How it's calculated

Each activity is converted to electricity used by the device, the network, and the data center, then to emissions at the selected grid intensity. eGRID rates are per unit of electricity generated, so electricity is grossed up for 5.1% transmission and distribution losses, as EPA does in its own equivalencies calculator.

Marketing email and advertising factors are published as CO2e per message or per 1,000 impressions. They are used as published and are not adjusted for the selected grid.

Every figure has three cases: every factor at its low published value, every factor at its central value, and every factor at its high value. Totals sum each case across activities. The results show the central case. Show ranges adds the span from the smallest of the three to the largest. For a ratio the low case can come out larger than the central case, so the range always shows the smaller number first and doesn't say which case produced it. The ranges are wide.

The model is attributional. It charges each activity its share of network and data-center energy, most of which would be spent whether or not the activity happened. See Limitations.

• Screen time: hours x (phone W (F15) + MB per hour (F16) x network kWh per GB (F3) + data center Wh per hour (F17)).
• AI tasks: data center Wh per task (F23). Full stack adds device minutes and MB sent per task (F27). Tasks that equal your year: your data-center kWh / Wh per task, or your full stack kWh / full stack Wh per task.
• AI video estimate: AI videos a day x 365 x attributed generations per view (F29) x Wh per clip (F23).
• Streaming, devices, phone, cloud, and web: per-item power or data volume (F30 to F40) at F3. Phone manufacturing: CO2e per year of ownership (F42), no grid factor.
• What changes it most: each line sets one input to zero (or keeps the phone a year longer) and recalculates the year. Unsubscribing removes the marketing email CO2e of the stores whose rewards programs you use.
• Companies, retailers, and restaurants: email = emails per member per year x marketing email factor (F7). Advertising = impressions per member per year / 1,000 x gCO2PM (F9).
• Companies with no retail media network whose policies say customer data is used for advertising elsewhere: impressions = the sector median, marked imputed.
• Unnamed brokers that supply a selected company, and unnamed brokers a selected company sells or shares with, one stream per company: bytes per year = attributes (200 / 3,000 / 10,000, assumption) x F12 x F13, at F3.
• Cars, carriers, apps, brokers, and travel: records per year = records per driving hour x driving hours (F11), or records per device-day x 365. Profiles: bytes per year = attributes x bytes per attribute (F12) x deliveries per year (F13). GB = records x bytes per record (F10) / 10^9. Emissions = GB x network kWh per GB (F3), grossed up for losses, at the selected grid.
• Emissions = kWh / (1 - 0.051) x grid intensity + fixed CO2e factors.

A view pays the account that posted the video, in money or in reach, and the account uses both to decide whether to make the next one. The estimate assigns each AI video you watch a share of the generations behind the next AI video, in proportion to what one view is worth against what one generation costs.

The generation is other people's, not your tech's, so the estimate is shown beside your totals and not in them.

This is an accounting model, not a measurement. No study measures how many generations one extra view causes. The premise rests on creator testimony, platform policy changes, and one qualitative study.

views needed to justify the next video = cost per posted video / payout per view attributed generations per view = generations per posted video / views needed generations a year = AI videos a day x 365 x attributed generations per view energy = generations x Wh per generation

AI videos watched are estimated from short-video hours: hours x videos per hour (F18) x the AI share of the feed (F28). Feed apps and long-form YouTube aren't counted: no audit of their AI share exists and not every item is a video. The results page shows the count at your answers, under Where it goes, and takes your own count under Change your answers.

Limits

Every layer is converted at the selected grid with the same 5.1% loss gross-up, data centers included, so AI and apps are converted alike. The companies' own location-based rates for their data centers (Meta 0.3246 and Google 0.3476 kg CO2e per kWh) are within 8% of the U.S. average.

The research was compiled on September 25, 2026, in four streams: feeds, AI use, AI video, and everyday technology. Figures are the research's own, with three exceptions set out in the factors below. Network energy uses F3 rather than each research file's own bounds. Cloud gaming network energy uses LBNL's per-user power, which is not a per-GB figure. Phone manufacturing is reported in CO2e rather than the research's energy equivalent.

