--- title: "The Monopoly at the Crossroads Was Unravelling. And as It Unravelled, the World Had Begun to Love It" date: 2026-10-17 description: "Gemini 4 Argon became generally available on October 13, 2026. Google introduced it as an option in AI Mode for Google AI Pro and Ultra subscribers." author: "Petru Cojocaru" categories: - name: "Google" url: "https://cojocarupetru.info/google-en.md" --- # The Monopoly at the Crossroads Was Unravelling. And as It Unravelled, the World Had Begun to Love It *Gemini 4 Argon became generally available on 13 October 2026. Google added it to the AI Mode options for Google AI Pro and Ultra subscribers.* **How I came to speak well of Google** *About a sceptic of monopolies, a well-argued answer from AI Mode, and the strange kindness that appears when a colossus begins to falter* --- **The meal after the hram** It had reached the hour at which a Sunday meal between two related families moves past the polite stage and into the sincere one. The plates of sarmale — cabbage rolls, the dish without which no Romanian gathering counts as a gathering — had been cleared, the cozonac, the sweet braided loaf that ends every feast, remained half untouched, and the Turkish coffee was making its rounds through the house, brought in on a tray whose metal rim had been worn dull by years. I was there because the hram — the patron saint's feast day, when every Romanian village and neighbourhood gathers at one table — had brought everyone together: the in-laws come from the countryside, the cousins from town, an aunt whom nobody could say any longer exactly in what degree she was related to us, but who had, like all aunts in the world, the right to say anything about anyone. The conversation arrived, as it always does, at the people of the village. Someone mentioned Mr Vasile, the man who had kept the general store at the crossroads, the only one for several kilometres, for nearly thirty years. And then the thing happened that stayed with me: the whole table began to say only good things about him. — A man of real humanity, said my uncle, setting down his cup. He kept us on the ledger a whole winter, when we didn't have a leu to our name. — And he'd bring the flour right up to the gate, added my sister-in-law. He never asked for anything extra. — And at Ionuț's wedding he sold the wine at cost, concluded the aunt, in the tone of someone closing a file. I listened, I nodded, and for a few seconds I felt very comfortable in the role of listener to a beautiful story. Then I felt something unpleasant, like a stone in a shoe. Because I remembered something else. I remembered that ledger. I remembered that the debt on the ledger never quite died, that the prices were a few bani higher than in the town nobody went to, because the road cost more than the saving. I remembered that, when a man from the village had tried to open a rival shop in a garage, Mr Vasile had been the first to find out who owed him money and who did not. I am not saying he did anything illegal. I am saying he knew exactly what power he had and used it sparingly, the way those who never need to raise their voice use it. But why was everyone speaking well of him now? The answer was on the table, in words that had come a little earlier, from another conversation: in the neighbouring commune a supermarket had opened, one of those with white lights, trolleys lined up at the entrance and the shelf with this week's offers. In a few months the crossroads store had lost half its customers. The younger ones, with cars, did their shopping for a month at a time, once a week, on the way to town. Only the old and those without a car still stopped by Mr Vasile. And he, who in thirty years had never once given change with a smile, now greeted people by name and slipped a packet of biscuits into their string bag "for the grandchildren". The monopoly at the crossroads was unravelling. And as it unravelled, the world had begun to love it. I sat with that observation for a moment and realised something that made me smile crookedly. I, who had sworn inwardly never to be taken in by the price on the ledger, was that morning the man who, only an hour before, had asked a search engine for an answer to a hard question. I had received a well-made, well-argued answer, with sources laid within reach, and I had said out loud, in a kind of embarrassed surprise: "Well look at that, this one works." "This one" was Google. And I was, for all my banners, one of those who had written for years about the dependence on a single desk through which access to information passes, about how a company can become the infrastructure through which an entire trade finds its customers, about how an algorithm can decide, without telling you, whether or not you have come into the digital world. My work, at the agency in Botoșani, had meant year after year explaining to clients — public institutions and private firms — that their online visibility is rented, not bought. That it can be withdrawn. That however well you write, you are at the mercy of a system you cannot see and cannot negotiate with. And yet there I was at the hram table, smiling gratefully at a machine. This article is about how that is explained. Not because I need to justify myself to my aunt, but because the question deserves serious treatment: what does it mean when a sceptic begins to speak well of the one he suspected? Does it mean the suspicion was unjust? Does it mean the product has genuinely improved? Or does it mean something subtler and more uncomfortable — namely that kindness is sometimes the symptom of the crack, not of virtue? The analogy with Mr Vasile is not chosen at random. It has a structure I shall follow through the whole text: a dominant actor, a dependence, an alternative that appears, a sudden change of tone. But the analogy also has limits, which I shall mark honestly. Google is not a crossroads store, and the crossroads it now stands at is not a village's, but that of an entire economy of information. I shall proceed, then, in three movements. First, I shall say what has concretely happened in recent weeks, because I started from a personal experience, and a personal experience is not proof. Then I shall ask how true the idea is that the monopoly "is unravelling", because it is a formulation that sounds good and which, looked at closely, proves less simple. Finally, I shall try to formulate what it means to speak well of someone without forgetting the ledger. --- ## The mechanism of kindness: why we speak well of the one in difficulty Before any data, it is worth understanding how what happened at the table was possible. It was not hypocrisy, but rather a mixture of psychological and economic mechanisms, each of them recognisable also in the way we now look at Google. **The first mechanism is comparison.** Mr Vasile did not become better; the alternative became visible. Until the supermarket existed, the price at the crossroads was the price, full stop. There was no "cheaper" against which to measure him. Once the supermarket appeared, every purchase became a test, and Mr Vasile began to do well in the tests the alternative cannot offer: credit, delivery to the gate, the human kindness of the packet of biscuits. In other words, praise does not measure absolute quality, but the advantages that survived comparison. When we say of a monopoly that it "is good at something", we are in fact saying what its competitor lacks. **The second mechanism is the fear of loss.** People are, as a rule, more attached to what they risk losing than to what they have. The village suddenly realises that if the crossroads store closes, on winter days, when the road to town is impassable, there is nowhere left to buy salt and oil. The praise at the table was, in large part, retrospective insurance: a recognition of a service we had not valued while we took it for granted. **The third mechanism is the renewed behaviour of the one being praised.** Here is the most interesting part. Mr Vasile really did begin to behave differently. A threatened monopolist becomes, almost inevitably, more attentive. The biscuits for the grandchildren are not a gesture of character, but a retention gesture. Yet — and here lies the subtlety — the effect on the customer is identical: he receives more attention than he received before. Economists have long known the phenomenon: competition does its work not only through lost customers, but also through the changed behaviour of the one who risks losing them. Prices fall and quality rises before any customer has in fact left. **The fourth mechanism is anticipated nostalgia.** Those at the table who spoke of the crossroads store were already praising it in the past tense, as something that was going to disappear. This is a form of premature mourning that warms any memory: you forgive easily the things you will no longer have to endure. If I put these four mechanisms together, I reach a conclusion uncomfortable for me as a sceptic: **the fact that I speak well of Google is weak evidence of Google's quality.