Why better search algorithms will encourage better writing

So, it’s over. HCU has been implemented. What actually happened? 

For those not in the know, Google recently implemented their “helpful content update” to Search so that users can “see more results with unique, authentic information,” and to “make sure that unoriginal, low quality content doesn’t rank highly.”

That Google is taking steps to improve the quality of search results is great news for everyone: unscrupulous writers have been trying to game search engines with low-effort, low-quality content for as long as search engines have been in existence. Most of us have had the experience of opening a link and being disappointed to find that the content behind the link is a generic, repetitive, keyword-stuffed word salad clearly created for the purpose of grabbing clicks.

Writers and SEO specialists did, however, wonder what the helpful content update could mean for the future of AI and content creation. In particular, writers who use AI writing platforms to ideate, draft, edit, distribute, and repurpose their content were worried that Google might be targeting AI-generated content specifically — and that their content might be penalized as a result.

AI writing tools can empower good writers to greatly accelerate and scale content production. Unfortunately, unskilled or unethical writers can also use AI writing tools to scale the production of low-quality content. As it turns out, early analyses of the impact of the helpful content update suggest that the sites that have been punished so far are those that produce large quantities of low-quality, unoriginal content.

Now that we’re almost two months removed from the helpful content update announcement, it’s become clear that the update wasn’t specifically about targeting content created with the aid of AI: it was about penalizing unhelpful and unoriginal content more generally. And as writers, we should be making content that provides real value to our target audience — whether that content is written with AI assistance or not.

Combating low-quality content

Readers aren’t looking for content that ranks high on search engines per se: they’re looking for written content that is well-crafted, informative, and insightful. Ideally, the role of search engines is to help readers find this kind of content. But when search engines can be easily gamed, there are perverse incentives to make content that caters to search engines, and not to people. 

Google has been fighting against black hat SEO techniques like invisible keywords, link farms, and sneaky redirects since long before AI writing tools became sophisticated and widespread enough to be a viable component of the content creation process. In 2011, Google’s “Panda” algorithm update aimed to help people find high-quality sites by reducing the rankings of low-quality content. In 2022, with their helpful content update, Google is simply attempting to do more of the same: by updating Search to prioritize original content, Google is trying to align the incentives of content creators to better match the needs of content consumers.

An important concept Google uses to identify high-quality content is “E-A-T”: expertise, authoritativeness, and trustworthiness. The exact definition of “expertise” can vary from topic to topic — for example, the standard for expertise on medical or financial topics is much higher than the standard for expertise required to relay one’s personal experiences — but generally speaking, high E-A-T content tends to be:

• Produced by people who are knowledgeable about a topic, as a result of having relevant life experiences, extensive research, or demonstrated formal credentials.
• Produced for a specific audience.
• Maintained and updated regularly to ensure factual accuracy.

Conversely, low E-A-T content can be:

• Created without adequate effort or relevant expertise.
• Irrelevant or unhelpful to its ostensible audience.
• Out-of-date, inaccurate, or fraudulent.

Importantly, E-A-T doesn’t just apply to the content, it also applies to the site on which the content lives. High E-A-T sites:

• Have a consistently trustworthy reputation.
• Present their content in a way that helps readers achieve a better understanding of the topic.
• Don’t overwhelm their content with distracting ads.

Whereas low E-A-T sites might:

• Have a negative reputation for being spammy, harmful, or deceptive.
• Present content with inaccurate or exaggerated titles
• Drown out their content with distracting ads.

If you are creating content targeted at specific audiences, providing net-new insights, and bringing real expertise, Google will rank your content more highly. In other words, the most sustainable and ethical way to ensure that Google identifies your content as “high-quality” is for your content to actually be high-quality — whether you use AI writing tools or not.

What AI can (and can’t) do

Like any other technology, AI writing tools are most effective when used by people who are already skilled writers. Skilled writers will always be needed to make the strategic decisions about what topics to write about, to come up with net-new thought leadership concepts or conduct original research, and to identify what makes a story compelling.

There are many aspects of the writing process that AI can make more efficient — like creating first drafts, rewording paragraphs to be more clear and concise, summarizing existing content, or iterating on ideas for titles, to name just a few. In particular, AI is very well-suited for predictable content types that follow repeatable formats, and for enforcing content consistency across platforms.

The foundational models on which AI writing tools are built, like the GPT models upon which Writer is built, draw from hundreds of billions of text tokens. By identifying patterns in their training data, these large language models (LLMs) use the power of prediction to generate original content from prompts. However, this means that what an LLM “knows” is limited by what it has access to: in particular, their default training data doesn’t include proprietary data, nor will it include very recent information.

Because of these blind spots, a writer using AI responsibly still needs to be able to bring their own expertise to the writing process: whether that’s by giving highly detailed prompts to the AI software, looking over an AI-generated first draft with a critical eye, or by training the model with their own relevant information and best-in-class content examples.

