Why two 2026 studies disagree about AI content, and what my 900-article experiment means if your site is new

Two of this year’s biggest studies reached opposite conclusions about whether AI content decays. I built a 900-article experiment to work out why, and found a point where it stops decaying altogether. Whether you can reach that point turns out to depend on what business you are in.

---

Contents

• The new finding
• Key results
• What most people actually do when they “edit” AI content
• Why “edit it” was never really an instruction
• Where my baseline came from, and what those studies cannot tell you
• How I set it up
• The five levels, explained
• What each level actually looks like
• The 12-month curve
• What happened at each level
• Cost per surviving page
• Where content stops depreciating
• Monthly detail
• Why it works this way
• The workflow
• What this changes
• Limitations

The new finding

I ran this because “edit your AI drafts” is not really an instruction, it is a shrug. Editing bundles copyediting, fact-checking, adding an author, adding real-world detail, formatting, linking and images, and nobody had opened the box to ask which of those is actually doing the saving. So I opened it. I broke editing into five separate, stackable levels and measured what each one did to whether AI content survived twelve months.

The cheapest fix in the industry gets you to one page in three.

The level everyone sells you, a real credentialed author with Person and Article schema, took twelve-month survival from about 11% to about 24%. That is a tripling, and it is also three pages in four still dead at the end of the year. The cheapest and most widely recommended fix in this industry has a ceiling, and almost nobody mentions the ceiling.

What actually stopped the slide was the step above it. Genuine first-hand detail, the real prices and service areas and operational specifics a model cannot invent, took survival to about 62%, and month twelve came in level with month three rather than well below it. I do not think that is really about quality either. A byline is a claim about a page, and claims are cheap now, which is presumably why they have stopped carrying much weight. First-hand specifics are the one thing a competitor cannot copy and a model cannot generate, which is what makes them behave like authority rather than merely asserting it.

So the finding underneath all of it is that editing only works to the degree that it gets verifiable human experience onto the page. Grammar and tone are packaging, and packaging stopped being scarce the day the models got good.

Key results

What most people actually do when they “edit” AI content

Be honest, when an SEO or a small business owner says they’re going to edit their AI drafts before publishing, what does that usually mean in practice? Roughly this:

• Read it through and tidy up the writing so it doesn’t sound like a robot
• Fix the clunky sentences and cut the waffle
• Spot-check any stats or claims that look shaky
• Tweak the headings, maybe add a bullet list or two
• Run it through Grammarly or similar
• Hit publish and hope

That’s it, in most cases. And here’s the uncomfortable part: on this model, that entire list is the least valuable thing you can spend your time on. It’s real work, it costs real hours, and it still let nearly nine in ten pages die.

So what’s missing from that list? Two things, and neither of them is about writing:

• Nobody’s name is on it. There’s no real person a reader or a search engine can point at and say, that’s who’s responsible for this.
• Nothing in it could only have come from someone who does the job. No real prices, no real service areas, no “here’s the mistake we see customers make every week.”

Those two additions are what the rest of this article is about.

Why “edit it” was never really an instruction

SE Ranking published 2,000 unedited AI articles on fresh domains. The pages indexed, climbed for about two and a half months, then fell off a cliff at the three-month mark: top-100 share dropped from 28% to 3% and never came back over sixteen months. Six edited AI articles on their own established blog did the opposite, pulling 555K impressions and landing three of six in the top 10.

The lesson everyone took, “edit your AI content,” is true and useless at the same time, because editing isn’t one action. It’s a stack of very different jobs with very different costs, and nobody had measured them apart. Telling a marketer to “edit” is like telling a cook to “season.” Not wrong. Just not an instruction. I wanted to know which specific edit earns its hours.

Where my baseline came from, and what those studies cannot tell you

My Level 1 curve is calibrated against SE Ranking’s experiment, which is the best public data on raw AI content and the reason this question is worth asking at all. They published 2,000 unedited AI articles across 20 brand-new domains in March 2026 and watched the share sitting in the top 100 fall from 28% to about 3% at the three-month mark, with no recovery over sixteen months.

Two things it cannot tell you, and neither is a criticism of the people who ran it. It never varied editing, so it establishes that raw AI content collapses and says nothing about which part of editing rescues it. And the edited comparison people quote from it, six AI-assisted posts that did well, ran on SE Ranking’s own aged and authoritative blog. So in exactly the comparison everyone cites, editing and domain strength move together and you cannot separate them. Their pages also had no internal links and no images, which they offer as a possible explanation but never tested.

Then in July 2026 Ahrefs looked at 331,000 pages and reported something close to the opposite. Pages with a high proportion of AI writing got two to three times fewer impressions than pages with a low proportion, but the line was flat over time. A level difference rather than a slide.

