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

TL;DR

My median forecast is 29% of all qualifying English-language webpages in the resolving snapshot (forecast model below). The central 80% forecast interval is 14%–57.5%, with a 15% probability of a result above 50% (distribution calculation). This forecasts the study-classified share of the whole sampled web, conditional on a numerical resolution—not the share of newly published articles or AI-generated words.

Context

The correct starting point is the July 18, 2026 all-page reading of 9.60%, not the 35.09% reading for dated pages published after ChatGPT’s release (Pew baseline, published August 20, 2026). Only about 10%–15% of sampled pages had identifiable publication dates, and that subset was nonrandom (Pew methodology). Using the larger number as the baseline would substantially overstate the quantity being forecast.

The target is a page-level text classification, not proven authorship. The sampled web also excludes some material available to human readers: paywalled and login-required pages are underrepresented (Pew methodology). The numerical distribution is conditional on a qualifying study; annulment is not assigned to any percentage bucket.

Evidence

The historical backbone consists of 49 crawl samples, each containing 10,000 English-language pages, or 490,000 page-snapshot observations (Pew sampling methodology). Below is the complete all-page history from the published interactive chart. Dates identify snapshots, not original publication dates; values are percentages of pages classified as showing meaningful AI writing or editing (Pew chart data, August 20, 2026 vintage). Read each row from left to right.

I fitted calendar-time linear trends and anchored their projections to the last observation, rather than to a fitted endpoint that misses recent acceleration. These calculations give materially different answers depending on the historical window (input observations):

These are extrapolation checks, not statistical prediction intervals. The long-window estimate preserves early slow adoption; the short-window estimate preserves recent acceleration. Neither establishes a future saturation level. I preserve that disagreement rather than treating several fits to the same data as independent confirmation.

Legacy content supplies real inertia. A study published May 17, 2024 checked nearly one million URLs drawn from 2013–2023 snapshots in October 2023. Only 38% of the oldest snapshot’s URLs were confirmed inaccessible after a decade (Pew digital-decay study). But URL survival does not prove text survival: an old page can be substantially rewritten without changing its address.

Nor does crawler discovery establish publication. The September 2026 crawl, collected September 4–17 and announced September 19, contained about 2.17 billion captures, including 587.2 million previously unvisited URLs—roughly 27.1% by division of the rounded counts (Common Crawl release). Those are all-language capture counts. A newly discovered URL can contain old text, while a repeatedly captured URL can contain new text. I do not insert this discovery percentage into an annual replacement model.

The incoming-content evidence is much more AI-heavy. A May 2026 analysis sampled 55,400 English-language articles and listicles with publication dates spanning January 2020–March 2026, requiring article markup and at least 100 words. It averaged three detectors (Graphite methodology). Its complete publication-cohort history follows; these are retrospective classifications of article cohorts, not comparable all-page snapshots (published source data).

The recent plateau argues against automatic exponential growth. But this sample excludes many page types, and its tests did not validate naturally occurring workflows in which people rewrite AI drafts (Graphite limitations). I use it to bound plausible incoming adoption, not to replace the broad-web baseline.

Publishing infrastructure provides a stronger leading signal than generic chatbot-use surveys. WordPress.com announced direct agent creation, editing and publication capabilities on March 20, 2026, with the announcement updated May 15 (agent publishing announcement). Its May 8 changelog expanded the opt-in editor assistant to all current paid plans (editor rollout). This is WordPress.com, not the entire self-hosted ecosystem. My inference is that native assistance makes substantive editing of existing pages cheaper, creating a growth path beyond new articles. Availability alone does not establish actual publication rates.

Search incentives provide a brake, not a ban. Google targets mass-produced content intended to manipulate rankings without adding value, including AI-generated material; it does not prohibit useful AI assistance (Google spam policy). Lower search visibility can reduce publishing returns without removing those pages from a broad archive.