As content marketing leader Robert Rose says in his recent whitepaper, “The Co-Created Future of AI in Creative Content Strategy,” “AI is not wise. … AI is good at helping humans assemble original content, but it is not yet (or maybe ever) going to help you find the deeper emotional connection or concoct a great story.”

Can AI writing be detected?

A concern that some writers have is that search engines might be able to “detect” content written with AI, and that if their content is detected, then it will be penalized.

It’s not clear that identifying a piece of writing as “AI content” is, by itself, a meaningful metric for determining whether that writing is valuable and informative. A piece of content can be written with AI while still abiding by Google’s E-A-T guidelines; a piece of content can be entirely human-written while still being unoriginal and spammy.

Furthermore, the truth is that the distinction between “written by an AI” and “written by a human” isn’t always so straightforward. If AI generates the first draft of a piece, and then a human heavily edits it later, inserting new information and replacing automatically generated quotes, should that be considered “AI content”? What about a piece initially drafted by a human, and then edited with AI?

There are several output detectors that can relatively accurately assess the probability that a piece of writing was generated solely by a particular LLM, without any human editing or reworking. (Writer has one that you can try for yourself.) At a high level, these detectors work by evaluating how “predictable” the text is: in other words, how similar is this text to other texts on the same subject that are already out there? If you input the raw results from an out-of-the-box model into a detector, the detector is indeed likely to identify the text as being created by an AI.

Because these detection tools evaluate the predictability of the text, the best way for a writer to avoid having their writing being flagged as “AI content” is by ensuring that your text is, well, unpredictable — that you’re bringing real value to your readers by saying something that hasn’t already been said before many times elsewhere. Also, the output of an AI model uniquely trained on your own content will inherently be less predictable than the output of an out-of-the-box model.

In general, the more effort a writer puts into making sure that a piece of content provides unique value to its target audience, the more likely it is that the piece will “pass” a detection check, even if the starting point for said piece was an AI-generated first draft.

Here’s some examples of how even slight differences in initial inputs can make a major difference for AI content detectors. (The following samples were generated in October 2022, and are accurately rated as of then.)

Raw AI output with low-quality prompt:
0% human-generated

If you give a prompt that’s general and vague, and don’t provide any additional information, this can lead to AI output that a detector will read as predictable — and therefore, AI-generated.

Raw output with better prompt:
95% human-generated

Providing just a few additional details can dramatically improve the quality of the AI output. In this example, specifying that the output should provide three reasons, using the phrase “top priority,” and specifying the tone in which the output should be written was enough to bump the human-generated content score from 0% to 95%. Not all minor adjustments will lead to such dramatic results, especially for longer pieces of content, but this demonstrates the importance of putting thought and care into your inputs.

Edited output with better prompt:
100% human-generated

Using the output of the previous prompt, we can then manually add in some specific facts, statistics, or anecdotes to make the content more compelling and useful. This additional content is enough to raise the human-generated content score to 100%.

How writers should use AI writing platforms responsibly and effectively

One of the most powerful benefits of an AI writing platform is that it allows you to scale your content production, and speed up the process of going from first draft to publication. Unfortunately, if your organization’s standards for content are already low, then AI just helps you scale the creation of large quantities of low-quality content.

And with Google’s helpful content update, a strategy of quantity over quality just won’t work. In their explanation of how the update works, Google states that “Any content — not just unhelpful content — on sites determined to have relatively high amounts of unhelpful content overall is less likely to perform well in Search.”

If we start from the premise that your goal is to produce high-quality content, rather than attract lowest-common-denominator search traffic with overwhelming volume, then the question becomes: “How can my content team scale sustainably with AI, while ensuring that all our content is high quality?”

In a recent webinar, leading marketing and brand thought leader (and Writer investor) Jamie Barnett proposed some requirements for ensuring high-quality writing when using AI writing software:

• Put good writers in the driver’s seat: to deploy AI writing tools in the most effective and impactful ways, good writers should be empowered to make high-level strategic decisions regarding their use.
• Fine-tune the AI model for specificity: to surpass the limitations of an out-of-the-box model, you need to train your model on examples of your best content, so that it learns how to write like your best writers, and knows the information that your best writers know.
• Abide by your style guide: to consistently deliver valuable content with a distinct voice and point of view, AI writing tools should automatically incorporate your brand guidelines.
• Have guardrails for bias: just as human writers need to check themselves and their own biases to ensure that they’re communicating in a healthy and positive way and using inclusive language, AI models need to be able to adjust for underlying biases in their training data.
• Validate claims: to avoid publishing inaccurate content, AI should either be able to identify if a fact is true, or be able to flag where a fact needs validation from a human fact-checker.

The goal of producing content shouldn’t be trying to outsmart search engines: it should be providing value and meaning to your audience. The fact that search engines recognize this and are taking action to encourage people-first content is good news for readers — and will hopefully raise the bar for content quality across the board.