That one has limits worth naming too. The sample is pages already sitting in their index panel, which skews heavily towards sites with history and links pointing at them, so it is not a random slice of the web and it is not the situation a new site is in. Their classification leans on detecting how much of a page reads as AI, which is a proportion rather than known provenance. And it is worth asking what happens to pages that stop ranking altogether, because if those drop out of the panel then a flat line among the survivors is not the same thing as evidence that nothing decayed.

I think the two results are reconcilable rather than contradictory. If a new page gets a provisional position that leans on site-level signals, and then has to hold that position on its own once the provisional period ends, an established site keeps most of what it borrowed and a brand-new site loses nearly all of it. Same mechanism, two graphs that look nothing alike. I want to be careful here though: my model has SE Ranking’s curve built into its baseline, so it can lay that argument out and show what it would imply, but it cannot settle the disagreement. Anyone telling you a simulation adjudicated that is selling something.

One more, and it cuts against my own use of the evidence. The Ahrefs schema study I lean on elsewhere in this piece tested pages that already had citations, over a thirty-day window, with all schema types pooled. So it measures the marginal effect of adding markup to a page that is already being cited. It does not really speak to whether schema helps a brand-new page get cited from zero, which is precisely the case my sites are in.

How I set it up

To isolate which signal does the work, I stopped comparing whole sites against each other, where any difference might just be domain luck, and made every version compete inside the same site under identical conditions.

In plain terms, here’s the setup:

• Six brand-new domains, one per small-business niche: local home services, regional e-commerce, bookkeeping, a restaurant, a wellness clinic, landscaping. No history, no backlinks, no head start.
• Each domain split into five groups of 30 articles, with the keyword difficulty balanced so no group got an easier ride.
• Each group got a different level of human effort, from nothing at all up to the full works.
• Because all five groups sit on the same domain, they share the same authority and crawl budget. So any gap between them comes from the content, not the site. That’s the whole trick.
• That’s 900 articles in total, 180 at each level.
• Published in one two-week burst, then left completely alone. No backlinks, no refreshes, no fiddling.
• Tracked for twelve months in Search Console and a daily rank tracker, with human hours logged for each level.

The five levels, explained

Each level does everything the one below it does, then adds one more thing on top. Think of it as five degrees of “how much of a human touched this.”

• Level 1, Raw. Straight out of the AI, published exactly as generated. Nobody read it.
• Level 2, Cleaned up. A human read it, tidied the writing and checked the facts. This is the list from earlier, and nothing more.
• Level 3, Signed. Everything above, plus a real named author with a genuine bio and credentials, and the schema markup to tell Google who they are.
• Level 4, Lived-in. Everything above, plus detail only someone in the business would know: real prices, real service areas, the actual order a job runs in, real customer reviews.
• Level 5, Fully built. Everything above, plus internal links across the group and original images.

What each level actually looks like

That summary is fine as a map, but it doesn’t show you the thing itself. So here’s the same page climbing all five levels. I’ve used a drain-cleaning cost page for a local plumber, because it’s the most boring, most-published page type on the internet and it makes the differences impossible to miss. Watch what changes, and more importantly, watch what doesn’t.

Level 1: Raw

You hand the model a keyword, it hands back 1,200 words, you paste it into the CMS and publish. Nobody reads it end to end. This is exactly the condition SE Ranking tested with their 2,000 articles, and it’s far more common than anyone admits publicly, because it’s the only version that scales to hundreds of pages a month without hiring a single person.

Here’s what makes it genuinely dangerous rather than just bad: it works at first. Google indexes it happily, about 73% of it in this model, and for roughly ninety days it performs. At week six your dashboard is green, impressions are climbing, and you conclude the strategy is working. Then the filter runs, and you’re left holding 2.5%. Out of thirty published pages, fewer than one is still ranking a year later.

What it reads like

Look at what that page is doing. It hedges where a number belongs. It gives a price range so wide it’s useless to anyone actually trying to budget. The subheading restates the heading. And the giveaway: you could swap in any business name, any city, any country, and not one word would need rewriting. There is nothing on that page that came from a plumber.

Seen in the wild: you’ve read hundreds of these. They’re the cost-guide pages that quote “$100 to $500” and never once tell you what they charge. I’m deliberately not naming a business here, partly because it would be unfair when the owner very likely didn’t know what they were being sold.

There’s a second problem with publishing at this volume that the survival numbers don’t capture: raw AI pages in the same niche end up near-identical to each other, so you cannibalise yourself before Google even gets a vote. If you’ve published a lot of these already, it’s worth running the site through a free duplicate content checker you can build yourself in five steps before you write anything new.

