Should You Use AI Content in 2026 - Episode 145

Grumpy SEO Guy

Abstract

An SEO practitioner spends ten minutes on the question saturating every forum in the field — can machine-written content rank, will it hurt rankings, is it safe — and answers it by reading two published studies against their own data rather than by asserting a position. He opens with a disclaimer that is unusual in the genre: he does not officially know how any search engine works, most published information in the field is wrong, and everything that follows is inference from testing.

• Two hundred machine-written pages across twenty fresh domains indexed fast and then decayed.
• The study's own first conclusion is the one that survives: visibility failed for lack of authority, backlinks and external validation.
• Author credentials and stated expertise are assertions a page makes about itself, and a signal that can be faked cheaply cannot be a ranking factor.
• A larger analysis found most top-ranking pages machine-assisted — neither punished nor rewarded.
• Purely machine-written content rarely reaches the first position, which is where authority concentrates.
• Free detectors already exist, so suppression is plausible rather than proven, and the practice that follows is to use very little.

Everything below follows the episode in order, through the experiment's design, its four conclusions taken one at a time, the second study's null result, and the listener reports that are the only operator evidence in it. For anyone building a personal brand the transferable part is the method rather than the verdict: the conclusions get separated from the data that supports them, and only two of four survive.

Chapter summaries

00:43 - The question saturating the forums is whether machine-generated content can rank

The question saturating every SEO forum is whether machine-generated content can rank, will hurt rankings, or is safe to use at all.

"SEO forum, everybody is like, should I use AI content? Will AI content hurt."

— 00:43

The question arrives in three forms that are really one: can it rank, will it hurt, is it safe. A fourteen-year practitioner opens by stating plainly that he does not officially know how any search engine works, and that everything asserted afterwards is inference from testing rather than from disclosure. He also notes that most published information in the field is wrong, which is why two case studies are named rather than a consensus.

02:44 - One study bought twenty fresh domains and posted ten machine-written articles on each

The first study bought twenty fresh domains, posted ten machine-written articles on each — two hundred pages total — and tracked them for sixteen months.

"The first article is called How AI-generated content performs."

— 02:44

A large, uniform, machine-only corpus on new domains isolates what machine content does with no other advantages attached. Twenty domains, ten articles each, all machine-generated, tracked over sixteen months.

03:17 - The pages indexed fast and then decayed across the following months

The pages indexed fast and then decayed: about 71% indexed within the first 36 days, impressions and clicks grew early, but by roughly three months only 3% remained in the top 100, recovering to 20% by month sixteen.

"So here's what they said. Month one, about 71% of new AI-generated pages."

— 03:17

About seventy-one per cent indexed within the first thirty-six days, impressions and clicks grew early, and then the curve turned down. Indexation is not ranking: the pages were found and kept, and the position they briefly held was not earned by anything that lasts.

03:58 - Visibility failed for lack of authority rather than for lack of quality

The study's first conclusion — that visibility failed for lack of authority, backlinks and external validation — is treated as obvious: you need authority to rank.

"They have four conclusions. They said why SEO visibility didn't last."

— 03:58

The study's own first conclusion is the one that survives scrutiny — visibility failed because there was no authority, no backlinks and no external validation behind any of the two hundred pages. Nothing about how the text was produced is doing the work here. Authority is the requirement the pages never met, and it would have been the same verdict on two hundred pages written by hand.

04:26 - Author credentials and stated expertise are spoofable, so they cannot be ranking factors

The study blamed missing author credentials and real-world expertise, and this is disputed: authors, credentials and real-world expertise are not ranking factors because they can all be spoofed.

"They said expertise and credibility. No authors credentials or real world expertise."

— 04:26

A signal that can be faked cheaply cannot be load-bearing, because the engine would be gameable by anyone willing to type a name. The study blamed missing author credentials and real-world expertise; both are assertions a page makes about itself, and an assertion is not evidence. The disagreement is not about whether expertise matters to a reader — it is about whether a machine can verify it from the page.

05:04 - Derivativeness is half the diagnosis, and missing structure is the other half

The study's differentiation point — that the content resembled what already exists — is half-right: the real cause is missing authority, but it is also true that machine content largely reproduces what is already out there.

"Content differentiation. They said much of the content resembled what already exists without unique."

