Everything we know about AI content for SEO (Plus what experts speculate)

I’m Paul, head of marketing at Verblio. I’m going to walk through everything we know about whether AI content can be used for SEO, plus the upsides, downsides, and risks associated with doing so.

Prepare for extreme whiplash.

Quick important note: SEOs and content marketers have been using AI-powered tools for SEO research for years (think Surfer, Clearscope, MarketMuse, et al).

In this post, I’m referring specifically to using an AI-powered Large Language Model (LLM) tool—like OpenAI’s GPT-4, ChatGPT, Jasper, or any of the other myriad tools using OpenAI’s API—to write content that has SEO intent behind it. That means using any of the above to produce full articles, paragraphs, landing pages, etc. that you intend to rank on Google.

There are other use cases for these tools, e.g. headline writing/brainstorming, writing meta descriptions, but this post is about writing the content itself using AI.

To attempt to somewhat-satisfactorily answer the twin questions, is AI content good for SEO? Is AI content bad for SEO? we need to start with what we know.

Thing we know #1: using AI content for SEO is still an uncertain thing

Anyone who tells you otherwise is lying, extremely naive, or arrogant to the point you should run.

There are zero people not under Google NDA who know how Google will react to AI content in the future. Google itself probably hasn’t decided, and will most certainly evolve its thinking over time.

What this means for you

While all SEO is full of ‘it depends’ and best practices that really should be tested, this level of uncertainty around AI = serious risk. If Google decides to slap down AI content, it could do so in a couple of different ways:

1. similar to how they’ve stated they’re beginning to deal with spam links: devalue any individual piece of content written by AI, or
2. websites with AI content (particularly a lot of it) could receive a sitewide penalty and disappear from the face of the planet Google.

So, yeah, worst case, each piece of AI content you publish is like planting a ticking timebomb with a broken countdown display on your website. Programmatic SEO, long a hotbed of risky SEO practices, has become a common use case for AI content. As we’ll see in the next section, results are… mixed.

All that said: while publishing straight AI output to your website comes with risk, that doesn’t mean you shouldn’t use AI to make you more efficient or give you a leg up in certain situations.

Thing we know #2: Google has already penalized websites with large amounts of AI content.

The theoretical future Google penalties described above aren’t all theory. We have examples already of websites with full, sitewide penalties. Here’s a search traffic graph from one such site:

I’ll let Kevin Indig, who did the case study, explain it here in more detail, but an important caveat on this one is this was bad AI content. Go read it in the article if you don’t believe me, but, like, barely readable bad (and funny, because it’s about celebrities).

Here’s another (undisclosed site)from Mark Williams-Cook (one of my favorite people to follow on this topic):

My caveat on this one: Mark published 10,000 pieces of AI content on this site. This means there are two reasons his penalty could have occurred that don’t condemn AI content as a whole:

1. As Ross Hudgens has been talking about a lot recently, there’s serious danger to publishing at extreme scale very quickly. This could have been the cause of Mark’s penalty, and it had nothing to do with AI. (Personally, I doubt this was the only factor, which leads to …)
2. This was pure, unedited AI content.

Point is: The precedent is here, now, folks.

Ross recently published another case study with AI content that didn’t do so well:

 

He’s careful to note that early results didn’t tell the whole story. When someone shows your a graph featuring AI content getting indexed and ranking (the big spike on the graph), that doesn’t mean the hammer doesn’t drop eventually. 

In this example, the AI content didn’t sink the website, but the growth was a mirage.

Ross also points out that while AI content can rank, it rarely converts. In his words “[AI] case studies will rarely show conversions, as this content will convert horribly due to poor quality.”

What this means for you

For the love of your website, if you’re going to publish AI content, don’t publish shitty AI content. AI needs a human editor for now. And, AI content gives nearly anyone the power to pump out huge volumes of content. That, without really knowing what you’re doing, is asking for trouble.

If you do publish hybrid content produced by AI and edited by humans, keep a very close eye on it.

Thing we know #3: folks are out there ranking pure AI content.

Wait, what?

Yeah, this is kind of the opposite of Thing #2, but with a healthy dose of nuance.

That nuance is some combination of this list:

• Scale isn’t massive (therefore the risk of being caught is lower)
• The content is better (if it’s pure AI, that means better prompt engineering and fine-tuned models)
• The sites are in less competitive niches (less competition, just easier to rank stuff)
• Content is shorter (the shorter it is, the better AI is at doing it. AI breaks down spectacularly on long-form)
• They haven’t been caught (yet)

My formative years in SEO were learning from Whiteboard Friday, where the thinking was if you’re not doing 100% white-hat SEO, you’re an idiot — Google is just too smart. I’ve now seen enoughto know there are plenty of people who win the cat-and-mouse game against Google for long periods of time and make good money doing it.

As I mentioned before, programmatic SEO is one big use case for this. Kevin Indig is working on this and getting results. (A reminder: programmatic SEO refers to creating thousands of pages based on a database of variants – [business keyword] + [location]; thing to do + [location]; Glassdoor ranking for [job title] + salary – stuff like that.) 

