How Much Electricity Does Bot Web Traffic Actually Use? A transparent order-of-magnitude bound (v1.2)

A transparent, reproducible order-of-magnitude bound on the electricity used by automated (bot) web traffic. No single honest figure is recoverable from public data. A naive calculation, multiplying a bot share by total data-centre energy at an average allocation, gives about 257 TWh per year, but that is inflated roughly tenfold by three implicit choices: multiplying a request-share by a pool full of non-request workloads, using a web-application bot fraction where a whole-network one belongs, and allocating by average rather than marginal energy. Two largely separate methods then bracket the answer at order 10 TWh per year, plausibly 1 to 70, while disagreeing on the central value by roughly five times. That disagreement is itself the finding. For scale, 10 TWh is 0.035 per cent of global electricity, against a denominator of 28,200 TWh, the IEA's global electricity consumption for 2025. The headline is the marginal quantity, meaning the energy that would be freed if bot traffic stopped; the attributional share-of-energy-used quantity is larger and is reported alongside it. Version 1.2 is an anchor refresh. A withdrawn headline framing was found still shipping inside the version 1.1 reproduction script, which labelled the top-down centre as the headline and printed the bound as 0.048 per cent of global electricity while the paper text said otherwise; the two files in the same record disagreed, and the script has been corrected. The global-electricity denominator is corrected from 30,500 TWh, an undated figure on an apparent generation basis, to the IEA's 28,200 TWh on a consumption basis, matching the numerator. Both bot-fraction anchors are now dated, and the whole-network figure is a first-party pull from the Cloudflare Radar API on 1 June 2026 at 32.8 per cent rather than a press snapshot. The bottom-up cross-check is rescaled on W3Techs at 24.1 per cent, which widens the disagreement between the two methods from roughly four times to 5.3 times, so refreshing an input strengthened the paper's own finding. A third and narrower bot frame now in wide circulation, covering HTML requests only, is documented but kept outside the bound, because the energy pool is not HTML-only. The claim that only hyperscalers and content delivery networks hold the joined traffic-and-energy telemetry is re-tested against Cloudflare Radar's May 2026 additions and survives, because every addition is on the traffic axis and none carries an energy axis. The headline bound is unchanged from version 1.1. An adversarial review before publication produced five further changes, none of which moves a number: the idle-power range is anchored to LBNL's 2025 Update rather than to Barroso and Hoelzle alone; the description of the two methods as independent is withdrawn as overstated, since they share the bot fraction, which cancels from their ratio and so cannot account for the disagreement between them; the reason the data-centre and transmission energy pools are not double-counted is stated explicitly; the 57.5 per cent figure is treated as a Cloudflare telemetry reading announced informally rather than as an uncitable relay, while remaining outside the bound because the pool is not HTML-only; and the honesty ledger gains the unsegmented joules-per-request range and three unmodelled upward mechanisms. Method-first, value-neutral, reproducible from free primary sources with a standard-library script. Independent analysis, not investment advice. Produced with assistance from Claude Opus 4.8, an Anthropic model; Anthropic is a participant in the AI industry whose energy footprint this work assesses, and it also operates one of the AI crawlers counted in the bot traffic measured here, so the assisting model is not a neutral party to the subject matter and the primary sources cited should be preferred to this summary.