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

TL;DR

My median forecast is 29% of English-language webpages, with a predictive mean of 31.7% (forecast calculation). The central 80% predictive interval is 15%–54.8%, and the probability of exceeding 50% is 13% (forecast calculation). This forecast includes old pages and is conditional on a qualifying numerical resolution, not annulment (resolution rules).

Context

<a id="resolution-rules"></a>The information cutoff is October 11, 2026, at 11:20 UTC. The supplied rules target all qualifying English-language pages in the latest qualifying snapshot from calendar year 2030, including older and undated pages; historical observations inform the forecast but cannot resolve it early. If no comparable study appears by December 31, 2032, the question annuls rather than resolving to zero (supplied resolution rules).

The outcome is a study’s page-level classification rate, not the percentage of words generated by AI or proof of individual authorship. Pew’s baseline measures body text and explicitly treats its detector as probabilistic (Pew methodology).

Evidence

The historical backbone is Pew’s study published August 20, 2026: 10,000 English-language pages from each of 49 Common Crawl snapshots, totaling 490,000 page observations. It flags meaningful AI writing or editing at an Open Pangram/EditLens score of at least 0.2. Only 10%–15% of pages have identifiable publication dates, and Pew says that subset is nonrandom (Pew methodology). (pewresearch.org)

The complete all-page history follows. Values are percentages of sampled pages, extracted from the article’s embedded chart data—not reconstructed from domain averages. Coverage runs from January 27, 2021, through July 18, 2026; the retrieved article metadata marks September 17, 2026, as its modification date (Pew all-page data). (pewresearch.org)

Pew warns that the early readings include false positives. Its comparison of 62,370 pages from seven crawls found 96% agreement with commercial Pangram 3.3, but agreement between detectors is not ground-truth accuracy (Pew methodology). I therefore exclude pre-ChatGPT observations from growth fitting and forecast the comparable reported classification rate, without an exact correction for latent authorship.

My equal-weighted, calendar-date fits show acceleration. The linear slope is 2.1 percentage points per year across all post-ChatGPT observations, versus 4.8 over the latest year. Fixed-ceiling sigmoid fits produce very different futures despite similar historical errors. These are my calculations from the full series, not Pew predictions (underlying observations).

The ceiling is not identified by the history. I use these fits as checks on the reasonable range, not as models whose small residuals justify narrow forecast intervals.

New-content evidence supports a higher incoming AI share, but also a slowdown in adoption. Graphite’s May 15, 2026 study sampled 55,400 English articles and listicles published from January 2020 through March 2026, requiring article metadata and at least 100 words. It averaged three detectors’ primarily-AI classifications; its full quarterly history appears below (study and definitions, complete source data). (graphite.io)

This is a plateau in a selected article population, not the entire web. Its primarily-AI definition also differs from meaningful editing, and the study did not validate detection after substantial human rewriting (Graphite limitations). An April 14, 2026 preprint provides weaker corroboration: its mid-2025 estimate was 35% AI-generated or assisted, but it used first archival dates, a one-URL-per-host cap, and long paragraphs; the retained analytical sample size is not clearly reported (Dolezal et al.).

Old material slows the transition. Pew’s May 17, 2024 digital-decay study sampled approximately one million URLs from 2013–2023 and checked accessibility during October 12–November 6, 2023. It found 38% of the 2013 cohort inaccessible a decade later. That measures disappearance, not unchanged text, and is not an English-page renewal estimate (findings, methodology). (pewresearch.org)

Crawler discovery is also not publication. Common Crawl’s September 19, 2026 release contained 2.17 billion captures collected September 4–17, including 587.2 million previously unvisited URLs. Those multilingual captures cannot be annualized into an English-text replacement rate (Common Crawl release).

Existing-page rewriting is a real channel. Shopify reported on March 24, 2026 that merchants created 19.8 million product descriptions in the first three weeks after its Winter ’26 release, which launched December 10, 2025. Merchant sample size, unique published URLs, language, and old-versus-new descriptions were not disclosed (Shopify production figures, release date). WordPress.com’s March 20 announcement, updated May 15, enables agents to update published pages, but requires enabled permissions and approval (WordPress.com announcement). These sources support scalable rewriting, not blanket conversion of the legacy web.

Economic incentives restrain a simple flood scenario. Google’s guidance, updated October 1, 2026, targets mass generation without user value, not useful AI-assisted writing generally. Search exclusion also does not mean removal from a public crawl. I treat this as a brake on indiscriminate publishing, not a cap on AI prevalence (Google guidance).