Network energy per GB is the largest single source of uncertainty. Published estimates span about 100 times, from about 0.002 kWh per GB (the Aslan trend carried to 2025) to 0.23 for 4G radio access and up to 1.0 in the DIMPACT review. At the low case the data center is the largest layer for every app. At the central and high cases the network is. F3 is applied to every gigabyte so the figures move together when it changes.

No double counting. Each hour of an app is device plus network plus data center. Advertising, ad decisioning, telemetry, and recommendation inference are shares of those three layers and are shown beneath them, never added. The streaming lines leave out the home router, which is its own line.

Tracking is counted, not charged. Telemetry, smart TV screen captures, location reports, operating system check-ins, and notifications are shown as events under What's being logged. Their transfer energy is already inside the network lines. The research found the energy of tracking small and its event counts large.

The questions. The questions on the first page write into the detailed inputs; there is one calculation. Instagram and YouTube hours are split evenly between their two feeds, the Typical adult answers' own assumption. Turning the smart TV on with no TV hours sets them to the streaming hours, as the Typical adult answers do.

Company profiles were compiled from public sources on September 25, 2026, by three researchers: grocery and mass retail; pharmacy, beauty, specialty retail, and restaurants; and companies outside stores. Each company's page lists its sources. Where the research says alleged, pending, or settled without admitting wrongdoing, so does this site. Enforcement statuses change: a matter shown as pending might have settled, been dismissed, or gone to arbitration since.

Impressions per member. The researchers converted each retail media network's reported or estimated ad revenue to impressions: revenue / CPM x 1,000 / members, at $20, $10, and $5 CPM for the low, central, and high cases (F14). Retail media mixes CPM and CPC pricing, so the result is an order of magnitude.

Email per member. Measured counts where the research found one (Costco, from a 42-retailer inbox study; Ulta and Sephora, from a beauty-category study; Bath & Body Works, from a secondary summary). Otherwise the researchers' labeled assumptions.

Imputation. Where a retailer's email count or impressions are null, the value is the median of the companies in the same sector that have one, taken separately for each case, and is marked imputed. Rite Aid is the exception for email: its stores closed in 2025 and its brand and loyalty data were sold, so it is counted at zero emails. Wegmans is still operating, so its missing email count is imputed.

Companies with no retail media network. Dunkin', Bath & Body Works, and Starbucks run no ad network of their own, but their policies say customer data goes to advertising partners or is used to serve ads. Each takes the sector median impressions, marked imputed. Rite Aid has no such statement in the research and stays at zero impressions. Wegmans runs a retail media platform with undisclosed revenue and shares customer purchase data with advertising platforms, so it takes the sector median, marked imputed. Outside retail, a company with no usable volume figure takes the median data volume of its sector, marked imputed. Automakers and apps with no published sampling rate borrow the GM and Arity rates. Medians are taken over all 52 profiles, brokers included, so an imputed value doesn't depend on the list.

Five components, each scored 0 to 3 by the researchers, for a total out of 15. The research files call this the worst offender score. A component the research couldn't score is null. Null means unknown, not zero, and the total is then out of the components scored.

Ranking of all 52 organizations

Ranked by data practices score. Ties share a rank and follow each researcher's own ranking. Where a component is unscored, the total is out of the components scored. S: sells or shares. E: enforcement. M: retail media scale. D: sensitive data. O: opt-out friction.

Codes F2, F4, F5, F6, F8, and F41 are not used.

Grid intensity

Everyday comparisons

The first Your year line on the results page ends with two comparisons: your year's electricity in months of a typical refrigerator's use (kWh / Q1 x 12), and its CO2e in miles driven (kg CO2e / Q2). Neither factor enters any total.

Per-year lines use the Typical adult hours.

One method covers the water line on the results page and the water per AI task: water in liters = kWh x a water factor in liters per kWh. Data-center kWh already include cooling and power-delivery overhead (PUE). The AI video estimate is left out, as it is from the other totals.

data-center water = data-center kWh x data-center factor devices and network water = (all kWh - data-center kWh) x grid factor water line = data-center water + devices and network water

The water line gives the total and names the data-center part. Water per AI task uses the data-center factor only.