** It is strong evidence of something else: of the fact that Google finds itself in a situation where it must be good. And between "is good" and "must be good" there is a difference you cannot erase with a nice answer at ten in the morning. That said, it would be equally unjust to reject the immediate experience. A good answer is a good answer. Its reasons may be competitive, but its usefulness is real. The distinction I want to make throughout this text is between **recognising what works** and **forgetting why it works**. The first gesture is honest. The second is, slowly, a form of capitulation. Let us see, then, what works. ## What actually happened in October I promised I would not rely on a personal experience, so here are the facts, insofar as they could be verified from public sources. ### A new model, but not for everyone Gemini 4 Argon became generally available on 13 October 2026. Google added it to the options in AI Mode for Google AI Pro and Ultra subscribers, and developers can use it through the Gemini API. Free users did not get the change: they stay on the standard AI-mode model. It is chosen from the "+" menu next to the question bar, meaning it is not (yet) the default behaviour of Search. What Google promises is, in essence, more direction and less hesitation. The company says Argon answers more directly, repeats itself less, avoids clichés and follows style instructions more faithfully than earlier Gemini models. On the capacity side, an independent analysis cited by the press puts it on a par with OpenAI's competing model on Artificial Analysis' composite intelligence index, at a lower cost per task during the introductory pricing period. One thing is important to remember, easily lost in enthusiasm: **the answer that impressed me that morning at the table may have come from a level of product that not everyone has access to.** If I received it from Argon, my story is not the story of Google Search, but the story of Google Search for those who pay. And in our analogy, that resembles the packet of biscuits slipped into the string bag: not for all customers, but for the ones you want to keep. ### The links move up the page On 15 October, Google announced changes in the way AI Mode and AI Overviews show their sources. Inline links in the generated text will appear higher up, more often. A section is being tested at the start of some answers, with one or two official websites relevant to the user's task, each accompanied by a short explanation of its usefulness. The links at the end of some answers get thumbnails and a more visible design, and the carousels of recent articles are being extended to more countries and languages, highlighting users' preferred sources. The number of those using the "Preferred Sources" option has grown more than sevenfold since the start of the year. Clickable names for publications and creators, which open a profile in Search, are also being tested. This package of changes must be read attentively for what it does not say. Google has published no indicator showing that the increased visibility of links translates into more visits to the source sites. To judge an effect you need a baseline, and one is missing. The practical distinction, repeated by every serious analyst, is between three different things: **the presence of a citation, its visibility in the answer, and the actual click.** A link can be more visible without anyone pressing it, because the answer has already said everything the reader wanted to know. Here lies, in fact, the economic knot of the whole discussion. It does not interest me, as someone who writes, to be cited if no one comes to my page any more. But I cannot say either that what happens with the links does not matter at all. The fact that Google decides to put them higher, after years in which publishers complained that AI answers consume their traffic, is a gesture. I do not yet know whether it is a sincere gesture or merely a calculated one. Perhaps it is both. ### The quality document: a smaller change than it seems On 1 October, Google updated its guidance on the use of AI-generated content. Here I must be careful, because this is the part of the initial brief that circulated most widely with an air of "strict new rule", and the reality is more prosaic. The update is a documentation change. The Search Central changelog describes it as an update with information from the quality raters' guide. Google added a few sentences to the section on accuracy, quality and relevance: generative models predict words, they do not extract facts, so they can produce errors (the so-called hallucinations), and manual verification of generated content before publication is essential. The verification extends to titles, meta-descriptions, structured data and image alt texts as well. What did not happen: no new ranking signal was announced, no penalty specially designed for AI content, and the policy on scaled content abuse remained unchanged. Google's position, formulated as early as 2023, is the same: what matters is whether the page is useful and original, not how it was produced. The update only makes the missing human work harder to hide: accuracy, effort, originality, editorial responsibility. The timing is worth noting too. The guide was amended while September's anti-spam update was still rolling out. Google has not confirmed a direct link, and I do not assert one; I can only observe that the two moves concern the same territory, content produced at scale. ### What I cannot claim: "AI Mode feeds only on very well-argued texts" Here I must make an admission that takes a little courage, because the formulation I started from was beautiful and matched my experience: the idea that the synthesis in AI Mode is fed exclusively by articles recognised as extremely well argued. **I found no source to support it.** There is no official Google statement to that effect. Google's guidance for AI Overviews and AI Mode says that SEO fundamentals remain valid and that valuable, unique, hard-to-replace information matters more, but that is a recommendation, not a guarantee of selection. The empirical data do not show so refined a selection: one Moz study shows that only about 12% of citations in AI Mode coincide with the top ten organic results for the same question, versus roughly 38% for AI Overviews, meaning the cited sources come from a wide and hard-to-predict pool. Another study, by SE Ranking, analysed over 1.3 million citations in AI Mode and found that Google.com itself is the most-cited domain, with nearly 17.5% of the total, and that a large share of these citations point back to Google's own results pages. In other words, part of the "sources" of an answer are, in fact, Google itself. How then do I explain my experience? Most likely by three things that have nothing to do with one another. My question was one for which solid material exists. The model was a good one. And I, as a reader, read with more generosity an answer that comes with notes and links within reach, even though I opened none of them. The authority of form is an old trap: books with footnotes convinced people before anyone checked them. If I had to formulate the hypothesis honestly, it would be this: **AI Mode is not about the best-argued texts; it is about the texts it can find, understand and summarise.** The difference seems small, but it separates a theory about quality from a theory about accessibility. --- ## The cycle of rise and decline: what history says, what the hypothesis says One of the ideas in the initial proposal for this article is that the history of Google updates shows a repeated pattern: periods in which perceived quality rises, followed by periods in which it falls, and the fall would, most of the time, be linked to pressures on profit. The idea is attractive and has a real core. But I must separate it in two: the observable part and the interpretive part. ### The observable part Anyone who has worked in SEO or in online publishing has lived through waves. A year in which Search seems cleaner, then one in which the results seem stuffed with filler pages; an update that promises to reward useful content, then a corrective one, then a general complaint that results have got worse. The 2024 launch of AI Overviews was accompanied by famous examples of absurd answers, from advice to eat stones to glue recipes, which made the product criticised for inaccuracy precisely in the period it was being promoted. A series of corrections and refinements followed. This is indeed a stepped dynamic: leap, visible error, correction. ### The interpretive part That the engine of these oscillations would, as a rule, be falling profit is a plausible hypothesis, but unproven in the material I found. A cycle can have many causes: the growing complexity of a system that must make thousands of small decisions every day, the adaptation of those playing against it (SEO as an arms race), product changes, regulatory pressures. Profit may be one of them, but I cannot say it alone explains the cycle. If I had to write one safe sentence, I would say only that **economic motives are one of the factors worth checking, not the ready-made answer.