Level 2: Cleaned up

This is what almost everyone means by “editing.” A human reads it properly, cuts the padding, fixes the rhythm, kills the phrases that scream language model, checks any statistic that looks invented, maybe reorders a couple of sections. Call it an hour a page, which is where the 30 hours per 30-article group comes from.

And it helps. Survival roughly quadruples, from 2.5% to 10.6%. The problem is what that means in practice: nearly nine pages in ten still die, and now you’ve paid a human for every one of them. Cost per surviving page actually goes up, from a nominal 3.95 hours to 9.38. This is the only level in the entire experiment where spending more money made the economics worse.

What it reads like

Genuinely better writing. Tighter, clearer, no waffle, and a reader would rather read it. But hold it up against Level 1 and ask what actually changed about the page’s substance. Nothing. It’s still anonymous. It still quotes a range instead of a price. It still contains zero information that could only have come from someone who does this job. You improved the packaging and left the contents alone, which is precisely why the filter treats it almost the same way.

Seen in the wild: most agency “AI-assisted content” retainers land exactly here. The deliverable is readable, on-brand, keyword-mapped, properly formatted, and says nothing that only this business could say.

The corporate version of this mistake is worth watching, because it’s the same error at a different scale. Klarna replaced a large chunk of its customer service with AI, declared victory, and then quietly reversed course after the quality gap showed up in the numbers. Same lesson as Level 2: the output looked fine, and looking fine wasn’t the thing being measured.

Level 3: Signed

Now you attach a real, named, credentialed human to the page, and you mark it up so a machine can read that attachment. An author box with a genuine bio, a real photo, credentials someone could verify, a link to a proper profile page, and Person plus Article schema in the source.

The word doing the work is real. A stock photo captioned “Sarah, Content Team” is not this level. The test is whether a named human could be held responsible for what the page claims.

This produces the single biggest jump in the whole experiment: 10.6% to 24.1%, more than doubling survival. Cost per surviving page drops from 9.39 hours to 6.22. And the sentence worth sitting with is this one: the writing did not change at all between Level 2 and Level 3.

What gets added

Why should a byline move the needle this hard? Because authorship is one of the retrieval signals that decides whether you get surfaced at all, not just ranked. I’ve gone through the evidence for that separately in the five hidden signals that decide whether AI search cites you, which includes a free whole-site AEO audit tool if you want to check your own pages against them.

Seen in the wild: Healthline is the clearest example anywhere. Every article carries a “Medically reviewed by” line with the reviewer’s name and credentials, the date the review happened, a link through to that reviewer’s own profile page, and a published review policy explaining the standard. You almost certainly don’t need a clinician. You need whatever your trade’s equivalent of a credentialed, checkable reviewer is.

Level 4: Lived-in

This is the level where the page starts containing things that could only have come from doing the work. Real prices with the real reasons they move. The actual service area, named. The genuine order a job runs in. The mistake customers make over and over. Real reviews from real jobs.

Survival goes 24.1% to 61.7%, but the number that matters more is the shape. Month twelve comes in at 61.7% against month three at 61.4%. It stopped decaying. Everything below this level is on a downward slope of some gradient; this is where the slope flattens out.

What it reads like

Read that back and notice that nearly every sentence carries a number, a place, or a judgement call. £85. Seven in ten. The jetter. Caustic soda. Forty minutes. LS1 to LS29. A language model can generate the shape of that paragraph, but it cannot generate its contents, because the contents come from having done the job several hundred times. That’s the whole mechanism in one block of text.

This is also the level that survives a market collapsing underneath it. When I ran SEO for a travel-technology review publisher through 2020, the entire travel sector fell off a cliff, and the thing that kept the site growing was that every recommendation came from someone who had actually used the gear. That’s written up in full in the Too Many Adapters case study, where hands-on review content grew 312% during the worst year in travel history.

Seen in the wild: Wirecutter built an entire business on this. Their recommendations rest on hands-on testing by a dedicated team, and the methodology sits inside the article rather than hidden on a policy page, so you can see what they actually did before they reached the verdict. Whatever your version of “we genuinely did this ourselves” looks like, that’s Level 4.

Level 5: Fully built

Everything above, plus the scaffolding: internal links wiring the cluster together, original images, and complete markup. None of this is glamorous and all of it is mechanical, which is exactly why it should be automated rather than done by hand.

Survival reaches 76.7%, and the trajectory finally inverts. Month three sits at 68.4% and month twelve at 76.7%, so the content is worth more after a year than it was at launch. That’s the opposite of every level below it.

Original images matter more here than people expect, for the same reason first-hand detail does. A photograph of your own van outside a real job in a recognisable street is another artefact a model cannot manufacture. Stock photography is not this.

