YT YouTube2Text

YouTube2Text → Transcripts → They Lied About AI Content (This Proves It)

They Lied About AI Content (This Proves It) — Transcript

by Caleb Ulku · 3,461 words · 512 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Here's the pattern nobody warned you
0:01about. You publish AI content and it
0:03works. Traffic climbs for 6 months,
0:05maybe 12 months. Rankings hold, you
0:07scale it up because why wouldn't you? AI
0:09content is easy and it's working. Then
0:12somewhere between month 9 and month 18,
0:14the whole thing falls out from under
0:16you. And this is not a dip. 54% of the
0:19sites that followed this pattern lost
0:2030% or more of their peak traffic. 39%
0:24lost half. 22% lost three quarters of
0:29all the traffic they'd spend years
0:30building. And a lot of them ended up
0:32below where they started before they
0:34published a word of AI content. The part
0:37that matters is the delay. If AI content
0:40just didn't work, you'd know it
0:42immediately. You'd publish, nothing
0:44would rank, traffic wouldn't go up, so
0:46you'd stop. This would be a cheap
0:48lesson. But this works long enough for
0:50you to believe in it. Long enough to
0:52build a strategy around it. Long enough
0:54to sell it to a client. And then it
0:57doesn't anymore. So what's happening in
0:59that gap? These numbers come from Lily
1:01Ray. She spent the last several months
1:03tracking more than 220 websites and she
1:06didn't choose them at random. She pulled
1:08them off the customer story pages of
1:10more than a dozen AI content platforms.
1:13Over a dozen companies publicly bragging
1:16about what their tool did for these
1:17sites. So this is a highlight reel,
1:20right? This is what the industry chose
1:22to show you. So she tracked organic
1:24traffic and organic page counts in
1:26Ahrefs and then checked it against
1:28Sistrix to make sure the pattern held
1:30across two data sets. Hey, what's wrong?
1:33You want to go? Let's get back to it. I
1:35want to be straight about what this is.
1:38Third-party estimates, this is not
1:39search console data. She's not claiming
1:41these tools caused anything and she
1:43names zero vendors and zero domains on
1:45purpose because the pattern is the
1:47focus. The pattern is the story and the
1:49pattern is remarkably consistent.
1:52Organic pages climb fast. 6 to 12 months
1:56of AI publishing, traffic peaks 3 to 6
1:59months after the content peaks. Then it
2:01goes over a cliff. And inside the
2:04following year, most of the gain is
2:06gone. Glenn Gabe coined a name for this
2:09chart shape. He calls it Mount AI.
2:12Straight up on one side and straight
2:15down the other. Great name. Now, here's
2:18the detail that tells you what's really
2:19going on. Look at a lot of these case
2:21studies today on the vendor platforms
2:23and they're still there. They're still
2:26alive. The vendor's still running them.
2:28The write-up still says what the tool
2:30did for the client. But the pages the
2:33case study references are gone, removed,
2:36redirected, or 410'd by the site owner.
2:38The success story outlived the actual
2:41pages. Okay, and a lot of these brands
2:44cut down their content footprints way
2:46down in 2025 and 2026. In some cases,
2:49traffic started coming back after they
2:51did it, which tells you they weren't
2:54guessing. They knew what was wrong. All
2:56right, back to the thing I said at the
2:58top. Why this delay?
3:01Because that gap is the whole trap,
3:04right? If this fails immediately, no one
3:06would get hurt. It fails slowly. And
3:09slow failure is what gets you to commit.
3:12You publish 100 pages, they rank, you
3:15get more traffic, so you publish 1,000
3:17pages. By the time the drop comes,
3:20you've built your entire content
3:21operation on top of this and you
3:23probably told the client that this is
3:26the plan. This delay is not Google being
3:28slow. Something specific has to happen
3:31first and it doesn't happen on your
3:33site. This happens on everyone else's.
