What percent of webpages will be significantly written or edited by AI in 2030?

TL;DR

My forecast for the qualifying all-webpages estimate is a mean of 32.8%, with a median of 29% (forecast calculation). The central 80% forecast interval is 15%–55.4%, and the probability of exceeding 50% is 14% (forecast calculation). These probabilities are conditional on a qualifying numerical resolution and concern the whole webpage stock, not just newly published articles (target methodology).

Context

The target is a page-count share of publicly accessible English-language webpages. Old pages remain in the denominator, while substantial AI rewrites can move surviving pages into the numerator. The baseline method classifies page body text using Open Pangram’s editing-aware model and a score threshold of at least 0.2; it does not measure the share of words, traffic, or writers using AI (Pew methodology).

The latest all-page observation I verified is 9.60% for July 18, 2026 (Pew chart data). The forecasting problem is how fast that stock changes through new publication, replacement, substantial editing, and changes in what the archive captures.

Evidence

The historical backbone is Pew’s random sample of 10,000 English-language pages from each of 49 crawls, or 490,000 page observations. These are observations, not necessarily unique URLs (Pew methodology). The complete all-page history follows, using the chart data from the study published August 20, 2026; percentages are page-level positive classifications, not verified authorship (Pew, source for every table entry).

The fitting window matters. Anchored extensions of the post-release and last-year fitted slopes give roughly 19% and 30.9%, respectively, at year-end 2030. These are my trend diagnostics, not validated predictive intervals (Pew inputs). I retain the slower path rather than assuming recent acceleration must continue.

The strongest reference class for the legacy stock is URL persistence. Pew’s study published May 17, 2024 tested nearly one million URLs sampled from 2013–2023 for availability in October 2023. About 62% of URLs sampled in 2013 remained accessible a decade later. But accessibility does not establish that their text remained unchanged (Pew digital-decay study). This argues against rapid disappearance of the human-written stock, while leaving room for substantial rewriting.

Incoming content is much more AI-heavy than the stock. Graphite’s May 15, 2026 study examined 55,400 English articles with publication dates from January 2020 through March 2026 and reported 49.9% primarily AI-generated among articles published in the latest quarter. Its authors also report saturation rather than continuing exponential growth. Article markup, publication-date requirements, and its primarily-AI definition make this a selected population, not a replacement baseline (Graphite study).

The Internet Archive study provides a separate check: roughly 35% AI-generated or assisted text in newly appearing material by mid-2025. Its sampling covered August 2022–May 2025, targeted 10,000 URLs per monthly cohort before filtering, capped representation at one URL per host, and required a sufficiently long paragraph. First archival appearance is not necessarily first publication, and the achieved usable sample size was not established from the retrieved material (Dolezal and colleagues). These studies support an AI-heavy inflow without establishing near-universal AI involvement.

A newer, larger number does not overturn the stock baseline. A September 30, 2026 preprint analyzed 310,000 document observations covering January 2021–August 2026 and reported 31.1% AI-labeled tokens in its latest filtered corpus. Its separate 10,000-document audit found that FineWeb filtering retained AI-labeled documents 2.3 times as often as human-labeled documents. Token weighting and selective filtering change the quantity being measured; the paper’s acquisition procedure also changes during its historical coverage (filtered-token study). I use this as evidence about selection and an upside risk, not as an all-page estimate.

Archive discovery is another weak proxy that can mislead. The September archive, collected September 4–17 and released September 19, 2026, contained 2.17 billion pages and 587.2 million URLs not visited in prior crawls. Those URLs were newly discovered by the archive, not necessarily newly written (Common Crawl release). Access rules can also change the sampled mix. Common Crawl honors robots exclusions, and Cloudflare’s September 15, 2026 announcement recommends training restrictions for new advertising-supported sites. Neither source establishes the resulting authorship bias, so I allow for selection changes without applying a fixed correction (Common Crawl FAQ; Cloudflare announcement).