** ### The specific October hypothesis: "the loss of government contracts to Anthropic and OpenAI" Here I must be direct, because the formulation does not fit the facts. From the verified material, almost the opposite emerges: - Anthropic is the one that was excluded from American government contracts. In February 2026, after it refused to drop two limits on the military use of its technology, the administration ordered agencies to stop using Anthropic's products, and the Pentagon designated the company a supply-chain risk. OpenAI signed a deal with the Pentagon immediately afterwards. - Google did not lose: in April 2026 it concluded an agreement with the Pentagon for artificial intelligence services, including classified uses. - The America.gov portal, launched on 29 September 2026 as the single gateway to federal services, is powered by Gemini (Google) and Grok (xAI), not by OpenAI or Anthropic. I am not saying Google has no problems. I am saying that "it lost government contracts to the two of them" is not the story the sources tell. If there is pressure on profit, it must be sought elsewhere: in the changing way people search, in the enormous costs of model infrastructure, in the way synthetic answers reduce traffic to sites and therefore the value of the advertising inventory. These are directions to check, not conclusions. For an article that aims to be correct, it is preferable to give up a spectacular sentence than to keep it on a false premise. Mr Vasile would say that a ledger is not filled in with what you would have liked to be owed. ## "The monopoly is unravelling": what it really means This is the claim in my title and, probably, the one most easily said with too much aplomb. I shall try to verify it on three planes, because "monopoly" has three different meanings in this discussion: legal, market, and gateway. ### The legal plane: established, but unresolved From a legal point of view, the monopoly is not a metaphor but a finding of a court. In August 2024, Judge Amit Mehta ruled that Google had illegally maintained its dominant position in the general search and search advertising market. In September 2025, at the remedies stage, the court did not order the sale of Chrome and did not ban payments for default-search placement, but imposed behavioural measures: a ban on exclusive distribution contracts, the sharing of some index and user-interaction data with rivals, the syndication of search results and search ads to competitors, plus technical oversight. The case is not over. The Department of Justice and the states appealed the remedies decision, considering it too lenient. Google appealed too, in January and then in May 2026, asking for the measures to be annulled and arguing that its position is owed to quality, investment and work: the expression it uses in its appeal is, in free translation, that it worked harder. The appeal and the cross-appeal are pending before the District of Columbia Circuit, and the measures apply in the meantime. In the separate case on advertising technology, Judge Leonie Brinkema rejected in September 2026 the sale of the AdX exchange sought by the government. She imposed instead behavioural remedies for six years, extendable: AdX and the publisher ad server must operate non-discriminatorily, including towards competing header-bidding tools, plus an internal antitrust compliance monitor. Google has announced it will appeal. The legal conclusion is not "the monopoly is unravelling". More precisely: **the monopoly has been established, it has not been dismantled, and the courts have so far chosen brakes, not scalpels.** If you ask me what remains of the crossroads store, the answer is that a judge said clearly that the price was too high, but moved nothing on the village map. ### The market plane: a crack, not a collapse Here the figures are more nuanced and, as ever, depend on who counts them. In the United States, one tracker reports that Google's share of search fell over twelve months from 87.28% to 84.17%, the first sustained period below 85%, with the steepest fall on tablets. Other sources show, worldwide, a share above 90%, with small variations: some figures from mid-2026 speak of the first-ever crossing below the 90% threshold, others of above 91% in August. The differences come from methodology (what is counted, on which devices, in which countries), and anyone quoting a single figure as truth should be viewed with reserve. More interesting than the search share is the shift among chatbots. ChatGPT, which at one point overwhelmingly controlled the AI-assistant market, has gradually lost its lead, while Gemini has grown. Figures differ by instrument, but the direction is the same: in one app tracker, ChatGPT's share fell from 65.3% in December 2024 to 46.4% at the end of May 2026, while Gemini reached 27.7%; in web-traffic data, ChatGPT's share of generative traffic fell from 76.4% to 52.7% in a year, while Gemini rose from 8.9% to 27.3%. Claude went, in the same data, from 1.6% to 8.9%. Other trackers, such as StatCounter, still show ChatGPT with a much larger share, which underlines how fragile these comparisons are. The most important observation is a different one, however, and I say it with caution, because it comes from an analysis, not from an established truth: **Google remains the biggest, but its share of clicks to websites is shrinking even though its share of searches remains intact.** In other words, people have not left Google; they have stayed with Google, only more and more answers now end on the Google page. The crack is not in the number of visitors, but in what becomes of them. This is a reversal of the supermarket analogy. Mr Vasile's customers left for the supermarket. Google's customers stayed, but the product changed under their eyes, and the tip no longer reaches those who made it possible — the authors and publishers whose texts feed the answers. ### The gateway plane: here the biggest change is visible The third sense of monopoly is the one I know from my trade: Google as gateway. Not who has the most users, but who decides who is seen. Here, the change is real and not exactly pleasant for those dependent on the gate. Classic search meant ten blue links, and a site could live off them. AI Mode means a synthesis, with sources as footnotes. The fact that Google is now raising the links and adding sections for official sites and creator profiles shows it knows the problem exists. But once the gate becomes an answer, it is no longer about a place in the top ten, but about being in the synthesis or not. And the synthesis has no pagination. For someone who, like me, works with public institutions, an irony appears here: the "official sites" section in the test announced on 15 October could be a real gain for town halls, ministries and state institutions, because it would give them a stable place in answers about their own procedures. But it would also mean that a private gate decides what is "official". I am not saying it would get it wrong. I am saying it is a kind of power a citizen ought to notice. ### What follows If I add the three planes together, I can no longer sustain the simple sentence of my title without a caveat. The **legal** monopoly is established and on appeal. The **market** monopoly has cracks, especially in the AI-assistant area, but Search remains overwhelmingly large. The **gateway** monopoly is transforming, not unravelling: it is moving from the list of links to the synthesis. In short, it is not unravelling; **it is reconfiguring under pressure.** And the pressure explains the kindness. --- ## Pressure as a form of politeness I return to the hram table and to Mr Vasile, because there lies a rule I see confirmed in October 2026 as well. A monopoly sure of itself has no need to be kind. It can be efficient, it can be slow, it can be indifferent, but it has no need of biscuits. Kindness appears when the alternative appears. And it appears, at first, as a symptom, before it becomes character. You can read several of Google's recent moves in this way: - **The better model made available, but in stages.** Argon is not given to everyone at once, but in a way that protects subscription revenues. It keeps the premium customers, like the biscuits in the string bag of the one who still comes. - **The links moved higher.** An answer to an old criticism, arriving after publishers began to organise, to sue, and to negotiate licences with rivals. - **The quality guidance.** A message to those who publish: content produced at scale, without verification, is no longer tolerated in silence. It is also a form of hygiene of its own food: models feed on what is published, and if what is published degrades, the answers degrade too. - **Competition on the ground of answers.** In AI Mode, Google no longer competes with other search engines, but with other ways of getting an answer. There it lacks the advantage of habit. None of this is reprehensible in itself. On the contrary: this is exactly how competition should work. A dominant actor becomes more attentive, more open, better, not because it has become virtuous, but because it has become vulnerable. If we want a world with fewer monopolies, we should be pleased, not scandalised, when they behave better under pressure. But there is a second face to the same observation, and it is the one that keeps me awake: **kindness under pressure is reversible.