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

TL;DR

My central forecast is 30%, with an expected value of 32.8% (calculation below). The central 80% predictive range is 14%–56%, and the probability of a result above 50% is 15% (distribution below). This forecast counts all qualifying webpages, not just new content, and is conditional on a numerical resolution rather than annulment (interpretation below).

Context

The target is a page-count share in the latest qualifying calendar-2030 snapshot. Old pages remain in the denominator; tokens, traffic, newly published articles, and non-textual AI use do not replace that denominator. I forecast the study’s classification statistic, not proven authorship, following the concept defined by Pew’s baseline and methodology.

The directly comparable starting point is 9.60% in the July 2026 all-page sample, rounded to 10% in the headline. The separate 35% estimate concerns dated post-ChatGPT pages, a nonrandom subset, and cannot serve as the all-web baseline (Pew, August 20, 2026; sampling caveat). (pewresearch.org)

Evidence

The historical backbone is Pew’s random sample of 10,000 English-language pages from each of 49 crawls spanning January 2021–July 2026: 490,000 page observations, not necessarily distinct pages. It classified extracted body text using Open Pangram at a score of at least 0.2. Only 10%–15% of pages exposed publication dates (Pew methodology, August 20, 2026). (pewresearch.org)

The complete all-page history follows. Values are percentages of sampled pages classified positive, decoded from the published chart rather than reconstructed from selected domain averages. The source was published August 20, 2026; its top-level modification metadata gives September 17, 2026 (Pew chart and article).

These observations support acceleration, but not one uniquely identified growth curve. I fitted straight lines anchored at the latest observation and bounded log-odds trends with assumed ceilings. The following are my December-2030 calculations from the full Pew series, not published predictions.

The ceilings are assumptions. These fits reuse the same evidence, so they are sensitivity checks, not independent forecasts to average. Their disagreement is much larger than the approximately 0.3-percentage-point sampling standard error calculated for the latest sample under an independent-binomial approximation (sample design).

Evidence about incoming content supplies a second anchor. Graphite’s commercially produced study selected 55,400 dated English articles and listicles, at least 100 words long, published January 2020–March 2026. It averaged three detector classifications. Its complete quarterly history shows rapid adoption followed by a plateau near half, rather than uninterrupted exponential growth (Graphite, May 15, 2026; March crawl and April–May detector vintage). (graphite.io)

That study targets primarily AI-generated articles, not Pew’s broader substantial-editing concept or all-page denominator. I use its plateau to support a slowdown scenario, not a universal ceiling (Graphite’s classifications and limitations).

Legacy content also matters. Pew sampled just under one million URLs across annual 2013–2023 crawls and checked accessibility in October 2023. It found 38% of the 2013 cohort inaccessible, leaving most accessible a decade later. That is evidence of persistence, but it measures URL accessibility—not unchanged text or denominator weight after new pages enter (Pew, May 17, 2024). (pewresearch.org)

Publishing tools provide a route around that legacy constraint. WordPress.com opened opt-in AI editing assistance across its current paid plans in May 2026. This is feature availability on WordPress.com, not measured uptake across the entire WordPress ecosystem (official May 8 announcement). Search incentives pull the other way: Google warns against scaled page generation without user value, rather than banning useful AI assistance. I read this as a headwind to commodity publishing, not a numerical cap on crawl prevalence (Google guidance). (wordpress.com)

Measurement changes deserve substantial weight. In a vendor-controlled evaluation of 14,990 substantially AI-edited student essays, the share labeled Mixed or AI rose from 20.8% under Pangram 3.3.2 to 58.6% under Pangram 4. This demonstrates that improved mixed-authorship sensitivity can produce large reclassification. It does not supply a correction to a random webpage sample or Pew’s open-model threshold (Pangram technical report, July 29, 2026). (arxiv.org)

I translate these forces into an effective stock-renewal model:

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

Here, $p$ is the positive share of the sampled page stock; $\lambda$ is an annual replacement, rewriting, and dilution hazard; and $q(t)$ is the positive share of incoming or substantially refreshed content. It rises linearly from $q_0$ to $q_T$. Starting from $p_0=0.096$, the endpoint is:

$$p(T)=q_0+(p_0-q_0)e^{-\lambda T}+\frac{q_T-q_0}{T}\left[T-\frac{1-e^{-\lambda T}}{\lambda}\right].$$

The horizon is a judgmental 4.25 years, favoring a later qualifying snapshot while allowing snapshot timing uncertainty in the predictive spreads. The assumptions below are my forecast specification, informed by the evidence above—not empirically identified parameter estimates (baseline input).

The central scenario initially grows by 4.56 percentage points annually, close to the recent fitted slope of 4.88 points. The slow case requires a slowdown; the fast case requires acceleration. Persistent pages and the article plateau justify substantial slow-case weight. Integrated editing, cheap page creation, and improved mixed-authorship detection justify retaining a meaningful fast-growth tail (historical input; publishing mechanism; measurement mechanism).

For each structural scenario, I use a bounded Beta distribution matching its endpoint mean and predictive standard deviation. Those spreads cover adoption, renewal, crawl mix, detector redesign, and snapshot timing. They deliberately preserve the disagreement between reasonable modeling approaches rather than narrowing around one fitted curve. The remaining tail reserve protects against structural changes those models omit.

The resulting distribution has a median of 30%, a mean of 32.8%, a central 90% interval of 11%–67%, and a 3% probability of a result at or below 10% (forecast specification above). I assign 70% probability to fine-precision reporting and 30% to whole-percentage-only reporting. These are reporting assumptions, not permission to ignore an available unrounded result. Rounding is applied before bucket allocation, giving integer outcomes the appropriate extra mass. The supplied code returns the complete normalized distribution without rounding its probabilities.

What's non-obvious

A plateau in new-content adoption does not imply a plateau in the web stock. In the central model, holding incoming-content positivity fixed at 40%, rather than increasing it, still raises the all-page share to about 24% as old material is displaced or rewritten. That is my calculation from the baseline and stock model. Persistent URLs are not a protected reservoir of unchanged human prose: AI can substantially rewrite a page without changing its address.

Higher headline estimates often concern a selected denominator. A September 30, 2026 preprint reports 31.1% AI-labeled tokens in August’s FineWeb-filtered text—not raw English-page prevalence. Its July filtering audit used 10,000 documents from 32 WARC files and retained AI-labeled documents 2.3 times as often as human-labeled ones; the audit also precedes English filtering, so its raw counts are not a direct English-page replication. This prevents rebasing the forecast to that headline while still showing how extraction and filtering can move the measured share (study and Appendix A). (arxiv.org)

Uncertainties

The largest missing measurement is the future weight of unchanged older text. URL survival, crawler discovery, and publication dates do not identify that quantity. A representative longitudinal panel measuring substantive text changes would sharpen the renewal hazard much more than another article-only estimate (accessibility study; publication-date limitations).

Detector agreement is not ground truth. Pew’s paired check covered 62,370 pages from seven crawls, with 96% agreement and a kappa of 0.61. Its early positive readings likely include false positives. No verified bridge establishes how a successor detector would classify the same representative web sample against independently known substantial editing. That gap prevents a fixed upward adjustment for improved detection (Pew validation; successor-model evaluation).

The evidence cutoff is the supplied October 5, 2026 date. The current positive share is not a hard floor: this is a snapshot proportion, and valid reclassification, page loss, or crawl-composition changes can lower it. The distribution is conditional on a qualifying numerical result under the supplied resolution rules; failure to obtain a qualifying study by December 31, 2032 means annulment, not zero. No prediction-market prices or public forecast aggregates enter the estimate (forecast interpretation and specification).