A friend of mine, a cabinet maker in the Marche region, asked me a question last week that I have been turning over in my head ever since. He said: my page is being cited inside Google now. I see it when I search. But the visits, the actual visits — they are dropping. How can both be true?
It is a question that, until very recently, would have seemed contradictory. More visibility, less traffic. Madness. And yet, the truth is, both are happening to a great many small businesses right now, and the cause is something quieter and more structural than most people realise.
The cause has a name. It is Gemini 3 Deep Think, the reasoning model that Google has been quietly extending across AI Overviews and AI Mode throughout the spring of 2026. And the shift it has introduced is, in my opinion, the single most consequential change to organic visibility since the launch of AI Overviews itself.
Allow me to explain — slowly, because the technical part matters less than the consequence.
A different way of reading the web
Up until late last year, the AI Overview block that appears at the top of a Google search behaved, more or less, like a very fast and very polite summariser. You would type a question. The model would scan a handful of top-ranking pages, pull a few sentences, stitch them together, and present the result. The citations underneath were, in essence, a list of the pages it had borrowed from.
The truth is, this was a fairly shallow operation. The model did not really think about what it had read. It compressed.
Gemini 3 Deep Think behaves differently. Before answering — and especially for questions that contain any element of judgement, comparison, or trade-off — it now allocates what Google calls thinking budget. This is, in plain language, additional compute time the model spends reasoning about the question before it begins to write. It considers multiple possible answers. It checks them against the evidence it has gathered. It rejects the ones that do not hold up. And only then does it produce the citation list.
The practical consequence is that the citations have become, let's say, more selective. The model is no longer just choosing pages that mention the query terms. It is choosing pages whose reasoning it can follow, verify, and reuse.
And this — this is precisely where the small business loses.
Why structure beats keyword density now
There is an old habit, deep in the SEO world, of stuffing pages with the phrasing one expects users to type. A page about, say, knife sharpening would dutifully repeat "knife sharpening service" and "professional knife sharpening near me" until the prose itself began to sound strange. This worked, for a long time, because the ranking algorithms were essentially counting matches.
A reasoning model does not count matches. It reads structure. It asks: does this page lay out a problem, examine it from more than one angle, and arrive at a defensible conclusion? Or does it merely repeat the question back to me in slightly different clothes?
Naturally, the second kind of page is now being quietly demoted in the AI Overview citation list — even when it still ranks reasonably well in the classic blue links below. We have observed this pattern across roughly two hundred small-business sites we audit on a rolling basis: pages that hold position seven, eight, nine in the traditional results, but which are no longer being cited above the fold by the AI Overview itself. The keyword presence is fine. The reasoning surface is thin.
This is, I think, the cabinet maker's problem in a nutshell. His page about how to choose a kitchen cabinet finish lists the finishes. It does not weigh them against one another. A reasoning model has nothing to chew on.
What "citable reasoning" actually looks like
When I say a page must be reasoned with, I do not mean it must become an academic paper. The opposite, really. The pages that Gemini 3 Deep Think cites most readily share three modest qualities — qualities a small business can produce without hiring anyone new.
The first is what I would call honest contrast. The page acknowledges that the question being answered has more than one reasonable answer. The cabinet maker's page, rewritten, would say something like: matte finishes look gorgeous in photographs but show fingerprints in a working kitchen; satin is the boring middle answer that almost always wins for families with children; high-gloss is wonderful in a showroom and a nightmare under direct sunlight from a south-facing window. That is reasoning. The model can extract that, compare it to other sources, and use it.
The second is grounded specificity. Real numbers, real timeframes, real names of materials and brands where appropriate. Not because the model fact-checks every claim — it does not, not really — but because specific claims carry the texture of someone who actually knows the field. The reasoning model is trained, in part, to prefer that texture.
The third, and this one surprises people, is conclusion-first paragraphs that then defend the conclusion. Not the inverted pyramid of journalism exactly, but close. State the answer. Then walk through why. The model can pull the conclusion sentence cleanly as a citation without having to assemble it from scattered fragments.
The thinking-budget signal
There is one more wrinkle that almost no one outside Google's own publications has been talking about, and it deserves a paragraph of its own.
Gemini 3 Deep Think uses more thinking budget for harder questions. For a simple navigational query — what time does the post office close in Asti — it barely thinks at all. For a comparative or advisory query — should I refinish my hardwood floors or replace them — it now routinely spends three, four, even five times as long reasoning before producing the answer.
The pages that get cited inside those longer-reasoning answers are, naturally, the pages that contain reasoning. The pages cited inside the short-reasoning answers are still mostly chosen for their pure keyword relevance.
This means a small business that wants to be cited inside the answers that actually matter — the advisory, comparative, problem-solving questions where a customer is genuinely deciding what to do — has to publish pages that survive being reasoned about. The factual little corner of the site, the hours-and-address page, will keep working as it always has. But the meaty content, the should I, which is better, what happens if — that is now a different game.
What to actually change this month
Let me be concrete, because abstract advice is, of course, the easiest thing in the world to write and the hardest thing to act on.
Take your three or four most important advisory pages — the ones that target should I, what kind of, how do I choose, or X vs Y queries. Read them out loud. If they read as a checklist of features, they are at risk. If they read as a calm voice considering the question from two or three angles and then making a recommendation, they are in good shape.
Where they fall short, the rewrite is usually small. Add one paragraph near the top that names the most common alternative answer and explains, briefly, why it is not your recommendation. Add one paragraph near the middle that includes a specific number — a price range, a duration, a measurement — that grounds the discussion in something verifiable. Add one paragraph near the end that states the conclusion in a single sentence the model can lift cleanly.
That is, more or less, the entire intervention. It does not require new pages, new tools, new agencies. It requires re-reading what you already have through the eyes of something that thinks before it speaks.
A word about the dashboards
I want to address, briefly, the dashboard question. Almost every SEO tool you have ever used reports rankings. None of them, as of this writing, reports citation share inside reasoning models with high thinking-budget allocation. The metric does not exist yet in commercial form. The visibility you are losing, or gaining, is invisible to the tools.
This is, I admit, frustrating. It means the only reliable way to know whether your pages are being cited in the deep-think answers is to actually run the queries yourself, manually, in Google AI Mode, and to read the citation list. Set aside an hour every two weeks. Pick the ten queries that matter most to your business. Note which pages are cited. Note when yours appears or disappears. It is artisanal work. There is no shortcut.
The truth is, this is how SEO used to be done, before the dashboards. People sat down and looked at the results. The dashboards will catch up — they always do — but the businesses that catch up first will be the ones doing the manual checking in the meantime.
The cabinet maker's verdict
I went back to my friend with the cabinet workshop and we spent an afternoon doing exactly what I just described. We picked seven advisory queries his customers would plausibly type. We checked which pages Google AI Mode was citing for each one. He was being cited for two of them. For the other five, the citations went to a competitor an hour away in Pesaro who, frankly, has a worse website but who happens to have written one extremely thoughtful blog post — three pages long, full of honest contrast and grounded specificity — about choosing a finish.
We are rewriting two of his pages this month. Not adding new ones. Rewriting two. We will check the citation list again at the end of June.
I will report back, in due course, on what we find. But the lesson is already clear enough to be useful to anyone reading this. The model that cites you in 2026 is no longer counting your words. It is following your argument. Make sure you have one.