Measurement can move the answer in either direction. In a July 29, 2026 vendor report, the same 4,826 synthetically and substantially edited texts received AI-or-Mixed classifications in 34.7% of cases under Pangram 3.3.2 versus 72.2% under Pangram 4 (technical report, Table 6). This demonstrates method sensitivity, not a correction factor for random webpages. Conversely, external research dated September 25 demonstrated detector evasion through agent-assembled base-model text, with experiments using 100 examples per benchmark and substantially higher generation costs (HALO study). Neither experiment estimates ordinary-web prevalence. Together they justify wide, two-sided measurement uncertainty.

My quantitative model treats the target as a stock renewed by new or substantially rewritten content:

$$\frac{dp}{dt}=r\,[q(t)-p(t)].$$

Here $p$ is the classified share of the sampled stock, $q$ is the comparable classified share of incoming or substantially refreshed text, and $r$ is an effective renewal hazard per year. Renewal includes new publication, major edits, deletion and changing crawl inclusion. Major legacy editing is included once, not added again as a separate conversion channel. The code starts at the measured baseline and lets $q$ change linearly over the horizon.

The following are my forecasting assumptions, not measured turnover rates or fitted adoption ceilings. Timing is centered on a late-year snapshot; the outcome spreads cover earlier or later qualifying snapshots (reproducible calculation).

The central case produces an initial increase of about 4.3 percentage points annually, close to the recent trend calculation. Its rising incoming share reflects wider substantive assistance; its moderate renewal rate preserves substantial legacy content. The slow case retains the plateau and persistence evidence. The fast cases retain the possibility of a structural change in publishing economics rather than assuming today’s constraints last indefinitely.

Each scenario becomes a bounded beta distribution matching its center and stated standard deviation. A further 2% weight is uniform over the allowed range for unmodeled but qualifying regimes. These widths cover parameter, detector, crawl-composition and timing uncertainty. They deliberately do not shrink merely because the historical sample is large.

The code also models reporting precision: 60% weight on hundredths of a percentage point, 15% on tenths, and 25% on whole percentages. These are judgmental reporting assumptions. Whole-number reporting puts more mass in integer-ending buckets. The resulting distribution has a median of 29%, a mean of 32.4%, and a central 90% interval of 11%–68%; the right tail raises the mean above the median (distribution calculation).

What's non-obvious

A quality filter can increase the apparent AI share. A September 30, 2026 preprint audited 10,000 raw crawl documents: FineWeb retained 29.3% of AI-labeled documents versus 12.8% of human-labeled documents. Even among English documents, its quality heuristics retained about 64% versus 39% (filtering audit). The paper is vendor-linked, and its labels are not ground truth. Still, it identifies a concrete selection effect: cleaner-looking training corpora need not represent the whole sampled web. Filtered token estimates therefore cannot simply replace a raw, page-weighted baseline.

Crawler access creates another selection effect. Cloudflare’s managed robots file explicitly blocks CCBot (managed-robots documentation), and Common Crawl respects those exclusions (crawler FAQ). If human-heavy publishers opt out more often than automated publishers, the archive’s classified share can rise without an equal rise across the public web; the reverse is also possible. I could not quantify that difference. This adds uncertainty about the denominator, not evidence for a fixed upward adjustment.

Uncertainties

The largest missing input is a representative panel measuring how often existing English-language pages receive substantial AI edits. URL disappearance and first crawler discovery do not answer that question. A second gap is independent calibration on naturally mixed webpages: the synthetic mixed-edit gains do not establish how many randomly sampled pages a newer detector would reclassify (Pangram evaluation). A third is changing crawl access and composition, including whether a future study can retain a sufficiently representative denominator.

Sampling noise is small beside these gaps. At the baseline prevalence and 10,000 independent pages, the binomial standard error is about 0.3 percentage points—a calculation using the reported sample size (Pew methodology). That does not cover correlated pages, classifier error or a changing sampling frame. I therefore keep a broad forecast, including a 4% probability of an outcome at or below 10%, rather than treating the current reading as a hard lower bound (forecast distribution). The precise probability array represents numerical precision in this model, not equally precise knowledge of the future.