3:37Okay, so think about what an AI content
3:39tool actually is. It's a template and
3:42it's a prompt. That's the product. And
3:45when it works, the company sells it to
3:47as many people as possible because I
3:50mean, that's what companies do. That's
3:52what they're supposed to do. So now, you
3:55have a thousand websites running the
3:56same template through the same prompt,
3:5910,000, and every single one of them is
4:02pumping pages into the index that are
4:04structured in the same way, ordered the
4:07same way, answering the same question
4:10with the same shape. When you're the
4:12first one doing this, there's nothing to
4:15detect. You look like a site that
4:17publishes a lot. Maybe you hired some
4:20new editors. When 10,000 sites are doing
4:23this with the same tool and the same
4:25prompt, you're not a site anymore,
4:28you're signature, a footprint, and
4:30that's the delay. Google isn't sitting
4:32there for 9 months trying to decide
4:34whether your page is good, right?
4:36Google's a machine, it makes these
4:37decisions much faster than that, but
4:39Google is waiting for enough of the same
4:41pattern to show up that it becomes
4:44obvious. And once it's obvious, it
4:46becomes cheap to act on. And everybody
4:48running that template, the same template
4:50with the same prompt, they all get
4:51caught at the same time, including you,
4:53even if your version was better than
4:55their version, which I'm sure it was.
4:56This is what Google means when it talks
4:58about content made for search engines
5:00instead of content made for people. This
5:02isn't a comment on how it reads, it's a
5:04comment on why this content exists. Now,
5:07Ray went through the top traffic URLs on
5:09the declining sites and found eight
5:11templates that kept showing up. Most of
5:13the sites in trouble were running three
5:15or four of them. The ones most impacted
5:17were running all eight templates.
5:20Comparison pages at scale, product A
5:22versus product B, every match up in a
5:24category. What is X? Glossary pages, one
5:27term each, built to get quoted by an AI
5:30engine somewhere. Sometimes
5:32grammatically translated into a dozen
5:34languages off of one template. Best X
5:38for Y listicles, straight out of the
5:40affiliate era. If you have been in the
5:43space long enough, you know what I'm
5:44talking about. We also have
5:45self-promotional listicles, the same
5:47thing except the company writing it is
5:49in the category and puts itself at
5:51number one with no evidence they ever
5:53even tested a competitor. Competitor
5:56alternative pages, one for every single
5:58named rival in the market. We also have
6:00programmatic location and language
6:02pages, one template times every city or
6:05country that Google will index around a
6:07business. FAQ farms, one question per
6:10URL, answer in the first paragraph,
6:13bullets in the body, schema at the
6:15bottom, and off-topic content at scale.
6:19B2B software companies publishing jokes
6:21and baby names because the keywords have
6:23volume. And I want to say something
6:25about that list because this is the part
6:27that people get wrong when they repeat
6:29it. Not one of those is on there because
6:31it doesn't work. Every single one of
6:33those works. That's the entire reason
6:3510,000 sites are running those
6:38templates. Ray says this outright and
6:40people skip right past it. These pages
6:43rank. They get cited by AI and that's
6:46exactly why enough people created them
6:49that it became a signature. Okay. Now,
6:51six of the eight have almost nothing to
6:53do with a local business, have nothing
6:54to do with local SEO. You're not
6:56building competitor alternative pages
6:58for a local plumber, but two of them are
7:01being sold to local operators right now.
7:03And one is the fastest way I know to get
7:06a local site in real trouble. Let's talk
7:09about the FAQ farm first. Local has been
7:12hit hard with this pitch because it
7:14sounds so reasonable. Customers ask
7:16questions, AI engines answer questions,
7:19so build a page per question and get
7:21quoted. The most common tactic is to use
7:24the people also ask questions that
7:26Google itself provides. There are two
7:27problems with this. It's the same
7:29footprint issue since every local site
7:31running the same playbook is going to
7:33generate the same questions with the
7:35same answer structure coming straight
7:37from Google's people also ask. And
7:39Google deprecated FAQ rich results
7:42earlier this year. So, the thing that
7:44made these pages visually attractive in
7:46the search results is now gone. What's
7:48left is a few hundred thin URLs on a
7:51site that should have 40.
