AI Detectors Are Ruining SEO

0:00An AI detector marked this content as 100% AI generated, but it was written before chat GPT existed. Now, if you're using these scores to decide whether your SEO content is ready to publish, a result like this can send you back to rewrite a perfectly good page. Now, I used to pay attention to these scores also, but getting the percentage down can actually make your content worse on the metrics that we actually care about for SEO. There are two checks that I would use before publishing, and neither

0:27one of them requires an AI detector. Now, this score is based on an email that my email copywriter wrote back in 2019. I know who wrote it. I know when it was written, but if this had been a new article from a writer I had just hired, that same result with a 100% score could have made me question their work or send it back for a rewrite. It looks clear. You have a percentage on the screen that gives you a verdict. It feels like something you need to fix before the content can be emailed out or go on a client's website. But if you run that same content through a different

0:59detector, you can get a very different answer. Here are a few other detectors we use for that same email. These detectors use different methods, and they don't all agree. In fact, you can often get a different score if you run the same content through the same detector multiple times. Which detector are you editing for? Which one's score do you care about? And none of this has actually told you whether the page is accurate or whether it answers the customer's questions. In addition to this, we saw another problem when we were comparing content that the core 30 agent was writing to the zero GPT scores. Some of the content with very

1:33low scores had awkward stop and start rhythm. Lots of very short sentences one after another that made it difficult to actually read. This example here that I have pulled up scored 17.3. Very, very low. But if we read it, this is the most common mixup. An owner says their dog is covered and they mean it. They bought a product that kills fleas and ticks, but fleas and ticks live on the outside of your pet. Heartworm lava travel inside different parasites. It's difficult to

2:00read content. Short sentences can work. They can work perfectly well, but when you keep doing it, the writing starts to feel choppy. Uh, it starts to be hard to write. I'd want to edit these passages before publishing, even though it's at 17% AI detection. Now, I'm not suggesting that everyone using a humanizer is deliberately making the writing worse. The problem is that

2:23getting a low score becomes the goal. That's the metric you're working toward because that's the number you have in front of you. So, you keep changing the wording until the tool says it's human without necessarily checking whether those changes actually improve the page. We used to have a function in the core 30 agent. It would automatically generate the content and run it through another step to make it sound more human. But when we reviewed what that step was doing, we found that it was actually reducing theformational additiveness of the content. Useful details were getting lost or removed in the rewrite. So, we got rid of that

2:58function from the generation tab. If a sentence is clunky, it should be fixed. If the content repeats itself, cut the repetition. I still want the writing to be clear and easy to read. But if the only reason you're rewriting it is that a detector gave you a high AI score, you really should be able to explain what the rewrite is supposed to accomplish, especially if you're doing all of this for SEO purposes. HRES actually studied the relationship between these scores and rankings. They analyzed 600,000 pages and compared their AI detection scores with their Google rankings position. The correlation was 0.11, effectively zero. But in a follow-up

3:37study, they found some association between higher AI scores and weaker performance. Right? Pages with higher scores tended to rank very slightly lower. They were also less likely to be indexed and they received fewer impressions. So the research doesn't prove that there's no relationship at all. But it is looking at the wrong aspect. If someone generated hundreds of generic articles without checking them or adding anything specific to the business, those pages would likely score highly for AI and they would also score poorly in Google. Someone who generates hundreds of generic articles without checking them or adding anything specific to the business or the location. Those pages would potentially

4:18score highly for AI, but they would also perform poorly in Google. One does not explain the other. It does not establish that the detector score explains the poor performance on Google. Neither of those studies tested whether taking the same page and improving its AI detection score actually improved its rankings. That's what I would want to see before

4:37adding that step to our own process. Because when I review content, I need to know what to change. A wrong fact gives me something to fix. A missing explanation gives me something to add. The percentage AI written by itself doesn't give me that direction. What I want to see isformationally additive content. That means the page gives the reader something useful beyond what's already available on the search engine results somewhere else. Now, you can ask AI to rewrite the top 10 results, changing all the wording, and that will give you something that passes a plagiarism check. And then you can ask the AI to reduce the AI score and have a

5:14very low AI score. But the reader still gets the same information. And Google has been working on this, right? Very frequently. The way AI actually works is it goes and finds information and then basically regurgitates it back to its users. So if you ask AI to write an SEO optimized article, you are almost by definition getting something that has noformational additiveness. Google wants to get AI slop out of its index. So they have a patent. They called it contextual estimation of link information gain. I thought I was bad at naming things. One version of this patent was granted in

