Every so often something comes along and is touted as the thing to finally put an end to Google’s search engine reign. Not long ago that thing was TikTok. Now, us marketers can’t look anywhere without hearing about ChatGPT. This artificial intelligence that generates new content (or generative AI) is the latest thing to be considered a Google killer. But can it really upend the search engine giant? When Microsoft introduced it as part of Bing in February, it definitely started changing things in the search world.
What is ChatGPT
ChatGPT is a language model. More specifically, ChatGPT is a chatbot interface built on the GPT-3 large language model. GPT-3 was first released by OpenAI in the summer of 2020. The language model is trained on the largest amount of text of any other model, so far. The New York Times reported that from this massive training, it identified 175 billion parameters, or patterns that map human language. In addition to its unprecedented scale, GPT-3 can also take in just a few examples for “few-shot learning.” OpenAI’s chief scientist, Ilya Sutskever, explained few-shot learning as “any layperson can take this model and provide these examples in about five minutes and get useful behavior out of it.” By overlaying a chatbot on this model and opening it to the public at the end of November as ChatGPT, OpenAI made it very useful to any person. As of March 14th, ChatGPT runs on GPT-4. However, these types of models come with a lot of caveats that need to be considered.
Problems with ChatGPT
First of all, ChatGPT is trained on data only up through 2021. When asked questions like “who is the current British prime minister?” or “who owns Twitter?” the model returns incorrect information or indicates it has information only through 2021.
Secondly, the model is biased. Machine learning tends to pick up on biases when it is fed data that hasn’t been vetted to remove it. Because this model was trained on data from the web, it is unlikely that any objectionable content was purged. OpenAI has been working to remove some of its biases and toxic language, but that task is also being handled in an unseemly way, as reported by TIME. Unfortunately, to train the model to remove toxic content that content must be surfaced. OpenAI outsourced low-wage labor to handle that without supplying mental health resources.
Bias is a common issue with machine learning that tends to get a lot of press. Another common issue with neural networks, which this AI is, that isn’t mentioned much is that these learning models suffer from hallucinations. That’s right; neural networks can output information that is not “faithful” to the input given , which data scientists call hallucinations. These hallucinations are likely where ChatGPT has completely made up books and fakes its sources when asked for citations
Finally, privacy has become a major concern for ChatGPT. Users reported seeing other people’s chat titles (but not the actual conversations) in their history. These reports caused ChatGPT to go down for a fix on Monday, March 20th.
What Does This Mean for the Future of Search
Microsoft and Google are racing to introduce a useful chatbot interface even directly in the search engines. Even before generative AI was integrated into search engines, creators and marketers were already testing how the technology could be used to scale content and grow audience reach on said search engines.
ChatGPT’s Integration in Microsoft Bing
On February 7th, Microsoft launched ChatGPT within Bing to select users. Quickly, the issues with ChatGPT were showing up in Bing as well. Users were reporting “unhinged” responses. These responses showed the problem with hallucinations. However, one might consider these conversations stemmed from prompts goading the AI into wild responses. Thus, Bing pushed an update that limited the number of prompts per session and day. Bing hasn’t stopped trying to improve the chat interface, adding the ability to switch the type of response to precise, creative, or balanced. When GPT-4 was announced, it was revealed that Bing chat had been running it the whole time, too.
Since the integration, even though it is still a limited rollout via a waitlist, Bing’s usage stats have improved. Bing has topped 100 million daily active users. A third of the preview users of Bing chat are using it on a daily basis.
My SEO perspective on Bing Chat
A plus for search engine optimization professionals (SEOs) and brands, Bing chat includes citations in its responses. However, I’ve found the links, especially those from ads (yes, I’ve seen Bing Ad links integrated into the responses) don’t always make the most sense to the context. In the screen below, the ad mention next to Autotrader points to a competitor which seems unfair to both brands (except maybe for the ones looking for clicks, although likely very low value ones from that experience) and users.
When it comes to trying to use it as a free and accessible version of ChatGPT to generate ideas or content, the options to pick between precise or creative seem like they’d be quite useful. However, it might not be that helpful, which you’ll see in comparison to Bard.
Google’s Generative AI, Bard
In response to Bing’s initial announcement in January that it would be incorporating ChatGPT into its search engine, Google’s founders were called in to form a plan to roll out its own chatbot by year’s end. Google promised its version would prioritize “getting facts right, ensuring safety and getting rid of misinformation.” Those claims sounded like being behind on a chatbot feature may actually help put Google ahead of the competition. (Spoiler alert: it didn’t).
On February 8th (yes, exactly one day after Bing started its preview rollout), Google announced its AI chatbot, Bard. In the demo, incorrect information was displayed in a response which caused Google’s stock to drop 8%! Because of that fail, it seemed like Google took an official rollout a little slow (compared to the speed at which everything had been changing up to that point for it and Bing). Bard didn’t start rolling out to some users until March 21st. Early signs are not promising, especially for marketers.
My SEO perspective on Google Bard
My first few conversations were testing prompts tied to my work sites. The answers were very similar to answer boxes and other search results features tied to the Knowledge Graph (shameless plug, I’ll teach you more about all of that at my Integrate session this June).
Big, terrible miss, though; it doesn’t include citations. Instead, after each response a button is provided to “Google it.” For people researching, the lack of links is terrible, but it might be even worse for brands and marketers. Microsoft folks have had a good laugh about it already. In its FAQ, Google defends its lack of links by saying its AI is meant to create original content and not replicate what already exists
That defense rings true, too. It wasn’t until I started trying to unlock the core of a generative AI tool, creating new content, that I was finding Bard’s responses go beyond that which you might find in search. Plus, Bard gives you other drafts of its response for you to review.