7:53Okay, the second one is the real one.
7:56Programmatic location pages. Now, here's
7:58the version being sold. Take your
8:00service pages, spin them across every
8:02city in a 50-mi radius,
8:05use AI to write these hundreds of new
8:07pages, and now you rank everywhere.
8:09Plumber Houston and plumber Katy,
8:10plumber Sugarland, plumber Pearland, on
8:13and on and on. One template, swap the
8:15city name, have AI crank them out,
8:17publish 200 pages. I've watched sites
8:20get taken out for this for years now.
8:23This is not new, and it's not just an AI
8:26tactic. AI just makes it fast, and AI
8:29means you don't need to find replace
8:30city names. But, the reason this fails
8:32isn't volume. It's that the pages claim
8:34something untrue. A page that says
8:36plumber in Sugarland is telling Google
8:38and telling a customer that there's a
8:40plumbing business in Sugarland. No
8:42address there, no coverage there, no
8:44jobs completed there, and the page is a
8:46claim with nothing behind it. Google has
8:48an enormous amount of local data. It
8:51knows where you are. You told Google
8:53where you are when you created your
8:54Google Business Profile. Now, I teach a
8:57system that generates a lot of pages for
8:59local SEO using AI, and someone is going
9:02to point at this and say it's the same
9:04thing. But, let me talk about the
9:05difference that's worth understanding.
9:08The structure that I focus on is bounded
9:10by the Google Business Profile. Your
9:12categories become category pages. Your
9:14services, they become service pages. If
9:17it's not on the Google Business Profile,
9:19it doesn't get a page. You end up with
9:21around 30 URLs, and every single one of
9:24them targets the city that your GBP is
9:27addressed in. Not the city that you wish
9:30you ranked in, the one that's on your
9:32profile. So, the page count comes from
9:34what the business actually does in the
9:36place it's actually located. This the
9:39programmatic version comes from what a
9:41spreadsheet can multiply and 199 of
9:44those pages describe a business that
9:46doesn't exist at all in that city. Count
9:49URLs and those look similar, but they're
9:51opposites. Okay, so that covers what not
9:54to build, but knowing what to avoid
9:56doesn't get you ranked. So, we don't
9:58want to just stop there. Okay? We don't
10:00want to just be someone with 30 pages in
10:03Houston. The test that separates a page
10:05that Google wants from a page that
10:07Google tolerates and it has nothing to
10:10do with whether AI wrote it. Ray puts
10:13this as a question you ask before
10:15publishing anything and
10:17it's very simple. Could a competitor
10:20publish a near identical version of this
10:22page tomorrow using the same prompt, the
10:25same methodology? That's the test. And
10:27if the answer to that is yes, then the
10:29page has no reason to exist. It might
10:31read beautifully. It might read like
10:33Hemingway wrote it. It might be more
10:35accurate than anything else ranking.
10:36Maybe it deserves a Pulitzer Prize, but
10:39if anything reproducible by anybody with
10:42the same software and the same input
10:45it's a copy and it just happens to be
10:47the first one to be produced and that's
10:50the dividing line in this whole study,
10:52which is why I want to be precise about
10:53it. The sites that got destroyed weren't
10:56destroyed by using AI. The sites that
10:58got destroyed weren't destroyed because
11:00they used AI. Google's own documentation
11:03says the production method isn't the
11:05issue. When they rolled out the scaled
11:08content abuse policy alongside the March
11:102024 core update, they named the
11:13practice specifically as generating
11:15pages to manipulate rankings regardless
11:18of authorship. Human AI, it doesn't
11:20matter. What killed them is that the
11:22pages carried nothing that wasn't
11:24already on the internet. We call this
11:25information gain. Does this page add
11:28something above and beyond what's
11:30already ranking? And we compare this
11:32with the second question, does this page
11:34exist because a real customer needs it
11:37or because a search engine or an LLM
11:39might cite it? Okay, you know the answer
11:41to this. You always know. And here's the
11:43part that I want the local SEO people to
11:46hear because you have an enormous
11:48advantage over every B2B SaaS software
11:52company, okay? Those companies were
11:55writing about software. Everything true
11:58about their product is already published
12:01on their own website, on G2, on 100
12:04review sites and comparison posts. They
12:06were trying to manufacture information
12:07gain in a category where there's almost
12:09no information left to gain. This is how
12:12you end up writing about baby names,
12:15okay? But in local SEO, that's not the
12:19position you're in. A local service
12:21business generates brand new information
12:23almost every day and almost none of it
12:25goes on the internet. Every job is a
12:27data point that other people don't have.
