Eight statistics that appear in most 2026 writing about content volume do not survive a check against their primary source. Two are misquotes of studies that say something different. Three have no traceable source at all. One is a 2022 forecast being cited as a measurement. Two are superseded by newer data from the same company that published the original. Below: each one, why it fails, and what to cite instead.
How I checked
Every claim below was traced back from the article quoting it to the study behind it. Where a study exists, I read the methodology for sample size, collection date, and instrument. Where no study exists, I say so. Checked September 2026.
Which content volume statistics failed verification?
1. “Google does not penalise AI content. Ahrefs found a correlation of 0.011 between AI content and ranking position.”
Verdict
Superseded by the same company’s newer data.
Ahrefs published that figure in 2025 from 600,000 pages. In July 2026 it ran the question again on a fresh sample of 331,000 pages and found the opposite direction: AI use correlated with lower ranking positions, lower indexation rates, and two to three times fewer impressions. Ryan Law’s own reading of the 2026 data is that quality tends to fall as AI use rises. The 2025 number was never fabricated. It has been overtaken, and quoting it now misstates where the evidence sits.
Cite insteadIndexation rate fell from 49.28% for low-AI pages to 40.35% for very-high-AI pages. Ahrefs, data collected June 2026.
2. “86.5% of top-ranking pages are AI-generated.”
Verdict
Misquote. The study measured something else.
The 2025 Ahrefs figure counted pages in the top 20 containing any detected AI text, which includes a single AI-assisted paragraph in an otherwise human article. It was never a claim that those pages were written by a model. The 2026 data points the other way: 54.7% of pages in positions one to three contain under 20% AI content, and pages under 50% AI account for 82.2% of top-three rankings.
3. “Over 50% of the internet is AI slop.”
Verdict
A headline three steps removed from the finding.
It comes from Graphite’s earlier single-detector study. Graphite’s May 2026 update averages three detectors and puts Q1 2026 at 49.9% of sampled articles classified as primarily AI-generated, against 50.1% human. The sample is 55,400 English-language Common Crawl articles of at least 100 words, published between January 2020 and March 2026. It measures newly published articles rather than the internet, and the three-detector method returns a figure 3.3 points lower than the one the headline was built on.
Cite instead49.9% of sampled newly published English-language articles were classified as primarily AI-generated in Q1 2026. Graphite, published May 2026.
4. “90% of online content will be AI-generated by 2026.”
Verdict
A forecast from 2022 being cited as a measurement.
2026 arrived, so the prediction can be tested. The measured figure for Q1 2026 is 49.9% of sampled new articles. A forecast that missed by roughly 40 points belongs in a piece about forecasting, not in a statistics list.
5. “19.10% of top Google results contain AI-generated content.”
Verdict
No primary source exists.
The number traces through several posts, none of which names a study, a sample, or a detector. It also conflicts with both organisations that have measured this directly. Graphite found 14% of articles ranking in Google were AI-generated, and 7% at position one.
6. “Deindexed sites collectively lost over 20 million monthly visits.”
Verdict
No primary source exists.
No named sites, no sample, no date range, no traffic tool. Documented cases exist, and they are stronger. Glenn Gabe’s post-mortem on the August 2026 spam update records one site losing rankings for over 200,000 queries and another for over 14,000, with the methodology visible.
Cite insteadCase-level query losses from the August 2026 spam update. Glenn Gabe, GSQi, August 2026.
7. “The March 2026 core update caused 50 to 80% traffic drops.”
Verdict
Assertion presented as measurement.
The same applies to the widely repeated “February 2026 caused 40 to 60% drops.” Google confirmed both updates. The percentage ranges arrive with no dataset, sample size, or method attached. What is documented and checkable: the March 2026 spam update completed in under 20 hours, the fastest confirmed spam rollout on record, and the August 2026 spam update ran from 18 to 21 August.
8. “45% of SaaS companies use programmatic SEO, at $5 to $50 per page, producing 300 to 500% growth in six months.”
Verdict
Sources cited as “case study range” and “industry consensus.”
Those are not sources. All three figures come from one aggregator page. If a number about the economics of programmatic content matters to your plan, it has to come from a company that published its own numbers, and almost nobody has.
Why do these numbers keep circulating?
Statistics pages cite statistics pages. Each hop keeps the number and drops the sample size, the collection date, and the instrument. By the fourth hop the figure has come loose from the study that produced it, and it gets quoted with more confidence than the researchers used.
This is the failure mode that volume creates, applied to evidence rather than to content. Restating a number costs nothing. Checking one costs an hour. When the incentive is to publish, the check is the first thing cut, and the same dynamic is visible in how content systems get built.
How do you check a statistic before you cite it?
- Find the study, not the article about the study. If you cannot reach a study, you do not have a statistic.
- Read the methodology. Sample size, collection date, and the instrument used. A detector-based finding is a probability estimate, not a census.
- Check whether the same organisation has published on the question more recently. Two of the eight failures above are stale rather than wrong.
- Record the collection date separately from the publication date. An article published in 2026 can be reporting data gathered in early 2025, which matters when the thing being measured moves every quarter.
That fourth step catches more errors than the other three combined. It is also the step that shows up in almost no marketing citation, which is part of why the provenance question is getting harder rather than easier.
FAQ
Does Google penalize AI-generated content?
The often-quoted Ahrefs 2025 finding of a 0.011 correlation is out of date. In July 2026 Ahrefs re-ran the question on 331,000 pages and found AI use correlated with lower ranking positions, lower indexation and two to three times fewer impressions. Indexation fell from 49.28% for low-AI pages to 40.35% for very-high-AI pages.
Is more than half of the internet AI-generated now?
Not the internet. Graphite’s May 2026 update, which averages three detectors, puts Q1 2026 at 49.9% of sampled newly published English articles classified as primarily AI-generated. The sample is 55,400 Common Crawl articles, so it measures new articles, not the whole web.
Will 90% of online content be AI-generated by 2026?
That figure is a 2022 forecast, not a measurement. The measured figure for Q1 2026 is 49.9% of sampled new articles, so the forecast missed by roughly 40 points.
How do I check a marketing statistic before citing it?
Find the study, not the article about it. Read the methodology for sample size, collection date and instrument. Check whether the same organisation has published newer data. Record the collection date separately from the publication date, because that step catches more errors than the other three combined.
Written by Tanja Gavrilović. All claims traced to primary sources in September 2026. Human-researched, human-edited, AI-assisted in drafting.