is the part that changes the size of the problem. AI Mode is a destination you choose. AI Overviews is what appears above ordinary results for people who chose nothing. At Google I/O on 19 May 2026, Google put AI Overviews at "over 2.5 billion monthly active users" and said AI Mode had "surpassed 1 billion monthly active users". Moving a feature from the second to the first is roughly a 2.5× change in who can see it, on a population that never opted in.
It is not new technology, only newly placed. Generative UI shipped in AI Mode in November 2025. At I/O in May, Google said Search would "build custom experiences just for your individual questions, like dynamic layouts and interactive visuals". The thing has been visible for nine months. What changed on 19 August is the surface it appears on.
The 90%, and what it was measured against
Every write-up of this story quotes a 90%. It is a real number with a real source, and it is being quoted in a way that inverts its meaning.
The source is a Google Research paper, "Generative UI: LLMs are Effective UI Generators", arXiv:2604.09577v1, submitted on 24 February 2026. We read the paper rather than the coverage, and the shape of the experiment matters more than the headline. The team took 100 prompts from LMArena, excluded 8, and collected pairwise human preference on the remaining 92, "sending each result to 2 raters" on a three-point scale: left, right, or neutral. Each prompt produced five candidates: a site built by a human expert, the top Google Search result for that query, the generated interface, and two plain text outputs.
| Generated interface compared with | Preferred | What that means for a page owner |
|---|---|---|
| The top Google Search result | 90% generated interface | Ranking first is not protection |
| Plain LLM markdown output | 82.8% generated interface | Not relevant to page owners |
| A site built by human experts | 50% human vs 35.3% generated | Being built well is protection |
Those two rows are the article. The generated interface beats whatever happened to rank first nine times out of ten. Put in front of a page somebody actually designed for the task, it loses — and not narrowly. The paper's own abstract concedes the point: the generated results "are worse than those crafted by human experts", and are only "at least comparable in 50% of cases".
The rows do not add up to 100 because raters had three options, not two — left, right, or neutral. Against the top search result, 90% plus the 6.7% that went the other way leaves 3.3% neutral. Against the expert-built pages, 50% plus 35.3% leaves 14.7%. Raters hesitated more than four times as often — 14.7 against 3.3 — when the thing across the table had been built by someone.
And there is a condition nobody quoting the figure mentions. The method section states plainly that "Generation time is not a factor in the evaluation". In the limitations section the same authors write that generation "can often take a minute or two". So the 90% was measured in a world where the thing appears instantly. On a results page, a minute is an eternity. That will improve; today it is a real constraint on where this can be deployed.
One more thing worth saying out loud, because nobody quoting the 90% says it: this is Google Research evaluating a Google system against a Google search result, on prompts from a chatbot arena rather than on real search queries, with two raters per pair. It is a legitimate paper and we are treating the number as real. It is not a measurement of what happens to live traffic.
271 impressions, zero clicks: our own tool page
Here is the part we can only contribute by owning up to something.
We publish a free page-speed checker at cittago.com/tools/page-speed, in English, Italian and Romanian. It is exactly the class of page the paper and the announcement are about: it performs a measurement rather than describing one. It has been indexed since early July.
The figures below are from the Search Console property for cittago.com, pulled on 22 August 2026. Google reported data for this page group from 8 July to 21 August 2026, which is every day it has existed in the index.
| Page | Impressions | Clicks | Average position |
|---|---|---|---|
| /it/tools/page-speed/ | 161 | 0 | 66.2 |
| /tools/page-speed/ | 92 | 0 | 44.9 |
| /ro/tools/page-speed/ | 18 | 0 | 50.1 |
| /tools/email-exposure/ | 5 | 0 | 21.0 |
| All four tool pages Google showed at all | 276 | 0 | — |
Three of those rows are the same tool in three languages: 161 plus 92 plus 18 is 271, which is the total Google reports for that URL group over the period. Zero days with a click. Not one, across 44 days with reported data and 46 distinct queries that Google was willing to name.
In the tighter 28-day window of 25 July to 21 August, the same tool took 217 impressions at an average position of 56.6. The demand is not imaginary and it is not tiny for a page of that size: people are typing "test sito web google", "page speed test online", "velocità sito", "weboldal sebesség teszt". Google is putting our page in front of them, on page five.
If a generated interface had appeared today for every one of those 46 queries, our loss would have been exactly nothing. We were not earning anything to lose. Generative UI is a threat to pages that convert impressions into clicks. It is irrelevant to pages that do not — and most small-site tool pages are the second kind.
There is a limit on that number we should state rather than let you discover: it counts organic clicks from Google Search only. People do reach the tool from our own navigation and from links we send. Zero here means zero from Search, not zero visitors.
And a second, more useful reading. The English page sits at position 44.9 while the Italian one sits at 66.2, on a larger volume of Italian demand. We already know part of why: on 21 August we found that the Italian and Romanian tool pages were declaring the English URL inside their own structured data, so each localised page was telling Google, three times over, that the canonical thing here was the English one. That was fixed the same day. The positions above are from before the fix, and we will publish what happened by the end of the month whichever way it goes.