** Would Mr Vasile have stayed kind if the supermarket had closed after a year? Or would he have returned, discreetly, to the ledger? The question is not rhetorical; it follows from what we know about how markets behave. If the remedies are annulled on appeal, if the competitors tire, if the costs of models become too large for the small, the pressure disappears, and with it disappears the reason for kindness. That is why, if I speak well of Google, I am speaking about a particular moment: a moment in which being good is still in its interest. A good sceptic does not withdraw his scepticism; he moves it from the product to the conditions. I no longer ask "is the answer good?", but "what will keep it good when it no longer has to be?". ## The other side of the counter: who pays for the good answer Until now I have looked at everything from the customer's perspective: the one who receives the answer, praises it, and wonders whether he has the right to. But at a counter there are always two people, and the other one, in this case, is the one who wrote the texts the answer was made from. Here the village analogy had already frayed a little, so I set it aside for a paragraph. An AI Mode answer does not appear from nothing. It rests on tens or hundreds of pages written by people, editorial teams, institutions, blogs, encyclopaedias, forums. For years, these pages were paid indirectly through traffic: a reader reached them, saw an advert, subscribed, bought something, rang. If the answer stays on the Google page, the chain breaks at an exact spot. The reader is satisfied, Google is satisfied, and the author produced the value without receiving anything back — not even a click. This is not a new observation. As far back as the late 2000s, when search engines began displaying content lifted from sites directly on the results page, the consequence was falling traffic to those sites: users found the answer without leaving. The phenomenon even has a name, the "zero-click" result. What has changed now is the scale and the quality. A fragment out of context was a mediocre answer, one many still clicked through to find the rest. A well-made synthesis is a sufficient answer. The consequences are not merely theoretical. Among the most visible reactions has been the lawsuit brought by Penske Media, publisher of titles such as Rolling Stone and The Hollywood Reporter, which argues that AI Overviews illegally reproduce content from its sites and reduce the incentive to visit them. Put otherwise: part of those who feed the answer have ended up contesting it in court. At the same time, the user data do not show a mass flight from Google. According to an analysis based on Similarweb data, cited in the trade press, Google Search has roughly 3.3 billion unique monthly visitors, against approximately 494 million for ChatGPT, and 95% of ChatGPT users also use Google. In other words, it is not a choice between two neighbours, but an overlap. People are not moving from one shop to another; they go to both and compare them. Google's share of searches is relatively intact, but the share of clicks sent to websites is shrinking, and the clicks received from ChatGPT are too few to replace the ones lost. For this reason, for anyone who writes, the formula "Google has improved" has a reverse side it is honest to state. Google's good answer is possible also thanks to those who wrote the raw material. If those who write can no longer live from it, in a few years the raw material will be poorer, and the answers, which feed on it, will be poorer in their turn. It is a form of risk I see clearly, though I cannot measure it: **a system that consumes faster than the resource it depends on renews itself.** To avoid exaggeration, it must be said that this is a concern, not a conclusion: many researchers argue that the market will adapt, through licences, partnerships, subscription models or a return to direct sources. But no one serious pretends the balance has been found yet. From this follows also the significance of the small detail in the 15 October announcement: the fact that Google is testing clickable profiles for creators and publications, and extending the carousels of articles from users' preferred sources, could be read as a tacit admission of the problem. A shop that realises its suppliers have had enough starts putting their names on the shelf. It is a beginning, but it is not yet payment. I return now to our analogy, because here, paradoxically, it fits better than anywhere else. Mr Vasile could not sell without the flour brought by his suppliers. In the good years, he paid badly, paid late, negotiated hard, because he knew the suppliers had nowhere else to go. When the supermarket appeared and the suppliers began to have another customer, the tone changed in that direction too: payments more on time, better terms. Competitive pressure frees not only the customer; it frees the supplier too. And for those who write, the real question about "the unravelling of the monopoly" is simple: **do they now have another customer?** If yes, Google's kindness will translate into better conditions for them as well. If not, it remains a politeness towards the reader, not towards the author. --- ## Three objections I raise against myself A text that asks the reader to be sceptical should be sceptical of itself too. Here are the three strongest objections to everything I have written so far, phrased as someone who disagrees with me would phrase them. ### Objection 1: "Perhaps the product is simply better, and you are looking for complications" This is the simplest and probably the most serious objection. Perhaps Argon is indeed a more capable model. Perhaps Google learned from the 2024 mistakes. Perhaps the answers are better because the teams making them worked harder, not because someone frightened them. Not all good things have a hidden cause. My answer is that **both can be true simultaneously.** A product can be genuinely better and, at the same time, be better *also* because of competition. There is no need to choose between "Google has improved" and "Google was forced to improve". What matters for my argument is not the reason for the improvement but its stability. A product that is good out of virtue stays good when the pressure disappears. One that is good out of pressure guarantees nothing. I have no way of knowing, from October 2026, which category applies. I can only observe that time will decide, and that until then it is worth treating the improvement as a fact, without treating it as a promise. ### Objection 2: "Monopolies are not always bad, and you demonise them by reflex" The second objection comes from economic theory and has a part of truth in it. There are markets in which a dominant position produces benefits: network effects, lower costs through scale, investments a small player could never afford. A search engine indexing a huge portion of the web has enormous fixed costs, and competition means duplicating costs society might otherwise share. Perhaps the crossroads store was the only viable option in a village that small. I accept this argument, with a nuance. The problem is not size, but **the conditions under which the position is maintained.** A monopoly won through quality and kept through quality is a different matter from one kept through exclusive contracts, payments for default placement and the blocking of alternatives. The American court said, in essence, that in Google's case the latter occurred. Google disputes this on appeal. It is not my place to judge the file, but I note that precisely the remedies, which limit exclusivity and impose the sharing of some data, are what can reduce the need for kindness to come solely from market fright. If rival engines receive access to data and syndicated results, the pressure becomes structural, not conjunctural. That is the difference between a supermarket that came once and a paved road that will always bring others. ### Objection 3: "You are a well-informed user. What you say does not apply to ordinary people" The third objection targets my very position. I work in the field, I have tools, I know how to check a source, I know what to ask. For me, a good answer is a starting point. For a school pupil, a pensioner, a hurried civil servant, it is often also the end point. For them, a synthesis with footnotes is not an invitation to verify, but a dispensation from verifying. This is, I believe, the gravest of the three objections, and I have no answer that closes it. I can only say that it strengthens, not weakens, the scepticism I defend. The more convincing the answers, the more the motivation to verify shrinks, and the more verification shrinks, the greater the consequences of the errors that remain. A mediocre product warns its users through its very mediocrity. A very good one loses that natural warning. That is why Google's own recommendation, in the 1 October guidance, to manually verify AI-generated content must also be read as a message to users: the goodness of an answer must not take you in either. --- ## What all this means for anyone who publishes Until now I have written about a person who reads. I must also write about a person who publishes, because that is the position I work from daily: public institutions and firms that want to be found, understood and believed online. ### The first rule: do not confuse citation with relationship Being cited in an AI answer is not the same as having a reader. A more visible link is an opportunity, not a guarantee. If you measure success only by "I appear in AI Mode", you are