— 05:04

Derivativeness is half right: content that resembles what already exists adds nothing an index needs a second copy of. The other half is the study's own next point, conceded without argument — no internal linking, no topical organisation, no clear hierarchy. That is a structural failure rather than a writing one, and it would have sunk the same pages in any era.

05:37 - Most top-ranking pages are machine-assisted, so the content is neither punished nor rewarded

A larger analysis ran a content-detection tool over top-ranking pages and found most of them machine-assisted, concluding the search engine neither punishes nor rewards machine content.

"The next article I want to tell you about. It's called AI generated."

— 05:37

If detection of machine assistance does not predict rank in either direction, the production method is not itself the lever. Six hundred thousand pages were analysed, and most top-ranking ones were judged machine-assisted.

06:16 - Purely machine-written content rarely reaches the first position

The same analysis found that purely machine content rarely reaches the number-one position — flagged as its most interesting result.

"This is very interesting. They said purely AI content rarely reaches the number one."

— 06:16

It is flagged as the most interesting result, and it is consistent with everything above it — the top position is where authority concentrates, and pure generation is cheapest exactly where authority is scarcest.

06:45 - Practitioners report indexing trouble that eases once a human edits the page

Listeners report directly that they struggle to get machine content indexed, and that in several cases the problem eased slightly once they switched to writing themselves.

"You what I have been told. This is people that I."

— 06:45

In several cases the problem eased slightly once they switched to writing themselves — which is anecdote rather than measurement, and is offered as such. It is anecdote rather than measurement and is presented as such, but it is the only evidence in the episode that comes from operators rather than from published studies.

07:55 - Free detectors already exist, so suppression is plausible rather than proven

A plausible-suppression argument: free detectors already exist and the engine likely has better, so if it judged machine content a quality problem, the odds it scans for it and lets those sites rank less easily are, in this view, good.

"I want you to think about this for a minute. Can probably determine."

— 07:55

If free detectors already exist, the engine plausibly has better ones — and if it judged machine content low quality it could suppress it. The argument is offered as plausible rather than proven, and the practice that follows from it is stated without being defended: the agency uses very little machine content, if any. The unproven risk is priced rather than argued.

Personal Branding Lessons

The subject is machine-written content, and the method underneath it is the useful part: read a study's conclusions against its own data, and keep only the ones the data supports.

Treat the authority gap as the cause before blaming the writing

Two hundred pages on twenty fresh domains had no links, no history, no external validation and no promotion. When they failed, the readily available explanation was how the text was produced — and the study's own first conclusion says otherwise: visibility failed for lack of authority. The same two hundred pages written by hand would have produced the same curve. Before attributing a failure to quality, check whether the thing ever had the one requirement that actually moves position. 03:58

Reject any signal that can be faked cheaply

A search engine cannot be built on something anyone can type. Author credentials and stated real-world expertise are assertions a page makes about itself, which makes them spoofable and therefore useless as ranking signals — however much they matter to a human reader. The test is worth carrying beyond this subject: whenever something is proposed as the reason a page ranks, ask what it would cost to fake, and treat a cheap answer as a disqualification. 04:26

Give a machine-written page the structure and internal links it lacks

The half of the diagnosis that survives is structural rather than stylistic: no internal linking, no topical organisation, no clear hierarchy. That is conceded without argument in the episode because it is not really about machine writing at all — it is the condition of any page published into a vacuum. The fix costs nothing and does not require rewriting a word of the text. 05:04

Expect indexation rather than ranking from a new page

Seventy-one per cent of the pages indexed within thirty-six days, and impressions and clicks grew before the curve turned down. Being found and kept is not the same as being ranked, and the early growth is the part most likely to be mistaken for success. Watch the shape over months rather than the first reading, because the first reading of a new domain is almost always flattering. 03:17

Read a study's conclusions against its own data

Four conclusions were offered and two survive. The experiment isolated one variable and stripped out every other input a page would normally have, which makes it a clean test of exactly one thing — and a poor basis for conclusions about credentials, expertise or differentiation. The discipline is to check which findings the design could actually have produced, and to say plainly which ones it could not. 02:44

Edit the page by hand before asking why it will not index

The only operator evidence in the episode is a set of listener reports: people struggling to get machine content indexed, and the problem easing once a person edited the page. It is anecdote and is presented as such. It is also free to test on your own pages, which is more than can be said for most of what gets asserted about this question. 06:45

Assume detection is possible and price the risk accordingly

Free detectors already exist, so an engine plausibly has better ones — and if it judged machine content low quality it could suppress it. None of that is proven, and the argument is offered as plausible rather than established. The response is not to argue the point but to price it: the practice that follows is very little machine content, if any, which is a decision about exposure rather than about the evidence. 07:55

Questions

Each answer ends at the moment in the recording where it is given.