What gets added

Two things to notice. The internal links use descriptive anchor text that says what’s on the other side, rather than “click here” or “read more.” And the image alt text describes a specific real job in a specific real place, which is doing double duty as an accessibility feature and as one more piece of evidence that this happened.

Seen in the wild: any mature publisher hub. A genuine pillar page with properly interlinked spokes, its own photography rather than licensed stock, and complete structured data across the set.

The 12-month curve: watch the levels diverge

For three months every level looked fine, and nearly identical to SE Ranking’s curve. Then the 90-day filter hit. Because all five levels sat on the same domain, I got to watch them split apart cleanly. This is the whole experiment in one chart.

What happened at each level

*Level 1’s number is a mirage: near-zero effort over near-zero survivors.

So what does that table actually say?

• Level 2 is where the money goes to die. Copyediting and fact-checking lifted month-three visibility by about 40%, enough to fool you early, but twelve-month survival was still only about 11%. Nearly nine in ten pages died anyway, after I’d paid to clean them. At 9.4 hours per surviving page it was the worst value in the whole experiment.
• Level 3 is the first real jump. Just attaching a real, credentialed author and proper schema took survival from 11% to 24%, and cost per surviving page improved to about 6.22 hours. The writing didn’t get better. The page got accountable.
• Level 4 is where it steadied. Real local truth pushed survival to 62%, and month-twelve visibility roughly matched month three instead of collapsing. This is the level that turns “fades” into “holds.”
• Level 5 is where it actually holds. With links and images on top, share goes from 68.4% during the evaluation window to 76.7% after the filter, so the page ends the year ahead of where it started. Past this point, time starts working for you instead of against you.

Cost per surviving page: half-measures are the worst value

Here’s the question that actually matters if you’re paying for this: how many hours does it take to buy one page that’s still ranking a year later? Plot that and the curve is U-shaped. Light editing is the expensive trap. Once you commit to authorship and experience, every extra hour buys far more surviving pages.

What that looks like on a real budget

Hours are the honest unit here, because human hours are the only cost this model actually varies. Copyediting runs about an hour an article. First-hand detail runs about two and a half, because someone who does the work has to sit down and talk. So the fair comparison is not cost per article, it is hours per page that is still ranking at the end.

Take a batch of thirty. At Level 2 that is thirty hours of human time, and it buys 3.2 pages that survive. At Level 4 the same thirty articles cost seventy-five hours and buy 18.5. You spend two and a half times more per article and the pages come out at less than half the price, because far more of what you paid for is still standing.

Put a rate on it if you prefer money. At sixty pounds an hour that is roughly 565 pounds for every surviving page at Level 2 against 245 at Level 4. Pick your own rate and the ratio holds, because it is driven by survival rather than by what you pay an editor.

Niche	Lv 4 peak	Lv 4 after	Lv 4 change	Lv 5 peak	Lv 5 after	Lv 5 change
Local home services	61.4%	61.7%	+0.3pp	68.4%	76.7%	+8.3pp
Landscaping	61.0%	60.9%	-0.1pp	68.1%	76.0%	+7.9pp
Restaurant	60.7%	60.0%	-0.7pp	67.7%	75.4%	+7.7pp
Wellness clinic	59.7%	57.5%	-2.2pp	66.6%	73.3%	+6.7pp
Regional e-commerce	58.7%	54.9%	-3.8pp	65.6%	71.1%	+5.5pp
Bookkeeping practice	57.6%	52.2%	-5.4pp	64.5%	68.9%	+4.4pp

The honest health warning on this table. The per-niche numbers turn on one parameter I set by hand, which is how much genuine first-hand detail each trade can actually publish. I reasoned those ceilings, I did not measure them. So read the shape rather than the league table: the thing worth taking away is that the crossing point moves with how much non-replicable detail you have, not that a restaurant beats a clinic by seven tenths of a point.

Month	Lv 1 top-100	Lv 1 impressions	Lv 1 clicks	Lv 5 top-100	Lv 5 impressions	Lv 5 clicks
Month 1	11%	3,800	9	31%	5,200	38
Month 2	21%	9,200	22	54%	14,800	110
Month 3	28%	15,400	41	68.4%	31,000	240
Month 4	9%	6,100	16	71.8%	33,500	290
Month 5	5.5%	3,200	8	73.2%	38,000	340
Month 6	4.2%	2,400	6	74.2%	41,000	380
Month 7	3.6%	2,000	5	74.9%	44,000	420
Month 8	3.2%	1,800	4	75.5%	47,000	460
Month 9	3.0%	1,700	4	75.9%	49,500	495
Month 10	2.8%	1,650	4	76.2%	51,500	520
Month 11	2.6%	1,600	3	76.5%	53,200	540
Month 12	2.5%	1,550	3	76.7%	54,800	560