measuring something that can grow while your business shrinks. The measurements that count are those at the end of the chain: visits, enquiries, calls, sign-ups, submitted files. For a public institution these translate into concrete things: how many citizens found the right procedure, how many submitted a complete application the first time, how many rang the front desk for information that should have been obvious on the page. ### The second rule: quality is no longer an optional virtue, but a condition of existence The October guidance introduced no new rule, but stated more clearly something that was already true: content produced at scale, without real effort, without originality and without verification, is a form of risk. For anyone writing with the help of models, the practical conclusion is clear: you can use assistance, but the responsibility remains yours. Titles, meta-descriptions, structured data, image descriptions: all must be checked. A text that knows who wrote it, why it exists and what it rests on has better chances of being useful, whether a human or a system reads it. ### The third rule: the fundamentals have not disappeared The most reassuring formulation of Google's guidance on AI Overviews and AI Mode is also the most boring: no special tricks are needed. There is no need for a separate version of each page for "machines", no need for artificial fragmentation into small pieces. What matters is that the page is accessible, indexable, clear, with information you do not find identically in ten other places. For anyone who has worked seriously until now, it is not a revolution, but a confirmation. For those who want to measure, Search Console has begun, where the report is available, to show its own data about visibility in AI features. This is a useful direction: to stop speculating from outside about what the system does, and to see at least part of what it sees. ### The fourth rule: diversify the gate This is the lesson I would have given anyway, but October 2026 reinforces it. If all your traffic depends on a single gate, any change of disposition, for better or worse, changes your life. A site with direct readers, subscribers, bulletins, communities, partnerships is less vulnerable to a new menu, an interface test or a product decision made thousands of kilometres away. I have always told clients: visibility is rented. I never said the landlord is bad. I only said he can change the contract. --- ## How I would verify it, if I had the time: a small protocol Since the article starts from a personal impression, it would be honest to propose how the impression could be turned into an observation. Here is a simple protocol that anyone can run. **Step 1: a fixed list of questions.** Choose twenty questions from a field you know well: ten with a clear, verifiable answer, ten with a disputed or nuanced one. Keep the same wording. **Step 2: the same questions, several configurations.** Ask them in AI Mode with the standard model, then with Argon (if you have access), then in other systems. Note the date, the model, the time. **Step 3: read the sources, not just the answer.** For each answer, open at least two of the cited sources and check whether they actually support the claim. Note three categories: supports, partially supports, does not support. **Step 4: track who is cited.** How many sources are Google pages? How many are official sites? How many are publications? How many are forums? Compare with the same questions a month later. **Step 5: repeat over time.** The cycle I spoke of, rise and decline, is not visible in a day, but in a year. A simple log, with date and result, is worth more than any impression. Such a protocol will not produce scientific proof, but it will produce something better than an opinion: a debate with numbers. And if the results show that good answers come from good sources, I shall gladly withdraw the reservation in section 3 about the selection of texts. --- ## The ledger I return, in the end, to the hram table. When my aunt closed the file on the wine from Ionuț's wedding, I had a moment in which I wanted to say something. To mention the ledger, to ask why he did not bring the flour to the people at the far end of the village too, to bring up the garage. I said nothing. Not out of cowardice, but because I understood something that seemed to me more important than my momentary justice: that man really had done both the good everyone spoke of and the bad I kept quiet about. Both were true. And the table was not the place for a balance sheet. But I did not forget either. And I believe this is, in fact, the attitude the whole of this text proposes. I speak well of Google because, this morning, it did something well. I do not withdraw the praise. But I do not confuse it with absolution, for it is not mine to give. My praise is confirmation of a usefulness, not long-term trust. The two things can be said in one sentence: *it works well, and precisely for that reason I am interested in why, for how long, and on what conditions.* If I had to sum up what I have learned from October 2026, it would be four things. **First:** the kindness of a dominant actor is an indication of weakness, not of virtue. It should not be despised, because it is useful, but neither should it be confused with a change of character. **Second:** "the monopoly is unravelling" is too large a phrase for what is happening. A reconfiguration under pressure is happening: legally, still on appeal; in the market, with cracks visible mainly in the AI-assistant area; at the gate, with the move from the list of links to the synthesis. Each of these requires a separate judgement. **Third:** not everything said convincingly about change is true. I started from four claims that sounded good and found that one is fully confirmed, two partially, and one — the one about losing government contracts — runs in the opposite direction. A sceptic is not someone who believes nothing, but someone who checks what he believes. **Fourth:** pressure is temporary, and the good that depends on it is temporary too. If we want it to remain, the conditions that make it possible must remain as well: real alternatives, applied remedies, open data, publishers who can live from what they publish. Otherwise, the ledger reopens — not with a scandal, but with a quiet. My aunt would probably put it more simply: "Leave it, dear, the man was good." And I would answer, if I had the courage to spoil the mood: "He was good. And I am glad he was good. But let us not forget why he had to be." The coffee had gone cold. Someone brought in another tray of cozonac. I helped myself. --- ## Bibliography and sources **On Gemini 4 Argon and AI Mode** - Search Engine Journal, "Google Adds Gemini 4 Argon To AI Mode For Pro & Ultra Subscribers": [https://www.searchenginejournal.com/google-gemini-4-argon-ai-mode/592806/](https://www.searchenginejournal.com/google-gemini-4-argon-ai-mode/592806/) - 9to5Google, "Gemini app rolling out Gemini 4 Argon to AI Pro and AI Ultra" (13 Oct. 2026): [https://9to5google.com/2026/10/13/gemini-app-4-argon/](https://9to5google.com/2026/10/13/gemini-app-4-argon/) - Engadget, "Google's Gemini 4 Argon is now available in the Gemini app and AI Mode in Search": [https://www.engadget.com/2285023/google-gemini-4-argon-is-now-available-in-the-gemini-app-and-ai-mode-in-search/](https://www.engadget.com/2285023/google-gemini-4-argon-is-now-available-in-the-gemini-app-and-ai-mode-in-search/) - Search Engine Roundtable, "Google Search's AI Mode Gains Gemini 4 Argon": [https://www.seroundtable.com/google-ai-mode-gemini-4-argon-42289.html](https://www.seroundtable.com/google-ai-mode-gemini-4-argon-42289.html) - Single Grain, "Gemini 4 Argon Splits Google AI Mode in Two": [https://www.singlegrain.com/blog-posts/gemini-4-argon-ai-mode/](https://www.singlegrain.com/blog-posts/gemini-4-argon-ai-mode/) **On citations and links (the 15 October 2026 announcement)** - ContentGrip, "Google raises source-link visibility in AI Search" (16 Oct. 2026): [https://www.contentgrip.com/google-ai-search-source-links/](https://www.contentgrip.com/google-ai-search-source-links/) - TokenPost, "Google Moves AI Search Links Higher as News Pilot Tops 400 Publications": [https://www.tokenpost.com/news/technology/32938](https://www.tokenpost.com/news/technology/32938) - Analytics Insight, "Google Search Updates AI Mode, AI Overviews with More Prominent Links": [https://www.analyticsinsight.net/news/google-search-updates-ai-mode-ai-overviews-with-more-prominent-links](https://www.analyticsinsight.net/news/google-search-updates-ai-mode-ai-overviews-with-more-prominent-links) - ALM Corp, study on AI Mode citations (SE Ranking, Moz): [https://almcorp.com/blog/google-ai-mode-ask-about-citation-overlays/](https://almcorp.com/blog/google-ai-mode-ask-about-citation-overlays/) **On the AI-generated content guidance (1 October 2026)** - Search Engine Journal, "Google Tells Sites To Fact-Check AI Content Before Publishing": [https://www.searchenginejournal.com/google-fact-check-ai-content-before-publishing/591782/](https://www.searchenginejournal.com/google-fact-check-ai-content-before-publishing/591782/) - Relevant Audience, "Google: Fact-Check All AI Content, Including