Detector evolution moves the result in both directions. Pangram’s July 29, 2026 vendor report scored the same 14,990 substantially edited student texts with two detector versions: Mixed-or-AI classifications rose from 3,113 to 8,789, or 21% to 58.6%. This is a selected benchmark, not a web audit or a comparison at Pew’s threshold (Pangram 4 report). (arxiv.org) Conversely, a September 25 experiment on 400 deliberately constructed responses reduced EditLens detection from 55% to 6% through base-model orchestration, using a separately calibrated threshold. That establishes an evasion mechanism, not its prevalence online (Dhingra and Pruthi).

<a id="forecast-model"></a>I combine this evidence through an effective stock-adjustment model:

$$\frac{ds}{dt}=\lambda\,[q(t)-s(t)].$$

Here, $s$ is the all-page AI-positive fraction, $q$ is the comparable classified fraction of entering or substantially rewritten material, and $\lambda$ is the annual adjustment rate. It summarizes additions, removals, rewriting, and crawl composition—not a measured page-death hazard. I start at the measured baseline and let $q$ rise linearly within each scenario. The following parameters, weights, and uncertainty spreads are my judgments, encoded in the accompanying calculation (forecast model).

The central scenario begins with growth of 4.2 percentage points per year, consistent with recent acceleration without extending exponential growth indefinitely. The slow scenario preserves the long-window extrapolation and persistent human stock. The upper scenarios preserve materially different futures involving faster rewriting, automated publication, and stronger editing detection; the downturn scenario allows a future measurement below today’s baseline (forecast model).

I integrate snapshot timing over March 31, July 18, and December 31 of 2030, with weights of 15%, 35%, and 50%. These are assumptions, not announced crawls. Each scenario has bounded beta uncertainty, followed by a nominal 10,000-page binomial sample. Reporting precision receives weights of 60% for hundredths, 10% for tenths, and 30% for whole percentages; this puts extra mass into buckets ending at integers when only rounded results are available (forecast model).

The resulting median is 29%, the mean 31.7%, and the central 90% interval 12%–64.3%. It assigns 13% probability above 50%, 4% at or below 10%, and 1% above 80%. These are predictive uncertainties about the eventual reported result, not sampling confidence intervals. No prediction-market signal was used (forecast calculation).

What's non-obvious

The physical web and the sampled web do not age at the same rate. Common Crawl says it recrawls slowly changing pages less often so it can focus on newly discovered and frequently changing pages. A persistent human-written page can therefore remain online while becoming less represented in snapshots (coverage explanation). Selection also explains some apparently conflicting AI estimates: a September 30 preprint’s audit of 10,000 raw captures found AI-labeled documents survived FineWeb filtering 2.3 times as often as human-labeled documents. Filtered-token estimates cannot be converted into all-page estimates with a fixed multiplier (filtering audit).

Human ideas can still produce AI-written pages. An October 5 paper separates idea provenance from prose provenance and demonstrates that detectors aimed at these concepts behave differently (IdeaLens study). This question counts substantial AI wording or editing even when a human supplied the argument. A future detector that measures only AI-origin ideas would change the target rather than improve its measurement. Likewise, old publication dates do not protect pages from becoming AI-positive through later rewriting (Pew’s editing definition).

Uncertainties

The biggest gaps are linked rather than independent:

• The rate of substantial rewriting on previously published English URLs is unmeasured. Platform generation counts do not establish acceptance, publication, unique-page counts, or human-to-AI conversion (Shopify disclosure).
• Detector upgrades need a representative, paired web evaluation with recorded writing histories. Current editing and evasion benchmarks do not supply that bridge, and detector agreement does not establish accuracy (Pangram report, Pew validation).
• Crawl coverage and page-type composition can change independently of authorship. Login-only and paywalled content is underrepresented, while changing pages receive more crawl attention (Pew coverage limits, Common Crawl coverage).
• Study availability, snapshot timing, and methodological comparability remain resolution risks. An unavailable or nonqualifying estimate leads to annulment, not a numerical outcome at the bottom of the distribution (resolution rules).

A repeated-URL panel combining publishing logs with old and new detectors would close the largest gap. Until then, the evidence supports a rising all-page share, but it does not identify a single renewal rate or detector correction. The forecast retains substantial probability on both slow accumulation and rapid rewriting instead of hiding that uncertainty inside a precise point estimate (forecast model).