A traffic source 'growing 693%' sounds huge — but what if it started near zero and is still tiny?

This explores how a big growth percentage can make a small change sound huge, mostly in the current debate over how much web traffic AI sends or takes away. The clearest case in the corpus is Adobe's holiday 2025 report. It found that retail traffic referred by AI assistants grew 693.4% year over year, and that those visitors converted 31% better than other traffic How much traffic do AI assistants actually send to retailers?. The math is correct. But Lipsman checked it against Sensor Tower data and found that AI referrals went from about 0.1% to 0.6% of ecommerce traffic Does Adobe's AI referral traffic growth statistic hold up under scrutiny?. When you start from nearly zero, any growth looks explosive as a percentage. The useful question is not 'how fast did it grow?' but 'what share of the total is it now?' Half a percent is a real trend, but it is not yet a channel that changes a retailer's business.

The same issue works in reverse when traffic falls. You need to ask whether a drop is relative (a percentage of the old figure) or absolute (percentage points). Pew's browsing data shows people clicked a search result 8% of the time when Google showed an AI summary, versus 15% when it didn't Do AI summaries on Google reduce clicks to actual websites?. The headline says 'nearly half.' The absolute gap is 7 points of click rate. Both are accurate, but they feel very different. The randomized experiment by Agarwal and Sen reports a 39.8% relative cut in clicks to publishers Do AI Overviews actually reduce clicks to publisher sites?. Meanwhile, the Wikipedia natural experiment finds a modest 5.45% drop in search referrals to English Wikipedia Does AI search summaries divert traffic away from Wikipedia?. Publishers surveyed by the Reuters Institute expect a 40–43% decline over three years Will AI Overviews reduce search referral traffic to publishers?. That is a forecast, not a measurement, so it sits on a different level of evidence. Put these side by side and the 'AI is killing traffic' story turns out to range from about 5% to about 40%, depending on the site, the measurement method and whether the number is observed or predicted.

A second trap is when the headline and the data measure different things. Cloudflare's report is framed as 'non-human traffic now exceeds half of all internet activity.' Its key number, though, is that 52% of *crawler requests* are for AI training, up from 22% Is non-human traffic really half of all internet use?. A share of crawler traffic is not a share of all traffic. Before you trust a headline percentage, ask: a percentage of *what*?

The lesson goes beyond traffic statistics. Elsewhere in the corpus, the same pattern shows up in AI evaluation: a single summary number hides what matters underneath. Strong overall accuracy can hide confident wrong answers clustered in the rare cases where they cause harm Why do confident wrong answers hide in standard accuracy metrics?. Search answers can stay accurate while the sources behind them quietly collapse into synthetic content Does synthetic content in search results hide ecosystem decay?. Averaging a model's confidence across a whole reasoning chain can hide the single step where the reasoning broke Does step-level confidence outperform global averaging for trace filtering?. Growth rates, averages and aggregate accuracy all compress the data into one number, and each throws away the context you need to judge it. When you see a dramatic number, look for three things: the starting value, the absolute size, and what exactly is being counted.