Detector changes matter, but the obvious correction is wrong. Pangram’s July 29, 2026 report tested 14,990 substantially edited texts. Counting either Mixed or AI as positive, its commercial detector comparison rises from about 21% to 58.6%. That is a vendor-controlled benchmark comparing commercial versions—not Pew’s open editing-aware checkpoint at its chosen threshold (Pangram technical report). I therefore apply no automatic uplift to the baseline. Independent editing research shows both that substantial editing can be detected and that human revision of AI text can reduce the signal, supporting uncertainty in both directions (MixDetect, September 26, 2026).

My forecast uses a stock-flow model:

$$\frac{dx}{dt}=r(t)[q(t)-x(t)]+h(t)[1-x(t)].$$

Here, $x$ is the positive-classification share of the sampled stock. The effective renewal rate $r$ combines entry, replacement, growth, and changes in the sampled mix; it is not a measured URL-deletion rate. The share $q$ describes incoming or routinely renewed material. The additional rate $h$ converts otherwise nonpositive surviving pages through substantial AI editing, excluding changes already counted in renewal.

The model starts at the observed baseline and runs for 4.25 years, representing a late-year target snapshot. Initial assumptions are $r=0.09$, $h=0.005$, and $q=0.40$; each changes linearly toward the terminal values below. These are judgmental assumptions, not fitted web-turnover coefficients (reproducible calculation).

All table parameters and outcomes are forecast assumptions or calculations (model specification). The slow branch preserves the evidence from persistent URLs and lower-growth trend fits. Continued diffusion receives the largest weight because the measured stock is growing while incoming content is already AI-heavy. The higher branches require faster renewal, more editing, or a changed crawl composition—not merely better writing models.

I represent each scenario with a bounded beta distribution. The mechanistic mixture receives 98% of the probability mass; a uniform surprise component receives 2%. The widths include adoption, renewal, detection, sampling composition, and snapshot timing rather than just regression residuals. This gives a central 90% interval of 13%–67.6% and a 23% probability of finishing below 20% (distribution calculation).

The code also models reporting precision: 80% effectively unrounded reporting, 10% one-decimal reporting, and 10% whole-percentage reporting. These are format assumptions. Rounded outcomes are placed in the bucket ending at their reported value, rather than spread evenly over adjacent buckets (PMF construction).

What's non-obvious

A plateau in newly published AI articles does not imply a plateau in the whole stock. An AI-heavy inflow can keep raising the stock share while its own prevalence stays constant. Conversely, most publishers using AI does not imply most existing webpages qualify. The stock model captures this distinction, and substantial editing creates a second route that does not require a new URL (article evidence; target definition).

The larger methodological risk is not simply that detectors are inaccurate. Weighting, filtering, extraction, and coverage can move an aggregate estimate even with the same content and detector. Better editing detection can raise the result, while more human-like generation or human revision can lower it. The evidence supports wider tails, not a universal correction factor (filter audit; editing research).

Uncertainties

The main gaps are distinct:

• Text renewal: I did not find a representative panel measuring how often surviving English-language pages receive substantial AI rewrites. A paired content panel would constrain the renewal and editing rates far better than URL availability or first-discovery counts (limits of the persistence evidence).
• Measurement bridging: The retrieved validation does not provide a representative, provenance-labeled webpage sample comparing Pew’s frozen model and threshold with successor detectors. Benchmark gains cannot identify a population correction without that bridge (baseline method; benchmark comparison).
• Snapshot and archive selection: The qualifying snapshot’s timing and sampled composition remain unknown. Using an early-year horizon instead of the late-year model horizon lowers the model mean to 29.4%; this timing uncertainty is included in the distribution widths (sensitivity calculation).

The present estimate is not a hard lower bound: pages can leave the sample, text can change, and methods can correct earlier classifications. The submitted distribution is conditional on a qualifying study yielding a number; annulment under the client’s rules is not represented as a zero-percent outcome.