12:30What the pipe looked like when you open
12:32the wall? What the repair cost in that
12:34neighborhood last month? What which
12:37failure keeps showing up in houses built
12:39in the '90s? The failure mechanism that
12:42keeps showing up in houses built in the
12:44'20s? What permit is the city slow to
12:47give you right now, okay? That's
12:49information gain. It exists. It's free.
12:52It's sitting in your job history and in
12:53your text heads where no one's ever
12:56bothered to put it on a page. So, the
12:57question stops being whether to use AI
13:00and it becomes something more useful.
13:02What is the AI actually for? These tools
13:06work when an experienced person is
13:07holding them and they fail when someone
13:10sets them and forgets it. Raise words,
13:12true but vague. So, let me make this a
13:15little bit more concrete for a local
13:17business. AI does the parts that don't
13:19require you to have been there. It can
13:21build the structure. It can map your
13:23categories. It can map your services and
13:24do site architecture. It can write the
13:26brief. It can handle title tags, H1s,
13:28the internal linking plan, the schema.
13:30>> [gasps]
13:31>> It can take your notes and turn them
13:32into clean copy. It can research what's
13:35already ranking, so you know what's
13:36already been said, so you don't repeat
13:38it. This is real work, and without AI,
13:40it can take days. And there's nothing
13:42dangerous about any of this, because
13:43none of this is the part that makes a
13:45page worth indexing. Okay, but
13:46formatting isn't the ceiling, and this
13:48is where a lot of people stop. Okay,
13:50this is where those 220 sites also
13:53stopped. One prompt in and one page out.
13:58That's the architecture of the tools in
14:00that study, and that's why every page
14:02they produced was reproducible by the
14:04next guy with the same subscription,
14:07with the same prompt. The other way to
14:09build this is to stop treating the model
14:11like a writer and start treating it as
14:13something that goes and collects. Okay?
14:16That's what I built the Core 30 Agent to
14:17do. It's software focused on local SEO.
14:21Before it writes a single page, it pulls
14:23Census Bureau data for the actual
14:24market. It pulls Reddit threads where
14:26real people are asking about this exact
14:28service in this exact kind of house with
14:30this exact problem. It's pulling Search
14:32Console data, so the page is built
14:35around what people are already asking
14:38that business specifically. It's pulling
14:40in questions that come off of local
14:41forms, and it's pulling what's unique
14:44about that client and what makes that
14:46business different from the three other
14:48companies in town. Then, it comes up
14:50with a story. What story can it tell
14:53based on all of this information and
14:55data? And it creates an outline for that
14:57story. Each story is unique, and beyond
15:00that, it tracks that story. So, the next
15:03time you write an article, the next time
15:05you write a page, it never reproduces
15:07the same story again. And finally, it
15:10writes. And the page that comes out is
15:12not reproducible, because no one else
15:14has this combination of inputs. A
15:16competitor can copy the prompt, but they
15:18can't copy your Search Console, your
15:20form submissions, or your position in
15:22the market. Now, there's still a ceiling
15:24on this, right? AI can't know what
15:26happened on Tuesday on the house that
15:28Elm Street, at least it's not Friday,
15:29right? But, on a water heater
15:31installation page, the AI-written
15:33version says that water heaters
15:35typically last 8 to 12 years. Signs of
15:37failure include rust-colored water and
15:40popping sounds. Man, you should hire a
15:42licensed professional. All of that's
15:44true, but all of that's on the first 10
15:46results already. Zero information gain
15:49and reproducible tomorrow by anyone. The
15:52version that survives says what a
15:5450-gallon gas replacement actually runs
15:57in your city this year, what it pushes
15:59to the high end, it says the permit
16:01takes about a week, maybe two. It says
16:04the permit takes about this long with
16:06your city, and here's the inspection
16:08they always flag. It says in the older
16:10neighborhoods with the original
16:11galvanized, the install takes longer for
16:13this specific reason. It has photos of
16:16your truck in a driveway that somebody
16:17recognizes and names the tech who does
16:19them. Now, run the test on this. Could a
16:22competitor publish a nearly identical
16:23version tomorrow with the same prompt?
