From the Filter Bubble to the Answer Bubble: How AI Is Redrawing the Edges of What We See

The internet arrived with a promise: knowledge, within everyone's reach. Over time, a harder truth surfaced. Access to knowledge is one thing; access to the truth is another. Social platforms and search engines learned what we liked, and began showing each of us what matched our interests and past behaviour. And so the "filter bubble" was born — a digital environment that feeds our own ideas back to us and keeps what contradicts them at a distance.

But the story was never that simple. In a landmark 2015 study in Science, researchers examined the behaviour of 10.1 million Facebook users. Yes, algorithmic ranking reduced the cross-cutting news people saw by roughly 15%. The surprise was that people's own choices — what they actually clicked — narrowed their exposure even more than the algorithm did. In other words, the bubble was never the machine's work alone. We were partners in building it.

That was the anxiety for a full decade. Then AI changed the shape of the problem entirely.

Search once handed you a list. Ask about a topic and you'd get a government page, a news article, an academic view, a forum thread, several competing interpretations. The ranking was never neutral — but the alternatives stayed visible, and you were the one choosing between them.

Ask an AI system today — ChatGPT, Google, Perplexity, Claude, or Grok — and something fundamentally different happens. The system doesn't just rank your sources; it reads them, decides which ones deserve to appear, reconciles their contradictions, and hands you a single polished answer. You are no longer choosing among sources. You receive the result of choices already made on your behalf.

That shift deserves its own name. This is no longer a filter bubble hiding some of the doors from you; it is an answer bubble that walks through the doors for you and returns with one conclusion.

Is AI the cure, or a thicker wall?

At first glance, AI looks more like the remedy than the disease. It doesn't depend on your friends' posts, or on a feed engineered to keep you scrolling. In theory, it can read thousands of pages you'd never open, and draft a balanced answer in seconds. Many recent models are, in fact, designed to present more than one viewpoint and to flag where experts disagree.

But reality is more complicated. A system reading diverse sources is not the same as a system showing you diverse perspectives. When thirty pages become four paragraphs, something decided which disagreements survive the compression and which opinions fall away. The decision is no longer yours; it happens inside the box.

And this is where the real problem begins — with two faces. One concerns our behaviour; the other concerns our trust in the answer.

The first face: we stopped clicking

The clearest evidence for this shift isn't an opinion — it's a measurement of real behaviour. In a 2025 study by the Pew Research Center, researchers tracked the browsing of more than 900 people. When an AI summary appeared at the top of the page, users clicked a traditional link just 8% of the time, versus 15% when no summary appeared. As for the links inside the summary itself, they were clicked in only 1% of visits.

That last number is the most consequential. It means the answer has become the final destination, not the starting point. The user reads the summary, settles for it, and leaves — no longer seeing the competing headlines, the contradictory evidence, or the minority view that once sat on a page of search results.

In fairness, Google disputes this reading. The company has said that the total volume of clicks it sends to websites has stayed "relatively stable" year over year, and that click quality has risen. But it has not published data to prove this, whereas Pew measures actual behaviour rather than a claim. A fair reader places both figures side by side, rather than substituting one for the other.

The second face: is a polished answer a correct one?

The trouble is that this confident answer isn't always sufficiently supported. In a peer-reviewed 2023 study, researchers audited four generative search engines and found that only 51.5% of the generated sentences were fully backed by their citations, and only 74.5% of the citations actually supported the sentence attached to them.

Put differently: a source sitting beneath a sentence doesn't mean the source says what the sentence says. This is a new kind of risk. A wrong link in old-style search was one option among dozens; a wrong sentence inside a coherent summary borrows the authority of the entire answer, and reads like the final word.

How the five systems differ

It's a mistake to treat these tools as one thing. How each system searches, and how it surfaces its sources, differs in ways that directly affect whether an "answer bubble" forms. The table below summarizes each system's documented behaviour:

System	How it searches and summarizes	How it surfaces sources	Bubble-relevant note
ChatGPT Search	May rewrite your question into several targeted searches before summarizing	In-line citations plus a side sources panel	You may never see the searches the answer was built from
Google (AI Overviews / AI Mode)	Summarizes inside the results page; AI Mode splits a question into many sub-queries	Prominent links and sources accompany the answer	May search more broadly than you would, but the summary can satisfy you before you open a source
Perplexity	Searches the live web and composes a direct answer with multi-step search	Citations and links are a core interface feature	Visible citations aid verification, but reduce the need to browse
Claude	Web search across multiple sources; a research mode runs iterative searches	Attaches citations and links, and encourages checking them	Widens the angles, yet the final prose is still one mediated synthesis
Grok	Can search public X posts and the web in real time	May expose the sources behind an answer	Adds social freshness, but imports platform dynamics into the answer

So the important question isn't "does the system cite its sources?" but: how many independent sources did it read? How many viewpoints did it show? Which did it exclude? And does it make verification easier — or make it feel unnecessary?