What the figures measure

• Withdrawal: water taken from a river, lake, aquifer, or utility. Some of it returns, for example as cooling-tower discharge or wastewater.
• Consumption: withdrawn water that doesn't return to its source, mostly because it evaporates. Every factor here is consumption.
• Evaporation: water lost from a lake or reservoir surface, gross or net of the rain that falls on it.
• Lifecycle footprint: all the water behind a product, mostly rain stored in soil for crops. Only its irrigation share is comparable with cooling water.

Known limits

In short, on the water page

The water page opens with three lines. The first two come from this arithmetic, with the chatbot prompt's central energy and water factor and the LBNL and USGS rows in the water page's scale reference.

chatbot prompt: 0.3 Wh x 4.29 mL per Wh = 1.29 mL 1 U.S. teaspoon = 4.93 mL (1/6 U.S. fluid ounce, 4.92892 mL) 1.29 mL / 4.93 mL = 0.26 teaspoon, about a quarter U.S. data centers on site, 2023: 66 billion L a year / 365.25 = 181 million L a day share of consumption, three largest uses, lower 48, 2010 to 2020: 181 million / 314 billion L a day = 0.058%

The first comparison is consumption against withdrawal: USGS's 2015 compilation is the latest that covers every use, and it reports withdrawals. The second compares consumption with consumption, but covers only crop irrigation, thermoelectric power, and public supply.

Reference values, the water cycle, and local context are on Water, data centers, and AI.

• Your own AI use. The calculator shows AI tasks for scale and takes no AI inputs.
• AI model training, except as a note for chatbot prompts. AI music and voice generation, for want of a measurement. Deep research queries, for which no measurement exists.
• Embodied emissions of every device except the phone, and of network equipment.
• Real-time bidding requests that did not result in an impression, except as already included in the advertising factor's ad selection component.
• Advertising served on third-party platforms using data the companies say they share, since no per-member volume was published. The exception is a company with no retail media network whose policy says its customer data is used for advertising; it takes the sector median.
• Model training and scoring that companies run on data bought from brokers. No figure is published.
• Onward resale by data brokers and insurers after the first recipient. Each company line counts only its own copy.
• Storage of telemetry and profiles, including the five-year location retention the FTC alleged at InMarket.
• Records exfiltrated in breaches.
• Handset radio and screen wake-up energy for background location fixes, telemetry, and notifications. No source measured it.
• Server-side energy for smart TV content matching, voice assistant queries, multiplayer game servers, ride-hailing tracking, and dating-app ranking. No company publishes it.

• Network energy per GB is the least settled number in the model. Published estimates span roughly two orders of magnitude, depending on year, boundary, and whether mobile access is included.
• Network power is mostly fixed baseload. Mytton, Lundén, and Malmodin (2024) show that energy doesn't scale linearly with data, so an average kWh per GB overstates what stopping any single activity saves. The model uses the average factor anyway, in line with common industry practice. The same applies to What changes it most.
• Company impressions rest on revenue figures of mixed quality: some reported in filings, some from a vendor blog the research rates low-confidence, some stale. Member denominators range from weekly visits to monthly actives. Each company's page says which.
• Bytes per record, profile attribute sizes, and delivery frequency are labeled assumptions. Telemetry payloads are not published.
• Programs reported as discontinued are counted at zero energy: Verisk's driving-data product (ended April 2024), Oracle Advertising (ended September 30, 2024), and ARC's Travel Intelligence Program (ended November 2025, per an August 26, 2026, letter from Senators Wyden and Brown). Whether copies already delivered were deleted is not known, so the copies stay in the count.
• Organizations holding a copy of your data is a count of records, not of distinct organizations. The large ad platforms appear in many company records and are counted in each.
• Items per hour, ad load, time spent per user, and each platform's share of its parent's data centers are assumptions for most apps. Each is labeled where it is used.
• Commercial AI image and video energy are third-party estimates. No provider publishes energy per output.

Every source behind these figures is on the Sources page, with its URL, date, kind, and what uses it.