16:25No.
16:26They couldn't. They'd have to go and do
16:28the work. And that's the whole thing.
16:30That line is whether there's something
16:33on the page that could only have come
16:35from you. And every local business on
16:37Earth has that, but almost none of them
16:40are putting it on the site. Hopefully,
16:42this makes sense. What What we're
16:43talking about here is writing content
16:46that makes it sound like you actually
16:49live and work and do this service in
16:52this neighborhood. Okay, this isn't
16:54mind-breaking. This is kind of obvious
16:58in retrospect. If we want Google, if we
17:00want the AI systems to trust that you
17:02are a local business that services this
17:06neighborhood with water heater
17:07replacements, then write content that
17:10makes it sound like you service water
17:12heater replacements in this
17:13neighborhood. That's all I'm saying.
17:16Here's why I'm confident that this is
17:17the direction and not just my opinion.
17:20The industry already ran this exact
17:21experiment once. September 2023, Google
17:24shipped the helpful content update. This
17:26is maybe a year after GPT-3.5,
17:29the most aggressive move that they had
17:31made in years against content that, in
17:33their words, feels like it was created
17:35for search engines instead of for
17:37people." Six months later, March 2024,
17:39the longest core update in Google's
17:42history, and Google stated the goal
17:44outright, "Reduce unhelpful, unoriginal
17:46content in search results by 45%." This
17:49is Google desperately fighting against
17:52AI slop starting less than a year after
17:54ChatGPT-3.5 comes out. Two updates
17:57back-to-back, the same target both
17:59times, content [snorts] produced at
18:01scale to rank. And alongside the second
18:04one, they wrote the scaled content abuse
18:06policy, which named the practice and
18:09went out of its way to say authorship is
18:11irrelevant. Human, AI, some mix of the
18:15two. This was two years before most of
18:17these AI content tools existed. A lot of
18:20the sites hit in that round never came
18:22back. Ray spent 2024 working with owners
18:25trying to dig them out, and plenty of
18:26them were publishing original,
18:28human-written content. Their crime was
18:32running it through so much SEO structure
18:34that it stopped reading like a person
18:36had a reason to ever write it, much less
18:39read it. When she says that the content
18:41coming out of these AI tools looks and
18:43feels like the exact stuff that got
18:45wiped off the map in '23 and '24, she's
18:47not predicting. She sat in that wreckage
18:50for a year. The packaging changed, but
18:52the pattern didn't. And the sites that
18:55came through that cycle in the best
18:57shape were the ones that went narrow
18:59instead of wide. Fewer pages, tighter
19:02topics, real experience on the page.
19:04Same answer this cycle, same answer next
19:08cycle when people panic about AI-written
19:11content yet again. So, this brings up
19:14something that's about to make all of
19:15this a lot more visible. Anthropic
19:18announced in mid-August that it's going
19:21to start watermarking Claude's output.
19:23Now, I did a full breakdown on that one,
19:26how that watermark affects SEO, and what
19:28it does to AI-generated content on your
19:31website. It's linked right here. Go
19:33check it out.

About this transcript

This page contains the full transcript of They Lied About AI Content (This Proves It) by Caleb Ulku, generated from the public captions YouTube serves with the video. The transcript has 3,461 words across 512 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.