Many apps I've been using lately seem to have the same issue. HBO Max buffers all the time; Instagram stops mid-scroll. At the same time, internet providers promise the fastest internet we could ever have. How is this contradiction happening? How can the internet be in it's fastest and slowest era?
It's not just a feeling, but it's also not new
Web pages have been getting heavier for a long time, well before AI features were a factor. Most pages these days are bloated with JS, media and chatbots. Cloudflare's own performance research points out that web pages have grown 6 to 9 percent every year for roughly a decade. Separately, HTTP Archive data puts the median desktop page at around 2.3 megabytes today. So the starting point here is: the web has been trending heavier for years and the real question is whether AI is making an existing trend worse, and in what specific ways.
What AI features actually add to the weight
Chat widgets add real weight, and it varies a lot by vendor. Independent benchmarking has found that a lightweight AI chat widget can add as little as 79KB of JavaScript, while a heavier one can add over 500KB and nearly a second of script execution time.
This weight typically loads whether or not the visitor ever opens the widget. If a chat bubble sitting in the corner of a page is still costing something on every single page load, for every visitor, most of whom will never click it.
Search results now often think before they answer. AI-generated answers and summaries now appear in a large share of search results, with estimates putting AI-generated answers in roughly half of Google queries in early 2026. Search used to be near-instant: type, get a results page. Now, a visible "generating" moment has been inserted into an interaction that used to feel immediate.
Agent and crawler traffic is adding load most people never see. Beyond what a human visitor experiences directly, the sheer volume of automated AI agents and crawlers hitting websites has grown sharply. Cloudflare has reported that agentic traffic made up close to 10 percent of its network requests as of March 2026, up 60 percent year over year.
Instant interactions are quietly becoming round trips. Actions that used to be handled instantly, client-side, are increasingly being routed through an AI call instead: smarter autocomplete, AI-assisted form suggestions, AI-powered search-as-you-type. Even when the AI call is fast, it's still a network round trip standing where synchronous logic used to be. Multiply that across a page with several "smart" features, and small delays start adding up into something a user can actually feel, even if no single delay is dramatic on its own.
Some of this is self-inflicted, not inherent. Not every AI-related delay comes from the AI itself. Some products deliberately throttle how fast a response streams in, purely for the visual effect of "watching it type," even when the full answer was actually ready almost immediately.
Why it compounds instead of replacing anything
The important part is that none of this is replacing the older sources of page bloat, the images, the trackers, the ad scripts, the frameworks. It's stacking on top of a foundation that was already getting heavier every year. A page that was already carrying a few megabytes of scripts and media now often also carries a chat widget, an AI search layer, and a handful of AI-backed interactions that used to be instant and local.
What this means if you're building these features
- Lazy-load anything AI-related that isn't part of the core experience. A chat widget that loads asynchronously after the page renders shouldn't block anything a visitor actually came for.
- Don't let "smart" replace "fast" for things that were already fast. If a client-side interaction worked well without a network round trip, adding an AI call to it should have a clear payoff, not just because it's possible.
- Be honest about whether streaming delay is earning its keep. If a response is ready, showing it is usually better than manufacturing a typing animation purely for effect.
- Measure the actual Core Web Vitals impact of any AI feature before shipping it, the same way you'd measure any other third-party script. "It's AI" isn't an exemption from the same performance discipline everything else on the page has to meet.
The web probably isn't slower because AI is inherently heavy. It's slower because we're stacking new "thinking" steps and new scripts on top of a foundation that was already trending heavier every year, often without asking whether each addition is actually earning the delay it introduces.
Resources
- Shared Dictionaries: compression that keeps up with the agentic web
- Cloudflare on web bloat and shared dictionaries — data on the decade-long trend of growing page weight and rising agentic traffic: startuphub.ai/ai-news/technology/2026/cloudflare-shrinks-web-bloat-with-shared-dictionaries
- Median web page weight in 2025: 15 years of web bloat
- DebugBear's chat widget performance benchmark — a comparison of real-world JavaScript weight across popular chat widget vendors: debugbear.com/blog/chat-widget-site-performance
- HTTP Archive - Page weight — httparchive.org
- web.dev on Core Web Vitals — Google's official guidance on measuring and improving real-world page performance: web.dev/vitals



Top comments (0)