When an information problem becomes an economic one

Until now, we've been talking about individuals. But the deeper effect touches whoever produces the information in the first place.

The open web runs on a simple equation: someone pays to produce content, and in return receives visitors, subscribers, and advertising. If the answer reaches the user without a visit to the site, that equation breaks.

Cloudflare's 2025 figures show the scale of the imbalance. Some AI systems pull thousands of pages from websites for every single visitor they send back — a ratio nothing like traditional search, which returns a visitor for every handful of pages crawled. The system takes a great deal, and returns very little.

Here a troubling loop appears: AI reduces the need to visit the source, so publishers' revenue weakens, so small and independent sources disappear, so the system comes to rely on fewer large sources. With each turn, the circle of voices from which tomorrow's answers are built grows narrower. This is what some researchers call the risk of a "monoculture" — consuming one machine's integrated account instead of choosing among many independent voices.

But none of this is inevitable

Before surrendering to the dark version, something is worth remembering.

The "monoculture" risk deserves watching — it is not a proven fact. Different systems still retrieve different sources, and give varying answers to the very same question.

The habit of tying algorithms to radicalization also needs scrutiny. In a 2025 PNAS study, researchers ran four experiments on nearly nine thousand participants, feeding them deliberately slanted recommendations on a YouTube-like platform. The result? It changed what they watched, but produced no clear, consistent short-term shift in their political attitudes. The researchers themselves cautioned that they can't rule out long-term effects, or effects on particular vulnerable groups. But the message is clear: exposure does not automatically become conviction.

The tool, then, is as capable of widening our horizon as of narrowing it. The difference, in the end, is not in the tool, but in how we use it.

What this means for businesses and brands

Here the story turns very practical.

Digital competition once revolved around one question: how do I appear on the first page of search results? Today a more important question has been added: how do I become the source the AI quotes when it builds its answer?

Appearing in the list is no longer enough, because many users won't reach the list at all. What reaches them is the summary. And whoever wants to be present in that summary must invest in something the machine cannot fabricate: original, well-documented, accurate content — built on real experience and clear data — and a digital reputation that makes the brand a reference worth trusting.

This is precisely the essence of Generative Engine Optimization (GEO). A company that treats content as a cost to be trimmed will gradually vanish from the picture the machine paints. A company that builds genuine authority in its field becomes the source the systems return to when asked. The value has moved from "being found" to "being trusted."

In the end: whose bubble is it?

Perhaps the important question isn't "will AI create new information bubbles?" but "will we use it to discover what we don't know, or to confirm what we already believe?"

Algorithms, however intelligent, cannot substitute for human curiosity. If we stop searching, stop hearing the other view, stop questioning ready-made answers — then the bubble won't be technology's doing alone. It will be a choice we made.

The old filter bubble asked: are algorithms hiding part of the internet from us? The AI-era question is more unsettling still: what happens when we stop seeing the internet at all — and see only what a machine tells us it found?

Sources

• Pew Research Center (2025). "Google users are less likely to click on links when an AI summary appears in the results." https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
• Liu, Zhang & Liang (2023). "Evaluating Verifiability in Generative Search Engines." ACL Findings (EMNLP 2023). https://aclanthology.org/2023.findings-emnlp.467/
• Cloudflare (2025). "The crawl before the fall of referrals." https://blog.cloudflare.com/ai-search-crawl-refer-ratio-on-radar/
• Liu et al., PNAS (2025). "Short-term exposure to filter-bubble recommendation systems has limited polarization effects." https://www.pnas.org/doi/10.1073/pnas.2318127122
• Google / Liz Reid (2025). "AI in Search is driving more queries and higher quality clicks." https://blog.google/products-and-platforms/products/search/ai-search-driving-more-queries-higher-quality-clicks/
• Bakshy, Messing & Adamic, Science (2015). "Exposure to Ideologically Diverse News and Opinion on Facebook." https://www.science.org/doi/10.1126/science.aaa1160