Google and AI companies are eroding the internet: small sites are losing traffic, models are losing quality

Google and AI companies are increasingly reshaping how the web works—not just as data users, but as competitors to those who produce that data. The latest debate reveals an uncomfortable paradox: AI simultaneously drains value from the open internet and degrades the material it then learns from.

Traffic Under Attack

The main impact is visible in search. According to Chartbeat data published by Axios in March 2026, small publishers with 1–10k daily pageviews lost 60% of search traffic over two years. Mid-size publishers saw a 47% decline, large publishers 22%. A separate Pew Research study revealed the cause: when Google adds AI summaries, users click through to links noticeably less often. Without summaries, clicks on regular results happen almost twice as frequently, while links within the summary itself receive just a fraction of clicks.

• Small sites lost 60% of search traffic over two years
• Mid-size publishers lost 47%
• Large platforms lost 22%
• With AI summaries, users click on regular links notably less often

The problem extends beyond media. Recipes, instructions, blogs, educational sites, and any niche resources that relied on search traffic are all affected. Large brands partially offset losses through direct visits, apps, and email channels. Small projects lack such a cushion. When traffic collapses, advertising, subscriptions, and motivation to publish new material suffer—especially content requiring expertise, time, and manual work.

The Internet Loses Voices

Then the economic effect kicks in. Less traffic means less money for those creating original content. Less money means fewer reports, reviews, instructions, studies, and local stories. Small independent sites, which often provide the most unconventional topics and angles, are disappearing fastest. The internet isn't becoming empty—it's becoming flatter: instead of many different voices, a stream of interchangeable texts emerges, optimized for algorithms and quick answers.

AI undermines the very hand that feeds it.

Against this backdrop, companies and newsrooms increasingly fill gaps with machine-generated content. One estimate suggests that by mid-2025, more than half of new online content was already being created by AI. It's convenient and cheap, but this growth changes the web's composition: human texts written from experience, observation, or specialized expertise are shrinking in relative terms. As a result, the internet gradually becomes a monoculture where rare and genuine voices are drowned out by a sea of similar publications.

When Models Learn

For AI systems themselves, this is also bad news. Generative models train on vast arrays of web data, meaning they increasingly encounter texts, images, and summaries created not by humans but by previous models. A 2024 Nature study describes this risk as model collapse: under recursive training on synthetic data, a model gradually loses rare but important real-world patterns and begins to reflect reality less accurately. At first this looks like a slight quality drop, then homogeneity, errors, and outright nonsense.

The situation is complicated by unreliable filters. AI content detectors show wide accuracy variance and frequently mislabel human texts as machine-generated. These false positives disproportionately affect non-native English speakers and people with unconventional writing styles. Synthetic datasets, proposed as a solution, seem more like a temporary workaround: they can amplify models' weak points and don't replace fresh human data. In other words, the industry contaminates the source from which it then drinks.

What It Means

If this trajectory continues, the market faces a double effect: publishers lose their economics, and models lose data quality. For users, this means a poorer, more homogeneous internet. For AI developers, it means growing dependence on a limited pool of genuinely human-created content. Solutions theoretically exist—compensating creators, transparent labeling of machine-generated material, and stricter data-collection rules—but for now, the industry moves in the opposite direction.