5:48June of 2025. two years after chat GPT three and a half's release that describes a way to evaluate whether a document adds information beyond what a user has already been shown. Now, of course, the patent doesn't prove that this exact system is deciding where your page ranks on Google, but it does describe a distinction that matters when you're creating content. Changing the wording does not add information. And Google signals seem to show that it cares a lot more about the information, not the actual words. So, for a local business, we have access to information that isn't in the top 10 results. You can talk to the owner, find out how they do the

6:25work, what customers ask before hiring them, what affects the services in their local area. If all you give the AI is a service and a city, you're going to leave it to fill the page with general information that it found about that service. And if you run that result through a humanizer, it's still just

6:42generic information about the service. It might just sound different. Now, a customer who's reading that page is trying to decide who to hire, even if that customer is reading another AI's summary of that page, explaining the service as part of it. But they should also learn why hiring that particular company would mean. And it brings us to the two checks, right? The first is whether a competitor could take the page, change the business name, and put it on their own website. If they can, and it would make sense, then we haven't given the customer much reason to choose this specific business. Now, some information is of course going to

7:16overlap. Two plumbers can explain the same repair. You don't need to invent a different explanation to be unique, but once I've read the page, I should know something about the company who's doing the repair. How do they approach the job? What do they check before recommending a replacement? What's included in their service? Whatever is accurate and helps the customer make

7:36that decision. saying that you pride yourself on quality service. Regardless of how many times I see this on local business websites, it is not a differentiator. Unless you happen to live in a town where everyone else is openly committed to doing a bad job, they can all say the same thing. Tell me what you actually do that makes your service better. The second check, can we take this page and swap the city name with a different city a thousand miles away and have that page work just as well? If I can do that, we haven't given the reader much useful local information. Adding a list of

8:13neighborhoods does not fix that. Please stop adding lists of neighborhoods in your content. That doesn't make it local. I want to know whether something about the location affects the work or affects the customer's decision. An electrician working on older homes in the Northeast, and most of the homes in the Northeast are older, they might regularly deal with horsehair plaster and not drywall. If that's something this business encounters, explain how it affects getting wiring to the walls and what the homeowner should expect about damage and repairs. If you moved a thousand miles toward Florida, you would

8:43not have horsehair plaster issues. Florida's homes just aren't old enough for that. This tells the customer something useful about hiring this business to work on their home. A list of local zip codes does not. Okay. Other details you might talk about local requirements, conditions the company regularly deals with, or even practical limits of where they can get to and when. These details come from the business and/or reliable research. We're

9:09not asking AI to invent local problems. We don't want the AI to hallucinate. We don't want the AI to invent things to make the page sound locally specific. all we're talking about. We want the page to read like it's from a local company doing this work in this location. That's it. That's what we're trying to achieve. The vast majority of local business content does not clear that bar. And if the content doesn't clear the bar, then I need to know where to focus. I need to find out more about the business or what matters about serving customers in that location. The easiest time to get that information

9:44into the content is before it's generated. If you're using the core 30 agent, that's our proprietary tool for local SEO research and writing. Take the time to fill out the client settings. You have the opportunity to do an audio interview where the core 30 agent will literally ask you about your business. Tell it about the company, who's behind it, how they do the work, what makes them different, include local details, ramble, talk about things that affect your customers. And if you're not using the core 30 agent, if you're using another AI tool, it still needs that information. The model can't know what makes you different if you don't tell

10:19it. Then read the output and check whether it use those details properly. Filling out the settings, feeding this information into Claude doesn't mean every claim in the finished article is going to be correct. We still manually review ours. We check that the service descriptions match what the company offers. Fix anything that doesn't make sense, and we add missing answers or relevant internal and external links. We would never publish any AI generated

10:43content without reviewing it first. These are the things that we check before we publish along with all of the typical technical requirements and every chain should fix a specific problem. This says we offer a service that we don't provide. This answer is missing something that customers need to know. We have a related page that we should link to. You can act on these instructions. Get this below 10% AI detection can send you through a multiple rewrite phase without fixing a single problem for the customer, without doing anything to help performance on Google. Now, publishing AI generated slop can still do real damage to a website's traffic. In this next video,

11:20I'm going to break down a study of sites that grew and then lost traffic after scaling AI content. Watch that next to see how to do AI content