The SLOP equation: how much unedited AI text gets published every day

How honest is this number?

A figure this large and this round should make you suspicious. This is an estimate, not a measurement. S, L, and O rest on published figures; P — the share that goes out with no human review — has a slider that lets you use your own factor because no one publishes it. At the conservative 10% default the equation estimates 814 million unedited AI-generated words a day. At 50% it spikes to four billion. Treat it as an estimate, as you would the Drake equation that estimates extraterrestrial civilizations.

The factor nobody can measure

P — the fraction of AI text that ships without a human ever reading it — is a difficult one to measure. One survey by HubSpot found marketers published 7% of their content without review, albeit self-reported. Move it from 10% to 30% and the daily total triples, so the final estimate is very sensitive to this factor.

That is the point of the exercise. It is the Drake-equation move: you build the formula less to get an exact number than to find out which number is missing or most sensitive to the estimate. Here the variable P is also the thesis: slop is not merely a generative-AI problem; it is a human-review problem. The tells that give AI writing away — the buzzwords, the hedging — are what’s left when no one edits the writing. Only a human who cares about craft and quality can decide when a draft is complete and edited for publication. P measures how often nobody did that.

Where the numbers come from

S and L: the open web

The web’s daily output is best estimated as a range. Wix alone reports on the order of 7.5 million new blog posts a day, each its own page; global domain registrations add several hundred thousand fresh homepages; and platform pages — product listings, forum replies, localized news — run into the millions more. We use ten million indexable pages a day as a deliberately round, defensible S. For length, we weight three page types — long-form posts near 1,380 words (Orbit Media’s survey and Digital Applied’s trackers put the average at 1,333–1,427), thin dynamic pages near 250 (Yoast and WordPress VIP treat 200–300 as the floor), and new homepages near 400 — into a blended 1,100. The AI share, O, is the one hard number of the three: Ahrefs found in 2025 that 74% of newly created web pages contain AI-generated content.

Books and publishing

The pattern shows up off the open web too. By late 2025 Amazon was seeing roughly 300,000 e-books released a month, up from about 100,000 before ChatGPT — and titles that read as AI-authored draw fewer and worse reviews, according to the Economist. The market is absorbing the volume and appears to be penalizing the slop.

Academic writing

Academia is the best-measured corner. A PNAS analysis found AI-influenced language in 57% of arXiv papers in 2025, up from 12% in 2023 — a nearly fivefold jump in two years. The interesting part isn’t that researchers use the tools; it’s how fast unreviewed AI phrasing spread once they did.

Business and professional writing

This is our daily reality. However, LinkedIn posts, cold emails, internal memos, and marketing copy are the writing most likely to be generated and least likely to be tracked. Public per-day figures barely exist here, which is why it’s best to estimate P: the professional web is where “no one reviewed it” is most plausible. It’s the section we most want to strengthen — and one of the reasons we built UnSlopIt.ai.

Reader trust

The cost lands on the reader, and here the numbers are firmer. Gartner found 50% of consumers prefer brands that avoid AI in consumer-facing content, and 68% often wonder whether what they’re seeing is real. Klaviyo and Datalily put trust-loss at 31% when content is visibly AI-generated. Skyword found that when AI content conflicts with a brand’s own messaging, only 29% still trust the brand and 54% go hunting for third-party validation. And Cashew found 87% of people already assume brand content is at least partly AI-made, while just 13% are confident they can tell. The slop is being published faster than trust can survive it.

Methodology

Primary sources where they exist, and the more conservative figure when a range is reported. Every sourced term links to its origin; assumptions are labeled as assumptions, and P is left as a slider rather than a claim. Found an error or a better number? The equation is versioned (below) and corrections are made in the open — reach us through the About page.

Our own data

One number we will eventually own. UnSlopIt logs anonymous pattern counts on every edit — which tells appear, per thousand words, never the text itself. As that dataset grows it gives us a measured “tells per 1,000 words,” which extends the equation from words to tells: not just how much unedited AI text ships, but how much of it a reader can actually spot. That’s the next version.

Put the equation on your page

The calculator is embeddable — copy the snippet below. And if you want to do something about your own slop, start with what the tells are or how to cut them.

<iframe src="https://unslopit.ai/slop-equation/embed" width="100%" height="720" style="border:1px solid #E4DDCC;border-radius:2px" title="The SLOP Equation — UnSlopIt.ai" loading="lazy"></iframe>

The SLOP Equation, UnSlopIt.ai, 2026. https://unslopit.ai/slop-equation