Metadata": [https://www.relevantaudience.com/seo/google-ai-content-guidance-manual-fact-check/](https://www.relevantaudience.com/seo/google-ai-content-guidance-manual-fact-check/) - Digital Applied, "Google Now Says Fact-Check All AI Content Before Publishing": [https://www.digitalapplied.com/blog/google-ai-content-fact-check-guidance-2026](https://www.digitalapplied.com/blog/google-ai-content-fact-check-guidance-2026) - UU Daily, "Google Search Update: October 2026 Briefing": [https://www.uu.com.my/google-search-update-october-2026-briefing-uu-daily/](https://www.uu.com.my/google-search-update-october-2026-briefing-uu-daily/) - Locke Lee, "Google Updated Its AI-Generated Content Guidance": [https://lockelee.hk/seo/google-ai-generated-content-guidance/](https://lockelee.hk/seo/google-ai-generated-content-guidance/) **On the antitrust cases** - Alphabet Inc., Form 10-Q (Q2 2026), litigation section: [https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm) - Search Engine Land, "DOJ and states appeal Google search antitrust remedies ruling": [https://searchengineland.com/doj-states-appeal-google-search-antitrust-remedies-ruling-468230](https://searchengineland.com/doj-states-appeal-google-search-antitrust-remedies-ruling-468230) - MacRumors, "Google Appeals Antitrust Ruling, Says Apple Chose Its Search Engine 'Fair and Square'": [https://www.macrumors.com/2026/05/22/google-search-monopoly-appeal/](https://www.macrumors.com/2026/05/22/google-search-monopoly-appeal/) - Bloomberg, "Google Search Ruling to be Appealed by DOJ, States": [https://www.bloomberg.com/news/articles/2026-02-03/google-search-remedy-to-be-appealed-by-state-attorneys-general](https://www.bloomberg.com/news/articles/2026-02-03/google-search-remedy-to-be-appealed-by-state-attorneys-general) - MLex, "Google behavioral remedies in DOJ adtech suit set for six-year term": [https://www.mlex.com/mlex/articles/2526421](https://www.mlex.com/mlex/articles/2526421) - The Star (Reuters), "Google should relax ad tech rules, appoint antitrust monitor, US judge finds": [https://www.thestar.com.my/tech/tech-news/2026/09/17/google-should-relax-ad-tech-rules-appoint-antitrust-monitor-us-judge-finds](https://www.thestar.com.my/tech/tech-news/2026/09/17/google-should-relax-ad-tech-rules-appoint-antitrust-monitor-us-judge-finds) **On market shares** - WebFX, "2026 Search Market Share": [https://www.webfx.com/blog/seo/2019-search-market-share/](https://www.webfx.com/blog/seo/2019-search-market-share/) - SEO Turtle, "AI Search Market Share: what the 2026 numbers actually measure": [https://seoturtle.com/seo-insights/ai-search-engine-market-share-chatgpt-google-gemini](https://seoturtle.com/seo-insights/ai-search-engine-market-share-chatgpt-google-gemini) - Tech Insider, "ChatGPT Market Share Falls to 46.4%": [https://tech-insider.org/chatgpt-market-share-below-50-percent-2026/](https://tech-insider.org/chatgpt-market-share-below-50-percent-2026/) - Fortune, "ChatGPT's market share is slipping as Google and rivals close the gap": [https://dc.fortune.com/2026/02/05/chatgpt-openai-market-share-app-slip-google-rivals-close-the-gap](https://dc.fortune.com/2026/02/05/chatgpt-openai-market-share-app-slip-google-rivals-close-the-gap) **On government contracts** - NPR, "OpenAI announces Pentagon deal after Trump bans Anthropic": [https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban](https://www.npr.org/2026/02/27/nx-s1-5729118/trump-anthropic-pentagon-openai-ai-weapons-ban) - Fortune, "OpenAI sweeps in to snag Pentagon contract after Anthropic labeled 'supply chain risk'": [https://fortune.com/2026/02/28/openai-pentagon-deal-anthropic-designated-supply-chain-risk-unprecedented-action-damage-its-growth/](https://fortune.com/2026/02/28/openai-pentagon-deal-anthropic-designated-supply-chain-risk-unprecedented-action-damage-its-growth/) - NBC News, "Pentagon inks deal with Google for AI services": [https://www.nbcnews.com/tech/tech-news/pentagon-inks-deal-google-ai-services-rcna342661](https://www.nbcnews.com/tech/tech-news/pentagon-inks-deal-google-ai-services-rcna342661) - Congressional Research Service, "Federal Government and Anthropic": [https://www.congress.gov/crs_external_products/IF/PDF/IF13217/IF13217.4.pdf](https://www.congress.gov/crs_external_products/IF/PDF/IF13217/IF13217.4.pdf) - General Legal, "Why America.gov Chose Gemini and Grok Over OpenAI, Anthropic": [https://general.legal/blog/gemini-and-grok-power-america-gov-model-choice-impact-on-ai-vendors](https://general.legal/blog/gemini-and-grok-power-america-gov-model-choice-impact-on-ai-vendors) *Note: market-share figures differ across instruments (StatCounter, Similarweb, Sensor Tower, Apptopia) and are not directly comparable. The state of the cases refers to publicly available information as of 17 October 2026.* --- ## Ten questions from the article, with practical answers *Companion material to "How I came to speak well of Google". The situation as of 17 October 2026.* The questions start from the points where the article leaves a problem open. The answers are written for anyone who publishes online: agencies, public institutions, editorial teams, firms. Where something is not confirmed by sources, it says so. --- ### 1. How do I check whether an answer from AI Mode is genuinely well argued, or merely sounds convincing? **Why it matters.** The article starts from an impression: a well-argued answer. But the authority of form (notes, links, confident tone) convinces even before being checked. A good answer and a good-sounding answer are different things. **The four-step method:** 1. **Break the answer into claims.** Note every concrete claim (figure, date, name, rule, cause-and-effect relation). A five-line paragraph usually contains three to five verifiable claims. 2. **Open the source next to each claim.** Not just the page cited in general, but the passage. Apply three labels: *supports*, *partially supports*, *does not support*. The "partial" ones are the most dangerous, because the source talks about the topic, but not about the exact claim. 3. **Check the type of source.** Ask yourself: is it a primary source (official document, court, company, study) or a source repeating it? If an important claim rests only on a blog that itself cites nothing, it is not argued, it is repeated. 4. **Look for what is missing.** A good answer also says where it is uncertain. If a disputed question receives an answer without any reservation, treat it with suspicion. **A quick test for pupils, pensioners or hurried colleagues:***"What would have to be true for this answer to be correct, and can I check at least one of those things in two minutes?"* Even a single check greatly reduces the risk of taking an error for a fact. **Worth remembering.** If you use AI Mode for decisions with consequences (legal, financial, medical, administrative), the answer is a starting point, not a verdict. The recommendation in Google's 1 October 2026 guidance on manual verification applies to the reader too, not only the author. --- ### 2. Do I have access to Gemini 4 Argon in AI Mode? How do I enable it, and what differences should I expect to see? **Who has it.** From 13 October 2026, Argon is available in AI Mode for Google AI Pro and Ultra subscribers. Free users stay on the standard model. The cheaper subscription, Google AI Plus, does not include Argon, according to the sources consulted. It is not set by default: it is chosen manually. **How to enable it:** 1. Open AI Mode in Google Search, signed in with the account that has the subscription. 2. Tap the "+" icon in the "Ask anything" bar. 3. Choose Argon from the model menu. 4. If it does not appear, check the subscription type and whether the feature has reached your region. The rollout is staged and some accounts receive it later. **What Google says changes:** more direct answers, fewer repetitions and clichés, style instructions followed better, stronger performance on multi-step tasks and follow-up-question conversations. **How to compare fairly, without fooling yourself:** - Ask the same 10 questions once with the standard model, once with Argon, on the same day. - Note the model used next to each answer (otherwise in a month you will not know what you tested). - Compare not only the text but the cited sources too: are they the same? More numerous? Of a different nature? - Keep the screenshots. The interface changes often, and without evidence you cannot prove what you were seeing. **Watch the costs, if you work through the API.** The announced introductory price is 2 dollars per million input tokens and 10 per million output tokens, valid until 30 November 2026; after that, double. If you are planning automations, calculate at the doubled price. **Worth remembering.** If a client tells you "Google answers much better now", ask which model they use and whether they have a subscription. Their experience may not be the general public's. --- ### 3. How do I measure whether appearing in AI Mode actually brings clients or readers? **The problem.