What you need before you can judge your own exposure
The whole point of the two numbers above is that a traffic chart cannot tell you which situation you are in. A tool page at position 3 that is about to be replaced by a generated interface and a tool page at position 57 that nobody has ever visited both look like a flat line. Four things separate them, and you can gather all four in an afternoon.
- Search Console, filtered to the tool URL alone. Not the site total, not the folder. The page. You want impressions, clicks and average position for that one URL, and you want the queries it actually catches — which, as we found on our own service pages on 19 August 2026, are frequently not the queries you assumed.
- The language of the demand. Export the queries and sort them by language, not by volume. English demand is exposed to this feature today. Everything else is exposed to it later, on a timetable Google has not published.
- An honest label for the page: does it explain, or does it perform? A page that explains how to read a Core Web Vitals score competes with a summary. A page that measures your site competes with a generated interface. Most pages that feel like tools are actually the first kind wearing the second kind's clothes.
- What happens after someone uses it. If the answer is "they get a number and leave", the page was never an asset — it was a courtesy. If the answer is "some of them ask us to fix what the number showed", you have something a generated interface structurally cannot copy, because it ends at the number.
That last one is the only defence in this article that does not depend on Google. It is also the reason we keep the tool up despite the zero: the measurement is not the product. Deciding which of the fixes is worth buying is the product, and that is what our page speed optimisation work actually is. If you have already got a number from any free checker and still do not know which three things to change first, the number was never the hard part.
What none of this establishes
The strongest thing we can do with a story this loud is be precise about where the evidence stops.
- Nobody has shown a tool page losing traffic to a generated interface. Not Google, not the paper, not us. The paper measures stated preference in a lab; the announcement describes a rollout. The causal claim everyone wants — "this took my clicks" — has no published evidence behind it yet.
- The 90% is not "90% of searches". It is the preference share across 92 prompts from LMArena, two raters each, when the two are shown side by side. It says nothing about frequency, only about preference when both are put side by side.
- Google has not said which queries trigger it. What the Gemini 3 post publishes is a condition, not a list: “When the model detects that an interactive tool will help you better understand the topic, it uses its generative capabilities to code a custom simulation or tool in real-time and adds it into your response.” We searched that post and the 19 August one for “opt out”, “turn off”, “disable” and “setting”. None of the four appears in either.
- There is no stated schedule for languages other than English — and we checked the same post for one, because that post was perfectly willing to give a timetable for a different feature two paragraphs away.
- We have not seen a generated interface on our own queries. We do not run automated queries against Google; our own IP gets blocked for it, and results are personalised anyway. Everything here about the feature comes from Google's own documents and from the paper, not from a screenshot we could not honestly reproduce.
- Our zero is not a benchmark. It is one small site, four months old in this index, in a market where our brand means nothing yet. A tool page with an audience behind it will look nothing like ours.
The words, in a table
Four terms are being used interchangeably in the coverage, and two of them mean different things.
| Term | What it is | Where it appears |
|---|---|---|
| AI Overviews | The generated summary above ordinary results, shown to people who did not ask for it | Standard Google Search |
| AI Mode | A separate conversational search surface you choose to enter | A tab, opted into |
| Generative UI | The capability of building a visual layout, tool or simulation on the fly | Since Nov 2025 in AI Mode; since 19 Aug 2026 rolling into AI Overviews |
| Generated interface | The individual artefact that capability produces for one question | Inside either surface |
The practical consequence of the table: a report that says "we saw generative UI in AI Mode in March" is not evidence about your exposure, because AI Mode was never where the volume was.
Why the Italian page is not in the blast radius yet
Our Italian tool page carries 161 of our 271 impressions — the majority of the demand, in the market where we sell. On the reading above it is also, for now, the safer of the two, and that is an unusual thing to be able to say.
Generative UI in AI Overviews is English, by Google's own sentence. Italian queries such as "test sito web", "misurare velocità sito web" and "controllo velocità sito" are not served by it today. The English page, at position 44.9 on phrasings like "page speed test online" and "free web page speed test", sits in the language where the feature is live.
This is a window, not a reprieve. Anyone whose non-English tool page still earns clicks has some months to make it something other than a number dispenser. What we would do with that window is unglamorous: make the tool produce something the model cannot — a comparison against last month, a record of your own site over time, a result that depends on data only you and we have. A generated interface starts from nothing every time it is asked. It has no memory of your site. That is the only structural advantage on the table, and it is a real one.
If you want the underlying mechanics of the score itself, we wrote about why the same page returns 68 on mobile and 96 on desktop in why PageSpeed says 68 on mobile and 96 on desktop. And if the broader question is whether AI systems can see and cite your pages at all, the firewall side of it is in does ChatGPT respect robots.txt.
Where you stand, in three thresholds
First threshold — your tool page has impressions and no clicks. This is us. Generative UI is not your problem and worrying about it is a way of avoiding the real one. Your page is not being taken; it is not being found. The work is position and the phrasing you rank for, in that order, and it is ordinary work with a known method. Nothing announced on 19 August 2026 changes what you should do on Monday.