Zack Notes and the folks at Sandbox SEO are doing a similar thing:

As you can see from Zack’s post, this stuff is getting indexed, ranking, and driving traffic. But in the comments, it becomes clear that there’s not a lot of monetary or conversion value to this content. It’s not a space that’s monetized; Zack refers to “publishers vs publishers” as being the competition for this content.

Verblio has built hundreds of local landing pages for local SEO that also use this principle. The difference – and why we’re seeing sustained results – is that we’re taking our own advice and having humans edit every piece of copy before it’s published. The result? Our stuff gets indexed and stays indexed.

What this means for you

If you’ve got black-hat skills and want to duke it out with Google, go for it (at least on websites where you’re ok with seeing traffic go to zero).

If ya don’t, or if you don’t work on sites where you can afford to have that risk tolerance, steer clear of publishing pure AI (read: unedited Jasper/ChatGPT/etc) content.

Thing we know #4: Google has commented directly on AI content

Before you get annoyed at me for burying the lead, read this in context of what you read above. And then read the what this means for you bit.

On February 8th, 2023, Google published this blog post about AI content and SEO and updated its SEO fundamentals documentation to address AI content. This builds directly on their January 12th reply to a lengthy thread:

Here is what Google said on Twitter on January 12th, 2023:

There isn’t a whole lot of net new info in here, though it’s nice to hear it directly from the source. Talk continues of not using AI for SEO evil, and, notably, “it’s just content” and “… about 10 years ago, there were understandable concerns about a rise in mass-produced yet human-generated content. No one would have thought it reasonable for us to declare a ban on all human-generated content in response. Instead, it made more sense to improve our systems to reward quality content …”

Holy mackerel, you say. We’re in the clear! Google themselves said so!

Hold on to your hats, folks, things are about to get meta. In the screenshot below, check out what this gentleman Peter, who’s clearly seen this Google-makes-statements-on-SEO rodeo before, commented on the LinkedIn thread where this tweet was being discussed. Then check out how our own head of content (and host of the Content Bounce House podcast) Ryan Sargent responded to my 🙂 pointing out that comment in our internal Slack channel:

If I just got too meta on you,the point is this: There’s a long history of Google saying something about SEO and search ranking factors, then feeling little to no obligation to stand by it. The definition of “primarily for search engine rankings” is also way open to interpretation.

Finally, here is what Google has previously said about user engagement signals: (JK Google doesn’t like to talk about user engagement signals as a ranking factor, but everyone knows they take all that delicious Chrome and Android data and use it + pogo-sticking to understand what humans think of your content.) Point of that long parenthetical is: YOUR CONTENT NEEDS TO BE GOOD TO RANK. And the problem with AI content is…

What this means for you

Again, for the love of Pete,do not publish pure AI content. If Google doesn’t punish you, people who search and find you on Google will punish you by bouncing. And then, Google will punish you.

Thing we know #5: AI detection is an arms race, watermarking is coming (or here)

Here are three things to know about this topic right now (for more on specific tools, read our deep dive):

Here are three things to know about this topic right now:

Most AI detection tools are crap

Publicly-available AI detection tools just aren’t great. They’re not smart enough, which can give you a false sense of security. They’re hard to interpret (and interpreting scores across small volumes of content can be misleading, as scores fluctuate by industry, type of content, and more).

Can Google Detect AI content? Assume Google has built bigger, badder AI detection than is publicly available

Google is heavily invested in fighting webspam, and this includes spammy AI content. So it’s probable they’ve built their own tools to detect it, and some or all of these tools will never be released to the public. (And yes, these tools will still likely be able to catch Jasper content, even if you use their Surfer integration ‘for SEO’.)

At the same time, there’s a lot more investment in AI content creation these days than AI content detection. Consider this entire topic “TBD.”

If watermarking isn’t here yet, it will be soon

You’ve heard of a watermark on a document.OpenAI is working on a watermark for content written by GPT that’s based on cryptographic methods. I’d try to explain the inner workings, but I’m not technical enough. We will give you some more details in that upcoming AI detection tools post, though.

Point is: This watermark will work by varying the probability of the next word chosen in GPT-generated content according to a pseudo-random sequence. These variances will change the words slightly, but not in a way that would be noticeable to humans reading the content. OpenAI’s ‘key,’ however, will be able to recognize the watermark by matching the probability variances, so it can definitively prove something was generated by their model. And worse, you won’t be able to remove the watermark by just doing some light editing — in all likelihood, you won’t remove the right words, or remove enough of them. This is the mother of all AI detection tools.

Ok, how does this apply to SEO? The theory goes like this: OpenAI makes this detection key public, Google uses it, detects content that was written by GPT, and punishes that content.