Does machine-generated content hurt your search rankings?

On the evidence gathered here, no — not directly. A large analysis of top-ranking pages found most of them to be machine-assisted and concluded the search engine neither punishes nor rewards content for how it was produced. The method is not the thing being scored. What decides the outcome is whether the pages have earned external validation and a reputation, which is a separate question from which tool typed the words. 05:37

Why do machine-written pages get indexed quickly and then lose visibility?

Because fast indexing and durable ranking are two different things. In the sixteen-month experiment, most new machine-written pages were indexed within about five weeks, but by roughly three months only three percent remained in the top hundred results. Without any authority behind them, the new pages tended not to hold that early position. Indexing is a door opening, not a rank being held. 03:17

Do you need backlinks and authority for machine content to rank?

The episode treats this as the obvious lesson of the failed experiment: pages on brand-new domains with no backlinks and no external validation could not compete with established sites, however they were written. Authority is what durable ranking requires, and it is exactly what a batch of freshly published pages does not have. For anyone building alone, being vouched for by others matters far more than raw output. 03:58

Are author credentials and real-world expertise actually ranking factors?

The episode disagrees that they are, on a clear test: author bylines, stated credentials and claimed expertise can all be faked cheaply, and a signal anyone can spoof cannot be what a ranking system relies on. The study is not called wrong, and credentials are granted a possible larger role in fields like finance, health and law — but the general claim is that these are not the lever people assume. 04:26

Why won't my machine-written articles get indexed?

Beyond the studies, the episode reports that multiple operators have said the same thing directly: they struggled to get machine content indexed, and in several cases the trouble eased slightly once they switched to writing themselves. The hedges matter — only "slightly," and indexing is hard for everyone regardless of method. It is offered as a reported tendency worth weighing, not a rule, and not as an instruction to abandon machine tools entirely. 06:45

Can a search engine detect machine-generated content?

Almost certainly, on the reasoning given. Free software already exists that judges whether text was machine-written, so a large search operator would presumably have something better. The episode is careful to keep this as an observation about feasibility rather than a claim about behaviour: detection being possible does not mean it is being used to rank anyone down. It is the premise for the next question, not a conclusion in itself. 07:55

Could a search engine deliberately suppress machine-generated content?

This is put explicitly as a question about odds, not a claim of fact. If detection is feasible and machine content were judged a quality problem, then scanning for it and letting those sites rank less easily is a rational move the operator could make — and the episode's view is that there is a good chance it could happen, while stressing there is no evidence it is happening. The practical reading is to treat heavy reliance as a bet against the odds. 08:14

Should a small site publish a whole library of machine-generated articles?

The clearest answer is behavioural: the agency's own practice is to use very little machine content, if any, with the fuller reasons deferred to another episode. Someone who has weighed the studies and the reports lands not on a ban but on caution. For a solo builder, the safer play is to lean on work that is genuinely yours and keep an audience you own directly rather than betting a whole library on a channel that could shift. 09:06

Sources

Should You Use AI Content in 2026 - Episode 145

The sections above follow the episode's own order. Each timestamp is the point in the recording where the idea is discussed, so the post reads as a map of the conversation rather than a rearrangement of it.

Quotations come from the episode transcript with spoken filler ("okay", "right", "you know", "um", "actually", "kind of", "sort of") and false starts removed. No words are added: square brackets mark anything inserted for sense, and an ellipsis marks any cut. British spellings are restored, and speech-to-text errors were repaired against context only where the intended word is unambiguous. Where a line appears both in a cold-open teaser and in its place in the conversation, the timestamp cites the second — the point at which it is actually said.

Quotes are stamped with the moment they are said rather than with a speaker's name.