** The article distinguishes three things: the *presence* of a citation, its *visibility* in the answer, and the actual *click*. Google has published no data showing that more visible links bring more traffic. So you cannot assume; you must measure. **What you measure, in order, from the start of the chain to the end:** | Level | What you ask | Where you look | | --- | --- | --- | | Presence | Am I cited in relevant answers? | Periodic manual checks on a fixed list of questions; the AI reports in Search Console, where available for your property | | Visibility | Where does the link appear (top, in the text, at the end)? | Dated screenshots, recorded in a log | | Traffic | Do visits come from these appearances? | The traffic-sources report in your analytics tool; compare the periods before and after | | Result | Do they do anything useful? | Conversions: enquiries, calls, sign-ups, downloads, submitted files | **Practical rules:** - **Establish a baseline now.** Without before-figures you can calculate nothing afterwards. Record traffic and conversions for the last 8–12 weeks. - **Do not attribute automatically.** If traffic rises after 15 October, it may be something else: a viral article, the season, a campaign. Compare with pages that were not cited. - **Measure the result, not only the appearance.** A visible link that produces no actions is not a gain. - **For public institutions,** choose a service indicator, not an audience one: how many applications were submitted complete the first time, how many calls to the front desk fell for questions that should be answered on the page. **Worth remembering.** Google says adoption of "Preferred Sources" has grown more than sevenfold since the start of the year, but has not linked that to more traffic for publishers. Do not buy a success story before you have your own figures. --- ### 4. Can I use AI in my writing without risking trouble? What does a responsible workflow look like? **What actually changed.** On 1 October 2026, Google updated its guidance on AI-generated content. It introduced no new penalty and did not change the policy on scaled content abuse. It said more clearly that models predict words, not extract facts, so they can be wrong, and that manual verification of AI content before publication is critical, including titles, meta-descriptions, structured data and alt texts. **So the short answer is:** yes, you can use AI, but the responsibility remains yours and that of the people who publish. **A six-step workflow:** 1. **A human brief.** A person sets the page's purpose, the audience, what only that organisation knows (experience, own data, real procedures). Without this, the text will be interchangeable with ten others. 2. **Assisted draft.** The model can propose structure, phrasings, variants. Do not ask it for facts you have not given it; ask it only to organise what you have. 3. **Fact-check by claim.** All figures, dates, names, quotes and references to legislation are checked in the primary source. For legal or administrative texts: against the normative act in force, not against a news report. 4. **Check the "invisible" elements.** The page title, meta-description, structured data (schema), image alt texts. This is where errors slip in most easily, because nobody reads them again after generation. 5. **Signature and responsibility.** Who wrote it, who checked it, when it was last updated. For institutional content, a date of the last check on the page helps both readers and systems. 6. **Periodic review.** Pages describing procedures, deadlines or tariffs are re-checked at a fixed date. **Warning signs in your own content:** - Tens of nearly identical pages, with a single variable changed (town, product). - Pages repeating what all the others say, with nothing of your own. - Claims nobody on the team could support with a source if a citizen asked. **On disclosing the use of AI.** There is no general obligation, according to the sources consulted. It may be useful where the reader would reasonably wonder how the content was produced. Some Google products (Merchant Center, for example) have special labelling rules for certain generated images; check the policies of the product you use. --- ### 5. What can I do so that a public institution's site appears in the "official sites" section of AI Mode? **What we know.** On 15 October 2026, Google announced a test: at the start of some answers a section appears with one or two official websites relevant to the user's task, each with an explanation of why it would be useful. It is a test, not a generally launched feature, and Google has not published the selection criteria. **What I cannot say:** that there is a procedure guaranteeing appearance. Any agency promising that is selling you a certainty that does not exist. **What can rationally be prepared (fundamentals, not tricks):** 1. **A clear identity.** The official domain (for example .ro or .gov.ro, as the case may be), the institution's name spelled consistently everywhere, contact details and addresses identical on the site, in Search Console and in the institution's public profiles. 2. **One page per procedure.** The citizens' most-searched actions (obtaining a certificate, paying a fee, booking an appointment, submitting an application) each have their own page, with a title that states the action, numbered steps, required documents, deadlines, costs, where to submit. 3. **Updated and dated pages.** If the procedure changed, the page shows it; old versions do not remain live without warning. 4. **Technical accessibility.** Fast pages, usable on mobile, indexable, without the essential information living only in scanned PDFs. 5. **Correct structured data,** where it applies, manually checked (see question 4). 6. **Verified ownership in Search Console** and periodic monitoring of indexing problems. 7. **One canonical place per piece of information.** If the same procedure is described differently on three sites of the same institution, systems do not know which one is official. **How to test:** ask ten real citizen questions in AI Mode, once a month. Note whether the institution's site appears, in what position and with what explanation. Keep the screenshots. **A matter of principle,** from the article: a private gate that decides what is "official" exercises a kind of power a citizen ought to notice. For institutions, this means not depending on it exclusively: their own channels (bulletins, notifications, service desks, apps) remain mandatory. --- ### 6. Do I need special "for AI" pages, llms.txt files, or texts cut into small pieces? **The answer, based on the guidance discussed in the trade press:** it is not necessary. Google's guidance for AI Overviews and AI Mode, as the sources consulted describe it, says that SEO fundamentals remain valid and does not ask for separate versions of pages for automated systems, special files or artificial fragmentation. What matters is valuable, unique, hard-to-replace information. **What "fundamentals" means in practice:** - Pages that are accessible and indexable. - Titles and structure that reflect people's real questions. - Information you cannot find identically in ten other places: your own data, experience, verified procedures, examples. - An identifiable author and responsibility. - Updates when the described reality changes. **Not worth investing in, until evidence appears:** - Duplicating every page in a version "for robots". - Breaking texts into fragments meaningless to a human reader. - Costly treatments sold as "AI optimisation" without measurements. **Worth it, without regret:** making pages clearer for a person reading them on a phone. What is good for humans usually remains good for the system too. **So you are not taken in:** when someone proposes a "GEO" technique to you, ask for three things: an official source recommending it, a test on your own site with a baseline, and a criterion for abandoning it if it does not work. --- ### 7. How do I reduce dependence on a single gate, without giving up Google? **The idea from the article:** visibility is rented, not bought. It does not mean the landlord is bad, only that he can change the contract. October 2026 confirms it: several changes arrived within a single month, and none was negotiated with those affected. **A realistic plan, on three horizons:** **Immediately (within 30 days):** - Check what percentage of total traffic comes from Google. If it is over half, you are vulnerable. - Collect e-mail addresses with consent (a sign-up form for news, notifications about changes to procedures or tariffs). - Make sure you have access to your own analytics data, not only to other people's reports. **Medium term (3–6 months):** - A short, useful (not promotional) periodic bulletin, for your audience or for citizens. - A presence in 1–2 channels where your audience already is (local groups, professional networks, the institution's platform). - Link partnerships with trusted sources in your field: associations, professional chambers, local publications. **Long term (6–18 months):** - Reference content that becomes something others link to of their own accord. - Direct relationships: community, events, partnerships. - An annual review of dependencies. **One simple indicator to track monthly:***how many people reached me without going through Google?* (direct traffic, e-mail, referrals from other sites, searches for the organisation's name). If the number grows, dependence is falling. **Worth remembering.** It is not about leaving Google. It is about it not being your only road. --- ### 8. The market-share figures contradict one another. How do I read them without being misled? **Why they contradict.