What this means for you

In a story that parallels many of the black and grey-hat SEO strategies of the past, AI content creation and AI detection are in an arms race. One may gain an edge, the other will surpass… you know the story. In arms races with Google, your opponent is smart, well-funded, and brutal. So:

• Do not assume lowering an AI probability score on a public tool means you’re safe. The level of sophistication isn’t there.
• Do not assume that editing AI content renders it undetectable
• Assume watermarking of GPT content is already happening. (It may not be, but the conservative approach is to assume it is.)

Now that we’ve walked through what is known about AI content for SEO, let’s move to the land of crystal balls. It’s fun in there, and, whether these predictions are right or wrong, they provide valuable insight for marketers.

SEO and AI content creation: what the experts are saying

Some very smart SEOs have put their thinking caps on around AI and SEO. Here’s a roundup of my favorites.

Ross Hudgens, CEO of Siege Media

In this LinkedIn post, Ross is commenting on AI content, but specifically using AI content to push publishing velocity (literally, how fast you’re publishing) to extreme and unhealthy levels. This applies to human content, too, but it’s much easier to achieve with AI. The point? In the short-term, extreme publishing velocity may have a positive impact on your organic traffic, but this will very likely come back to bite you. Ross’ theory: If the ratio of content to backlinks starts climbing too fast, the ‘foundation’ of links supporting the content crumbles.

Lily Ray

If you’ve been around SEO for any length of time, Lily Ray needs no introduction. And this tweet needs little explanation. Lily is predicting that what happened to the sites from the case studies way back in Thing #2 will be happening big time in the near future—a specific Google AI content update, or a broader algorithm update that includes stuff designed to target AI-written-content.

Mark Williams-Cook

Mark makes another appearance in this article with his adaptation of the classic Gartner Hype Cycle. In this case, the trough of disillusionment is the trough where your website traffic goes to zero. To provide some balance to Mark’s pretty negative outlook here, just note that recent well-informed speculation has GPT-4 arriving sometime in H1 2023. So although Gartner Cycles are often measured in years, this one … may be much faster.

Lots of naysaying from our expert roundup. Do note, though, all three of them use “mass content/high volume” in their posts. We’ve seen Google’s dislike of super-high-volume (particularly low-quality, high-volume) before, so it’s hard to argue against the likelihood of this particular use case for AI content ending badly.

Kevin Indig

Kevin is willing to make some bold statements about SGE and the future of search in an AI-powered world:

Where this all nets out (what you should take away)

Phew.

Hanging in there? It’s a lot. If you’re currently feeling some version of, ‘Huh. Well, I want the benefits of AI, but this all freaks me out,’ you’re not alone.

Here’s where I am with all this, and what I think you should take back to your lab website(s):

Assume the following:

• All of this will keep changing, rapidly. I’ve been doing this awhile, and I’ve never seen things evolve this fast. Keep your ear to the ground.
• Google has better AI detection than you have available to you, and it will only get better.
• OpenAI will soon, or already is, watermarking all GPT content. Light editing will not remove the watermark. Heavy editing still probably won’t. Running it through another, different LLM (or maybe a content spinner) probably will. But like, why?

Just say no to pure AI content:

• Beyond the SEO risk, publishing (especially poor quality) AI content represents brand risk, too. Look no further than CNET’s decision to pause its use of AI content after public backlash
• It’s not good for your human users, either. And they matter because they’re the ones who actually buy things + engagement signals 🙂
• Pure AI is so cheap and so fast, it encourages bad behavior for search marketers: publishing too much, too fast, too low quality, too far from the core promise of your website.

Create AI-assisted content with care:

• Spending 5 minutes editing Jasper output won’t make it all better. (We tried.)
• Taking care early in the process (through prompt engineering and injecting humans throughout the process, not just at the end) and using editors trained in the specifics of editing AI are all necessary to make this stuff sing (or at least go from barely readable to good reading)

Where to use AI matters:

• First, consider your website and its place in the internet. Competing in a hyper-competitive niche and have a site where a drop in organic traffic would be catastrophic to you/your client? Yeah, maybe just don’t AI.
• Don’t let your shiny new AI hammer trick you into seeing all your content needs as nails waiting to be whacked. Like humans, LLM writing has strengths and weaknesses. Deploy it on the content where it makes sense, but don’t force it beyond that.

As for me? None of the websites I work on are getting AI content on them anytime soon. They’re in competitive niches. They’re highly dependent on organic traffic for the survival of the businesses. I need people to read the content to convert. AI content might be readable, but it’s not good reading. Last but not least, I’m scared of Google.

This is not a blanket condemnation of AI content. If I were an agency serving local clients and small SMBs that couldn’t afford high-quality human content and had a sad, empty blog? Heck yeah (heck yeah to well-crafted, AI-assisted content, that is).

If you’ve made it this far, thanks for the read. If you have questions or thoughts or disagree with me, paul[at]verblio.com.

Author’s note: this article was originally published on February 3rd, 2023, and updated February 8th, 2023 with new information on Google’s blog post on AI content and SEO. I updated it again on July 31, 2023. This space is changing rapidly, I’ll continue to update this article with new information as it becomes available.