Why AI Search Is Making Google Worse

For roughly two decades, the deal with Google was simple and mostly honest: you typed a question, it handed you ten blue links, and you picked one. The results degraded over the years — more ads, more SEO sludge — but the core mechanic held. You were still choosing which source to trust.

AI Overviews changed the deal. Now, before the links, Google generates a paragraph that answers the question for you. Sometimes this is helpful. Sometimes it is confidently, memorably wrong, and it is wrong above the ten blue links that would have told you the truth.

The greatest hits of getting it wrong

The launch of AI Overviews produced a genre of screenshot that will outlive us all: the feature suggesting glue to keep cheese on pizza, recommending people eat a small rock a day, and confidently misattributing quotes and dates. Google described these as rare edge cases and adjusted the system. The edge cases were funny. The structural problem they revealed was not.

The problem is this: a summary presented as an answer carries an authority that a list of links never did. Ten blue links are visibly a menu of opinions. A single grey box at the top of the page, in Google's own voice, reads as a verdict. When the verdict is wrong, the interface has laundered a low-confidence guess into something that looks like fact.

The interface removes your ability to judge

Part of what made the old results page trustworthy was that it displayed its own uncertainty. You could see the domain names, weigh the source, notice that the top result was a forum thread and the third was a government site, and calibrate accordingly. That calibration was the user's job and the interface handed you everything you needed to do it. The summary box takes that job away. It presents a single, sourceless-feeling paragraph and asks you to trust the aggregate, with the individual signals — who said this, how reliable are they, is this one crank or a consensus — dissolved into a smooth grey voice.

The links are still down there, of course, usually collapsed behind a small expander, which is its own quiet statement of priorities. The information you would need to evaluate the summary has been demoted beneath the summary itself. You are invited to accept the conclusion and discouraged, by layout alone, from inspecting the evidence. For a company whose original genius was ranking sources so you could choose between them, burying the sources under a machine's paraphrase is a strange inversion of the thing that made it worth using.

The click that no longer happens

The second problem is quieter and, for the wider web, more serious. When the answer is in the box, you do not click through. That is the entire point of the box. But the box is built from the pages it is discouraging you from visiting. Google is summarising the open web on a page designed to keep you from reaching the open web.

Publishers have watched referral traffic from search soften as more queries resolve on the results page itself. This is not a conspiracy; it is the predictable outcome of answering the question before the link. The sites that produced the information get the cost of being crawled and none of the visit. We wrote about where that road leads in our piece on AI Overviews and the open web, and it is not a road with many sign-posts back.

You now have to fact-check your search engine

The cruellest part of the redesign is the new chore it hands the user. The reason you used a search engine was to not have to already know the answer. But because the summary is sometimes wrong in ways that are impossible to detect without checking the sources, the responsible move is now to read the summary, distrust it, and then click the links anyway to confirm — which is exactly the work the summary claimed to save you.

So the feature adds a step. Best case, it is right and you saved a click. Worst case, it is wrong, you did not notice, and you have confidently repeated an invented statistic in a meeting. Average case, you read it, squinted, and scrolled down to the links you were always going to use.

This is not anti-AI. It is anti-this-implementation.

None of this means summarisation is useless. For genuinely settled questions — conversions, definitions, opening hours — a generated answer is fine, and a good use of the technology. The complaint is about defaulting to a confident summary for every query, including the contested, the ambiguous, and the ones where being subtly wrong causes real harm.

The incentive underneath is not neutrality

It is worth asking why the box exists at all, because the answer is not “to help you.” A search engine that sends you away to a publisher's site has, for a moment, lost you. A search engine that answers on its own page keeps you — keeps the session, keeps the next query, keeps the advertising real estate in view. The answer box is, among other things, a retention mechanism dressed as a convenience. Every question resolved without a click is a question that did not hand attention to anyone else.

Once you see that, the design choices stop looking like accidents. The summary sits at the top; the links are collapsed beneath it; the whole layout gently discourages the departure that used to be the entire point of the product. This is not a conspiracy — it is the ordinary logic of an advertising business under pressure to keep users on its own surface. But it means the interests of the search engine and the interests of the searcher have quietly diverged, and when they diverge, the layout serves the one that pays the bills.

What a good version would actually do

None of this means summarised answers are inherently bad. A search engine genuinely interested in the user would show a confident answer only where the question is settled and low-stakes, mark clearly when it is uncertain, put its sources where they can be seen and clicked rather than folded away, and get out of the way entirely on the contested, the ambiguous, and the consequential. It would treat the summary as one tool among several, deployed with judgement, rather than a default bolted onto every query to keep you from leaving.

That version is entirely buildable. The reason it does not exist is not technical difficulty; it is that a modest, honest answer box that cheerfully sends you elsewhere is worse for the business than a confident one that keeps you here. So we get the confident one, wrong often enough to matter, sitting on top of the links that would have told us the truth. The technology could serve the reader. It has been pointed, instead, at the session.

Trust is the asset being spent

The thing that made Google indispensable was never the technology alone; it was trust, accumulated over two decades of that first result usually being good enough. That trust is an asset, and it is the asset now being drawn down to underwrite the answer box. Every confident summary borrows against it, and every wrong one — the rock, the glue, the invented statistic repeated in a meeting — spends a little that does not come back. Trust is slow to build and quick to burn, and a feature that is right most of the time but wrong in ways you cannot predict is precisely the kind of thing that burns it, because the failures are the memorable part.

The gamble underneath is that people will keep trusting the box faster than the box loses their trust. That is a strange thing to bet a search monopoly on. The moment enough users learn, through a few burned afternoons, that the summary must be double-checked against the links anyway, the box stops saving anyone anything and becomes pure overhead — a thing you scroll past to reach the results you were always going to use. At which point the company will have spent a genuinely rare asset, degraded its own core product, and unsettled the web it depends on, all to win an argument about being first to the future. The technology was never the problem. The willingness to spend hard-won trust to prove a point about it is.

And the exit, once that trust is spent, tends to be quiet rather than dramatic. People do not announce that they have stopped believing the answer box; they simply start scrolling past it, adding “reddit” to their queries, or drifting to whichever tool has not yet disappointed them. Habits erode at the edges, one burned afternoon at a time, and by the time the erosion is visible in the numbers it is already well advanced. A product can look dominant right up until the moment its users have privately decided to route around it — which is precisely the position a search engine puts itself in when it repeatedly answers confidently and wrongly on the questions that mattered.

Google spent twenty years teaching the world to trust that first result. It is now spending that trust to prove a point about AI, and the exchange rate is worse than the earnings calls suggest. The search engine still works. It just asks more of you now, gives less back, and occasionally recommends the rock.