What happened to 2,000 unedited AI articles over 16 months

SE Ranking ran the experiment everyone argues about instead of running. Twenty brand new domains, 2,000 AI-generated articles, no human editing, no links, 16 months of observation. The results are the clearest picture available of what unedited AI content actually does over time, and the shape of the curve is the finding.

About 71 percent of pages indexed in the first month. The sites accumulated more than 122,000 impressions. By month three, only 3 percent of pages remained in the top 100. The August 2025 spam update briefly lifted that to around 20 percent, which did not hold. Almost all impressions landed in the first two and a half months.

The shape matters more than the numbers

That is not a penalty curve. A penalty is a cliff. This is a slow bleed, and the difference changes how you should read your own dashboards.

Early indexing gave a strong false positive. Pages got in, got impressions, and anyone reporting on month one would have called the programme a success. The failure only became visible in month three, by which point a real content operation would have published another few hundred pages on the same basis.

If you are running an AI-assisted content programme and reviewing performance monthly, you are reviewing it on the exact horizon where it looks good.

What Google's policy actually says

Worth being precise here, because the policy is narrower than people assume and broader than they hope.

Google's spam policies document, last updated in May 2026, defines scaled content abuse as many pages generated for the primary purpose of manipulating search rankings and not helping users, explicitly including generating numerous pages through generative AI tools.

The key phrase is the purpose test, not the tool test. Human-written thin content is covered identically. There is no clause anywhere saying AI-written content is prohibited. There is a clause saying content produced primarily to rank, at volume, without helping anybody, is prohibited regardless of who or what typed it.

The two siblings in the same document catch related patterns: site reputation abuse, where third-party content parasites a host domain's signals, and expired domain abuse.

So the experiment did not fail because a detector caught it. It failed because 2,000 pages of unedited output contained nothing worth ranking for longer than ten weeks.

Why the decay happens

Three mechanisms, and they compound.

Indexing is cheap, ranking is competitive. Getting into the index is a low bar and always has been. Holding a position requires beating pages that already exist on relevance and usefulness, which is measured over time through actual user behaviour.

The content has no unique fact in it. A model writing without sources produces a synthesis of what is already published. There is no reason for a system to prefer that synthesis to the original sources it synthesised.

Nothing accumulates. Real content programmes get better over time because pages earn mentions, get cited, get updated, and build topical depth. Unedited output earns nothing, so month twelve looks exactly like month one, except everything published in month one has aged out.

The threshold that separates this from working

The useful question is not should we use AI but how much human input flips the curve. The honest answer from the available evidence is that the intervention has to add something the model could not produce, and volume of editing is not the same as substance of editing.

Rewriting sentences for tone does not change the curve. The page still contains no fact worth retrieving.

What does change it: a number from your own data, a detail from a real project, a named threshold, a correction to something widely repeated and wrong, an opinion with an argument under it. One of those per page, genuinely present, is worth more than a full stylistic rewrite.

That is a research bar, not an editing bar. It is why the cost of good AI-assisted content is not much lower than the cost of good content, and why the promised savings mostly do not arrive.

What to do if you are already running this

If you have an AI content programme in market, audit it against the curve rather than against opinions.

1. 1.Pull every page published more than four months ago. Chart impressions by month since publication for each cohort.
2. 2.If the cohort curves peak in weeks four to ten and decline after, you have the pattern in the study.
3. 3.Sort those pages by whether they contain at least one fact that could not have been generated without your involvement. Most teams find this is under 10 percent.
4. 4.Consolidate. Take the pages with substance, merge the rest into them where topics overlap, and delete the remainder with proper redirects. A smaller library of substantial pages outperforms a large library of thin ones on every measure that matters.
5. 5.Change the review horizon to six months minimum for any new content programme. Month one performance is not evidence.

The honest counterpoint

This study used zero editing and zero links, which is a deliberately extreme setup. It tells you what pure automation does. It does not tell you that AI-assisted content fails, and the evidence elsewhere is that AI-assisted pages are cited by AI search at high rates.

The finding is narrower and more useful than the headline: automation without substance produces a ten-week asset. If your content plan assumes a page will still be earning in year two, unedited generation cannot get you there.

Publish fewer pages that contain something only you know, and the decay curve stops being your problem.