Bing Chat vs. Google Bard
Early signs point to Bing staying ahead of Google. Citations help give some credibility to these models that are trained on massive and mysterious datasets. It doesn’t help when Bard lacks answers without a lot of prompting to get to them.
I asked both chatbots “what foods are good sources of potassium.” Bard’s opening paragraph even has an example prompt of “help me incorporate more high-protein vegan options in my diet.” Yet, when I asked my question, it couldn’t answer it all.
It took me going about that prompt in a very roundabout way to finally get a proper response. I had to ask if it could help with recipes. It said it could and suggested I asked about a banana bread recipe. Instead, I asked what it could tell me about bananas. Within that response, it called out that the fruit has a lot potassium. I then asked “what else has a lot of potassium” to get a response. Oddly, Google Search has a full Knowledge Graph panel for “potassium” with the same prompt. So maybe Bard isn’t tapping into as I had originally thought or the language model was struggling to get to the main topic of my prompt without additional context. Meanwhile, Bing’s chatbot had an immediate answer with citations.
With Bing’s responses set to be “creative,” I tried seeing how these tools might help be more descriptive for say a product description. I prompted each how to describe “yellow.” As you can see, Bing gave me a run down of different shades of yellow in a very matter of fact way. Didn’t seem very creative.
Bard’s response seemed more creative. It offered up several options for descriptive pairings, like “stimulating and energizing.”
Finally, I asked how to describe yellow as Stephen King would. Not bad.
Google said Bard was meant to create new content and replicate, and that it does. It failed on this prompt across all drafts, too.
Takeaways from my small sampling
- If you are truly generating content, having options via drafts and “original” content (to everyone feeding it the same prompt atleast) gives Bard an advantage.
- If you are researching, use Bing.
- If you want derivative content, use Bing.
Will AI-Generated Content Rule Search Results?
With these tools allowing for writing a whole page of content with a simple prompt, the likelihood of content marketing and SEOs using it to scale website content is likely. In fact, it is already happening!
For some time now, the SEO world has believed it is a risky game to try to use AI content to rank in search engines. Google has been saying for years that machine-automated content goes against its guidelines. However, it was left open by one of Google’s search liaisons that potentially AI content will be accepted as long as it is quality and helpful. On Twitter, discussions would pop up asking whether or not a site was using AI content or some would deliberately talk about testing it for themselves. Chatter would show mixed results with epic traffic gains followed by swift losses, maybe due to a manual penalty.
Recently, before ChatGPT was released, the discussion started heating up about AI content with the Helpful Content Update release in August 2022. Google claimed the update focused on ensuring that “helpful content, written by people, for people” were seen in search results.
Not much longer after that update, Google announced a change to its search quality raters guidelines. These search raters are actual humans testing the performance of the search ranking systems. E-A-T– expertise, authority, and trust – was the guiding principle for raters to determine quality sites. In December, Google added another E to the concept to cover “experience.” Sharing actual experience with a product or service helps ensure originality. All of these recent updates seemed to be a set up for combating the forthcoming AI content rush. If everyone is using the same ChatGPT prompts, originality will be a thing of the past.
With the renewed interest in AI content, some SEOs noticed some major news publishers using AI-generated content and indicating such. With that revelation, SEOs finally asked for clarification on the guidelines. Google’s official search liaison tweeted back that the only real concern is content written specifically for search engines, instead of users.
Sites like Bankrate and CNET are still providing useful content to keep ranking with AI generation, right? Turns out that might not be the case. After SEOs started pointing out and analyzing the AI content on those sites the cards began to fall. All of the issues with neural-network models were exposed. CNET had to post that it had found factual errors in some of the AI-written posts and were reviewing them. Then it completely paused any more posts to be written in that way. The media coverage didn’t let up, though. Futurism reported how those posts were flagged for plagiarism. Of course, the model has to borrow from its training data. That Google quality rating update to encourage writers to share their experience seems really smart now, doesn’t it?
Another recent issue with AI content appearing in search is that review spammers are using it to automate fake reviews in Google. The spam is quite obvious (for now) since the response is indicating it is “an AI language model” in the review comment. However, these review changes at scale could boost or drop a business in local SEO ranking until it is identified by Google as spam.
How to Use ChatGPT for SEO
So if CNET has shown us that AI content may not be ready for its time in the spotlight, what can we do with generative AI? Think back to what these tools are at their core, a natural language processing (NLP) model. NLPs are built to do things like break down syntax, understand sentiment, classify text, and the like. Sticking to NLP tasks, SEOs can use these models to:
- Determine the syntax and sentiment of content and see if a pattern emerges for better ranking content
- Reorganize a website menu or other internal linking by classifying articles into specific categories.
Both ChatGPT (in Bing too) and Bard offer some additional skills beyond a typical NLP by understanding prompts, even prior prompts with short-term memory, and outputting text. This allows us to use these for:
- Keyword research
- Topic generation
- Meta description writing
The super-talented international SEO consultant Aleyda Solis goes into more detail about those uses and has even more ideas.
If you are interested in learning more about the use of AI in marketing or search engine optimization, check out IMC 619 Emerging Media and the Market and IMC 642 Web Metrics and SEO.
Interested in growing your career with a master's degree in Data, Digital or Integrated Marketing Communications? Request more information today!