AI content stopped growing in 2024 and nobody noticed

Graphite analysed 65,000 English URLs from Common Crawl covering January 2020 to May 2025. AI-written articles overtook human-written articles in November 2024, and then stopped growing. The split has sat near 50/50 ever since. Axios ran the update in May 2026 under the headline that AI has not overtaken human writers.

The flood narrative is wrong, and it has been wrong for over a year. That matters because a lot of content strategy in the last eighteen months was built on the assumption that volume was about to become infinite.

Why it plateaued

Two reasonable explanations, and both are useful.

The obvious one is saturation of the use case. Everyone who was going to switch to AI-assisted production had switched by late 2024. There was no third wave, because there was nobody left in the second.

The more interesting one is that the economics stopped working. Publishing at volume only pays if the pages earn. The evidence on unedited output says they earn for about ten weeks. Once operators worked that out, the marginal page stopped being worth the hosting, and production settled at the level where it makes sense.

Either way, the practical implication is the same: you are not competing against an infinite supply of content. You are competing against roughly the same amount of content as before, produced differently.

The decision this changes

If you believed the flood narrative, you had two logical responses. Publish more to keep share of voice, or give up on content entirely and move budget to paid.

Both are wrong under the actual data, and the first is expensive.

Publishing more to compete on volume against a supply that is not growing is a straightforward waste. Worse, it pushes you toward the production method with the ten-week half-life, so you spend more to build an asset base that evaporates.

The correct response is boring. The bar was never volume. It is the quality of the median page in your specific niche, and that bar has barely moved, because the same amount of content is being produced by people who mostly are not putting anything original into it.

What the competitive picture actually looks like

Stack the findings together and the shape is clear.

Volume is flat at roughly 50/50 human and AI. Unedited AI pages hold about 3 percent top-100 presence by month three. AI Overviews cite mixed human and AI pages 87.8 percent of the time, with pure human at 8.6 percent and pure AI at 3.6 percent. Google's documentation names the numbered-tips article as the example of what not to publish.

So: same volume, most of it commodity, citation going to pages with substance regardless of how they were produced.

That is not a hostile market. That is an unusually soft one, for anyone willing to fund actual research.

Where the advantage sits now

Three things, in order of how hard they are to copy.

Proprietary data. Nobody else has your numbers. An annual survey of your own market, or an aggregated anonymised cut of your own operational data, produces a page that cannot be replicated and gets cited for years. This is the single highest-return content investment available to a business with real customer volume, and almost nobody does it because it requires legal sign-off and a quarter of lead time.

First-hand testing. You ran the platform, the process, the product. Here is what broke. Cheaper than a survey, faster to produce, and still impossible to generate.

Practitioner specificity. The actual menu path, the actual field name, the actual threshold, the actual thing that goes wrong in month three. This costs nothing except having someone who has done the work write the piece instead of someone who has read about it.

Everything else is available to everyone, and half of it is already published.

What to do with the budget

Cut planned volume by half and put the savings into research, not into more editing on the same plan. Editing a commodity piece produces a well-edited commodity piece.

Set a floor: every published page contains at least one fact that required your involvement to know. Check it at brief stage, because checking at edit stage means the money is already spent.

Then extend the review window. Judge a content programme at six and twelve months, not at four weeks. Month one performance is the exact horizon on which the failing approach looks like it is working.

The caveat on the data

Common Crawl is a sample, not a census, and classifiers for AI-written text are imperfect in both directions. The 50/50 figure should be read as a shape, not a precise measurement.

The shape is what matters. Something that was rising steeply stopped rising in late 2024 and has not resumed. That is solid enough to plan against, and it is the opposite of what almost every content prediction from 2024 said would happen.

The flood never arrived. The opening it left is still there for anyone who publishes something worth reading.