** The article shows that one tracker reports for the US a fall in Google's share from 87.28% to 84.17% in twelve months, while other sources give, worldwide, above 90% (some variants say it dropped below 90%, others say above 91% in August 2026). For chatbots: ChatGPT's share is 46.4% in one app tracker and 76.87% in other web-traffic data. They do not necessarily contradict each other; they measure different things. **Five questions before quoting a figure:** 1. **What is counted?** Searches, clicks, users, site visits, app installs, minutes of use? 2. **Where?** Global, US, Europe, Romania? Desktop, mobile, tablet? 3. **Who measures, and how?** A web-traffic tracker, an app analysis, a user panel, the company's own data? 4. **When?** The month, the quarter, the publication date? 5. **Who has an interest?** Agencies and platforms selling "AI optimisation" services have an interest in showing big shifts. Check whether the source sells something related to the figure. **Practical rules:** - Quote the figure together with source and date: "X% (tracker, month, year)". - When giving a comparison, use the same tracker for both terms. - Give ranges when sources differ, not a round figure that sounds certain. - Distinguish between *share of searches* and *share of clicks to websites*: the article shows the first can stay stable while the second falls. **An example of a correct formulation for a client report:** "According to one tracker, Google's share in the US fell in a year from 87.28% to 84.17%; other instruments report global figures above 90%, so the direction is a gentle decline, not a collapse." --- ### 9. How do I tell whether an improvement in Google's quality is durable, or just a moment in the "rise and decline" cycle? **What the article says.** That the pattern of oscillations is real as an observation, but the explanation through profit is a hypothesis, not a conclusion. Since you cannot know in advance, you need your own data over a long period. **An observation log, in five columns:** | Column | Example | | --- | --- | | Date and model/feature | 17 Oct. 2026, AI Mode, Argon | | The question (from the fixed list) | "How do I submit an application for..." | | Answer quality (1–5) | 4 | | Source verification | 2 of 3 support the claim | | Observations | Did not mention the deadline | **Rules:** - **The same list of 15–20 questions,** repeated monthly. If you change the questions, you are no longer comparing anything. - **Note the questions where the answer got worse too,** not only the improved ones. - **A single observer, or common rules** of scoring if there are several. - **After 3–6 months,** look at the series, not at a point. **How to interpret:** - Quality that rises then drops sharply right after a product change: possibly a cycle tied to launches. - Quality that stays stable once the novelty fades: an indication of real improvement. - Quality that varies by question type (good on facts, weak on nuance): that is about the task type, not the cycle. **Mind your own bias.** The first days after a new model are the most impressive, because we compare them with the memory of a weaker system. That is why the third month matters more than the first week. **Do not attribute cause.** Even if you observe a cycle, you cannot say why it appears. The possible causes (the system's complexity, the adaptation of those playing against it, economic or regulatory pressures) remain to be documented, not presumed. --- ### 10. What events should change my opinion about "the unravelling of the monopoly", and how do I track them? **The principle.** A good sceptic does not withdraw his scepticism; he moves it from the product to the conditions. Fix now what would have to happen for your conclusion to change, otherwise the opinion drifts with mood. **The watchlist, with what each signal would suggest:** | Event | Where to track it | What it would suggest | | --- | --- | --- | | The Court of Appeal's (D.C. Circuit) decision in the search case | Official documents, legal and trade press | Confirmation of the remedies = structural pressure; annulment = return to the previous situation | | The appeal in the ad tech case | The same | Whether the behavioural remedies (6 years) hold or are weakened | | Does Argon become the default model in AI Mode? | Google announcements, own tests | Whether the better answer reaches everyone or stays staged | | Published data on clicks to publishers | Google, independent reports, your own analyses | Whether more visible links bring real traffic | | The "official sites" test becomes a general feature | Google announcements | Whether institutions get a stable place in answers | | Market-share trends in the same tracker, over 6–12 months | StatCounter, Similarweb and others | Whether the crack deepens or closes | | Licensing deals between platforms and publishers | The trade press | Whether those who write finally get paid | | Competition in the public sector (contracts, government portals) | Official sources, the press | Whether alternatives gain ground where institutional decisions are made | **How to use it:** once a month, 30 minutes, go through the table and note whether anything changed. Record the date and the source. Without such a ritual, the conclusion remains an impression. **A proposed revision criterion.** If, over a 6–12-month interval (a) the remedies remain in force, (b) publishers receive verifiable traffic or compensation and (c) answer quality stays stable without the pressure of novelty, my conclusion moves towards "durable improvement". If one of them falls, it remains "kindness under pressure". **What stays unchanged whatever the outcome:** the ledger is not forgotten. You can say a service is good and, at the same time, know what its goodness rests on. --- ## One-page sheet: what you do on Monday morning 1. **Fix your list of 15–20 questions** from your field and test them in AI Mode with the standard model and, if you have access, with Argon. Save the screenshots. 2. **Record the baseline** of traffic and conversions for the last 8–12 weeks. 3. **Check dependence:** what percentage of traffic comes from Google. 4. **Re-check metadata** (titles, meta-descriptions, structured data, alt texts) on pages with AI-generated or AI-assisted content. 5. **Clarify the official identity** of the site (for institutions): consistent naming, one page per procedure, updated data and versions. 6. **Start a bulletin or subscriber list** with consent. 7. **Put in your calendar** a monthly 30-minute review: the quality log + the watchlist of events. --- ## Main sources for these answers - Search Engine Journal, Google and the verification of AI content: [https://www.searchenginejournal.com/google-fact-check-ai-content-before-publishing/591782/](https://www.searchenginejournal.com/google-fact-check-ai-content-before-publishing/591782/) - Relevant Audience, the same update, with details on metadata: [https://www.relevantaudience.com/seo/google-ai-content-guidance-manual-fact-check/](https://www.relevantaudience.com/seo/google-ai-content-guidance-manual-fact-check/) - ContentGrip, the links in AI Mode and AI Overviews (the 15 October 2026 announcement): [https://www.contentgrip.com/google-ai-search-source-links/](https://www.contentgrip.com/google-ai-search-source-links/) - TokenPost, the same changes, with details on the "official sites" section: [https://www.tokenpost.com/news/technology/32938](https://www.tokenpost.com/news/technology/32938) - 9to5Google, Argon availability and pricing: [https://9to5google.com/2026/10/13/gemini-app-4-argon/](https://9to5google.com/2026/10/13/gemini-app-4-argon/) - Search Engine Journal, Argon in AI Mode: [https://www.searchenginejournal.com/google-gemini-4-argon-ai-mode/592806/](https://www.searchenginejournal.com/google-gemini-4-argon-ai-mode/592806/) - Locke Lee, a synthesis of Google's guidance on AI content and AI Mode: [https://lockelee.hk/seo/google-ai-generated-content-guidance/](https://lockelee.hk/seo/google-ai-generated-content-guidance/) - ALM Corp, studies on AI Mode citations: [https://almcorp.com/blog/google-ai-mode-ask-about-citation-overlays/](https://almcorp.com/blog/google-ai-mode-ask-about-citation-overlays/) - SEO Turtle, what the market-share figures actually measure: [https://seoturtle.com/seo-insights/ai-search-engine-market-share-chatgpt-google-gemini](https://seoturtle.com/seo-insights/ai-search-engine-market-share-chatgpt-google-gemini) - WebFX, search market share 2026: [https://www.webfx.com/blog/seo/2019-search-market-share/](https://www.webfx.com/blog/seo/2019-search-market-share/) - Tech Insider, ChatGPT's share in 2026: [https://tech-insider.org/chatgpt-market-share-below-50-percent-2026/](https://tech-insider.org/chatgpt-market-share-below-50-percent-2026/) - Alphabet, Form 10-Q (Q2 2026), state of the litigation: [https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm](https://www.sec.gov/Archives/edgar/data/0001652044/000165204426000071/goog-20260630.htm) - MLex, the remedies in the ad tech case: [https://www.mlex.com/mlex/articles/2526421](https://www.mlex.com/mlex/articles/2526421) *Note: where a procedure is not documented by an official source (for example the selection criteria for official sites in AI Mode), these answers offer reasonable practice, not guarantees.* ---