Second threshold — your tool page ranks on page one for English queries and earns clicks. You are in the population the paper is about, and give it an hour and one question: what does someone get from your page that they could not get from a thing generated on the spot? If the honest answer is "the same number, in our colours", the traffic was borrowed and the loan is being called in. Build the memory in — history, comparison, an account, an export — or accept that the page is a marketing cost rather than a channel.
Third threshold — your tool page is a lead source with a service behind it. You are in the best position of the three and you should still change something. Shorten the path between the measurement and the conversation, because that path is the whole asset. A generated interface will happily give someone a number. It will not tell them which of the four fixes to buy, in what order, on their budget. Make sure your page does, on the same screen as the number.
Questions nobody has actually asked us
Not one of these arrived in an inbox. They are the questions we had to answer for ourselves before the sections above could be written, kept here in the order they came up.
What is generative UI in Google Search?
Generative UI is Google's capability to build a visual layout, an interactive tool or a simulation on the spot, inside the answer, instead of listing pages that already contain one. Google's own illustration, from its Gemini 3 post of 18 November 2025, is a custom-built interactive loan calculator generated directly in the response.
Is generative UI live in AI Overviews right now?
It is rolling out. Google's wording on 19 August 2026 was that the capabilities "have launched globally in English in AI Mode and have started rolling out as well in AI Overviews". No completion date was published in that post, and none appeared in the reporting on 19 or 21 August 2026.
Does it work in languages other than English?
Not according to Google's sentence, which says "globally in English". That is a deliberate wording: the same post promises, for a different feature, rollout "across more than 180 countries in English, with local language support to follow in the weeks ahead". For generative UI in AI Overviews, no such timetable was given.
Does a generated interface really beat the top search result 90% of the time?
In one specific test, yes. Google Research's paper arXiv:2604.09577v1, submitted 24 February 2026, put a generated interface against the top Google result across 92 prompts from LMArena, two raters per result, and the generated interface was preferred 90% of the time. Generation time was excluded from that evaluation.
So should I delete my calculator or tool page?
No. In the same table, sites built by human experts were preferred over the generated interface 50% to 35.3%. What the study exposes is the advantage of merely ranking first, not the value of a page someone designed for the task. A well-built tool page still won.
How do I tell whether my tool page is threatened or just invisible?
Filter Search Console to that single URL and read three numbers: impressions, clicks and average position. A page on page one with real clicks is in the exposed population. A page with impressions and no clicks has a discovery problem that generative UI neither caused nor changes.
Can I opt out of having a generated interface shown for my queries?
Not that we could find. Google publishes a condition rather than a list: when the model detects that an interactive tool will help you better understand the topic, it codes one in real time and adds it into the response. We searched that post and the 19 August one for 'opt out', 'turn off', 'disable' and 'setting', and none of the four appears in either.
Is this the same thing as AI Overviews taking clicks from articles?
No, and the distinction matters. Search Engine Journal put it this way on 21 August 2026: "Summaries compete with the pages people read, while generated interfaces compete with the pages people use." A page that explains a calculation faces the first problem; a page that performs it faces the second.
What can a tool page do that a generated interface cannot?
Remember. A generated interface starts from nothing each time it is asked and ends at the answer. A page can hold history, compare this month with last, keep an account, export a record, and connect the measurement to whoever fixes what it found. That is a structural difference, not a stylistic one.
Has anyone measured traffic actually lost to generative UI?
Not publicly, as far as we can find. The Google Research paper measures stated preference in a controlled comparison; the 19 August 2026 post describes a rollout. Neither is a measurement of live traffic, and we have not seen a published dataset showing a tool page losing clicks to a generated interface.
Last updated: 22 August 2026. The rollout sentence is quoted from Google's post "New back-to-school study tools in Search", dated 19 August 2026, read in full on this date; the loan-calculator sentence is quoted from the Gemini 3 in AI Mode post of 18 November 2025, retrieved with a real browser on this date and reproduced verbatim. The preference figures, the sample of 92 comparisons and both stated limitations come from arXiv:2604.09577v1, submitted 24 February 2026 by authors at Google Research; we read the paper, not the coverage of it. The user counts for AI Overviews and AI Mode are from Google's Google I/O 2026 post of 19 May 2026. The framing sentence about pages people read versus pages people use is quoted from Search Engine Journal, 21 August 2026. All Search Console figures are for the cittago.com property, queried on 22 August 2026; Google reported tool-page data only from 8 July 2026 onward, and clicks are organic Google Search clicks, not total visits. We will update this page when Google publishes a completion date for the AI Overviews rollout, when it states a timetable for languages other than English, or when our own tool page records its first click.
Sources: Google — New back-to-school study tools in Search (19 August 2026) · Google — Gemini 3 in AI Mode (18 November 2025) · Generative UI: LLMs are Effective UI Generators, arXiv:2604.09577 (24 February 2026) · Google Research — Generative UI · Google — Sundar Pichai at I/O 2026 (19 May 2026) · Search Engine Journal — Google's Generated Interfaces Could Compete With Tool Pages (21 August 2026)
