Build this LLM workflow to spellcheck your entire website

Here’s a Slack message that never stops making my blood run cold, no matter how many times I receive it:

👋 I found a typo on this blog post you wrote.

P.S. Have you noticed that several product pages have outdated images?

Scanning an entire website for copy errors used to be impossible. With LLMs now it’s practically automatic. And the workflow you set up for spellchecking can also be applied to a lot more than typos. (More on that below.)

First, let’s look at a simple LLM workflow you could build on your own.

5 steps: build an AI workflow for content quality assurance

No matter which tool you use, the basic workflow is the same:

1. Pick a workflow platform. Or, if you don’t have a platform, you can do these steps one by one in GPT.

2. Identify the URLs that you want to scrape on your site. It might be every page, it might just be pages on your blog, or product pages, etc.

3. Set up a website scraping node. This will grab the text from every page so that it can be checked for eros.

4. Run each page through an AI parsing node. You set the rules here for what the AI is checking. Is it looking for spelling errors? Broken links? Outdated brand terminology?

5. Output the results to a spreadsheet. The spreadsheet should flag the pages that have errors as well as the specific errors on the flagged pages.

Again: use whichever tools you please. Different tools work better for different teams, budgets, and technical skillsets. For the purposes of this blog post, I’ll be sharing a workflow that I built on a platform called n8n.

Example: a spellcheck workflow built in n8n

Here’s a video walkthrough I did, if that’s more your speed 👇

If you've ever used HubSpot, you're already familiar with n8n’s drag-and-drop style of workflow interface.

All the bits and pieces here fit together and have the little arrows and do cool stuff.

Here’s a simple example workflow:

1. Identify a group of URLs from our website.
2. Scrape each page on the website to get its text content.
3. Send that text to GPT.
4. Prompt GPT to identify and flag any copy errors that we define.

It will output a spreadsheet containing the flagged pages and specific errors within.

Next, the workflow surveys the site and returns this spreadsheet. It’s a list of blog articles.

Note: for this exercise, I’m scraping Clay’s website. I love their product and I think their marketing team is doing awesome work… So they can withstand a teeny-tiny bit of scrutiny like this.

LLMs can run diagnostics on the quality of your content

We can run this little piece of our content system anytime we want on however many blogs we have. We've done it for clients who have thousands of pages on their site. And because we're in this new era of LLMs, we can define all kinds of crazy tests that scan for elements of content.

Previously, we could only audit technical elements of a website:

• Is there a broken image?
• Is there a 404 link?

Stuff like that, which is important but doesn’t quite apply to the content and messaging side of marketing.

Now, however, we can run automated tests that do apply to questions of content.

👆Like the workflow I’m running here: “Are there any misspellings in this article? Start your answer with Yes or No....”

The workflow returns that information to us for every single page on a site.

Sidebar: the possibilities go way beyond spellcheck

Let’s take a step back here and admire the possibilities for this simple but powerful workflow engine. We can use it to test all sorts of questions about content quality, style, and messaging:

• Does this fit our brand guidelines?
• Does this answer all the questions that people have?
• Does this include our latest pricing that we just published yesterday?
• Does this match our new messaging?
• Is this duplicative of some other thing we have on the blog?

This is quality assurance on a scale that we’ve never been able to manage before. I think that’s pretty cool.

Output from the spellcheck workflow

After three minutes, the LLM flow returned the spreadsheet containing a list of relevant URLs. Each one has been scanned and analyzed for misspelled words.

Here’s an example of the simple punctuation errors spotted by the workflow:

• “Overelying” should be “over-relying”

Pretty articulate analysis there, sourced from GPT. And if we look on the page to verify the results, there it is:

This is a tiny example, of course. Probably nobody's ever noticed those misspellings.

My point is: have a machine proofread your website for errors like this – and all the other tiny but important details that go into maintaining a content library.

As always, human oversight is required

And now, the mandatory AI caveat: not all of the suggestions made by the LLM will be great.

Some of the errors flagged by the LLM output will be overly cautious. Stuff like: "This word got used in a way that wasn't expected.” A decent note, but not what we’re actually looking for. So we’ll have to keep that in mind when we refine the workflow for the next version.

Content systems like this are influenced by devops

If you work in the software development world, you might be familiar with the concept of linting. That’s basically what this workflow is: a linter. It's kind of going through and pulling out errors from your work (like a lint-roller pulls fuzz from your sweater).

Developers use this sort of thing for code when they deploy it. Workflows like this one extend that programmatic ability to marketers.

We are moving toward a time and a place where we can treat content more like data.

We can parse it. We can move it around. We can use computers to help us understand and maintain it.

Let’s all start thinking in terms of content systems

This little LLM workflow is a tiny system. It fits into my broader content system.

I think every marketer should be thinking of content in terms of systems. The system requires inputs: data about our brand’s product, point of view, style, etc. We run those inputs through a variety of routines and tools (like this LLM workflow). With human oversight, they generate a consistent output: increased page views, social clicks, pipeline conversions, etc.

Producing content is hard!  When you think about your content as a system – a set of human-led and machine-assisted workflows – you can produce a wider range of material at a higher velocity with better results.

Let a machine proofread your website for errors. Let it check to make sure that all pricing on your site is current and accurate. Make it a standard part of your content system and spend more time on more creative challenges.