Stay human-first but use AI for the dirty work

Andy says: “Use AI to help do the dirty work and the boring bits that you don’t really want to spend time doing – but don’t forget to be human first with everything you’re doing.”

What elements need to be done by a human?

“First and foremost, the fact-checking. A very dangerous aspect of AI is that it can easily make things up that are completely factually incorrect.

One example of that is when Google’s AI overviews were saying that you should eat rocks as part of a healthy diet. If you searched for, ‘How many rocks should I eat?’, it gave a nice AI overview saying, ‘You should eat several a day to improve your diet.’

You also need a human to attach your own style to things, and the tone of voice in particular. You might have a client where you need to hit the mark with that, or you may need to give it your own tone of voice, style, and interpretation.

Otherwise, everyone is going to be using the same style of writing. Obviously, you can use different prompts or different AI models to try and give it some more flavour, but there’s no better flavour or style than your own writing.”

Do you use AI to start with, and come up with concepts and ideas, before embellishing it using a real human?

“Absolutely, but it’s not just with content that you want to be using an AI model. You can also use them for the boring parts of SEO, like trying to think up 155 characters for a meta description.

If you have 100 new products to add to your website, that takes a lot of time. However, if you can invest in a model to give you a starting point, and then you only need to fact check and proof those, you’ve got a good time saving on your hand.”

How do you prompt AI to get that done to the best possible standard?

“You want to give it as much information as possible. At the moment, we assume it’s clever enough to figure it out for itself, but it’s not.

You need to think about what additional information you can give to it. Are there related products or similar things to what you want it to write about? Spoon-feed it.

If you can make that into an automated process, you won’t have to manually provide that information each time. You could put it in a spreadsheet, for example, that runs a script, takes that information, and feeds it directly to the model. Then, you’re saving time, but you’re also giving it substantial information.

You can get it to read a specific document and then write in a tone of voice in accordance with that document, mention something specific like a brand name, or whatever it might be that you want it to do. Give it as much information as possible.”

Have you been able to completely automate the process from the creation of the meta description to publishing that content?

“We’re not quite into the publishing side of things yet, but that’s on the list of to-do’s. If you can automate it, go for it.

There are plenty of very good Google Sheets plugins that work with ChatGPT or other models. You can basically create a big spreadsheet of all the data you want it to know about, then drag that down your spreadsheet. You could be doing 1 meta description, 100, 1,000, or 100,000; it doesn’t really matter.

On big e-commerce websites, where you have a product catalogue that changes every spring and summer, someone has to write meta descriptions for that. This would save you a lot of time. We’ve currently automated it from a spreadsheet point of view, but not from a CMS point of view quite yet. That would be nice.”

Can Google just read the page and use whatever meta description it wants anyway?

“There is that side to it too, of course. I still think you’ve got to hit the mark with your basics and don’t assume.

In the same way you shouldn’t assume that the AI models know what you’re talking about, don’t assume that Google is going to pick a good meta description for you, because sometimes it won’t. Sometimes it’ll take content from the footer or something weird.

Give it the information, first and foremost, and then it’s up to Google whether it wants to use that or use something else. If you keep it blank, you’re leaving it up to Google to decide.”

Should you also use AI to assist with redirect mapping?

“Yes, and for basically anything that physically takes a lot of time. For us, we do a lot of website migrations. In the last couple of years, we’ve had more migrations than we’ve seen in nearly 20 years in the industry.

Redirect mapping (literally changing from one URL to another), can be really time-intensive – particularly when someone’s changing the website entirely in terms of structure, changing content, adding a lot of content, or removing a lot of content. Someone often has to sift through the new website manually, but you can get a robot to do that.

Now, you can ask AI to give you its best guess and a percentage of how much it trusts that guess. It can come back with a measurement to say that it’s 100% sure a redirect is correct, or 50% sure, etc. Then you can decide that, for anything that’s 50% sure, you need a human to go and redirect map it. Whereas, anything above 95% accuracy is pretty much good to go.

You might still check those manually, but it depends on the size of the site and the investment you have to make to do it properly.

We’re trying to develop a model that will be able to check this accuracy in more detail. First, you want it to identify whether the URL is exactly the same as it was before, in which case it would be 100% sure it’s a match. If the title is the same as it was before, it can be 100% sure that’s a match. However, if there are slight changes in nuance or the page content is slightly different, you’re starting to get less accuracy.”

How do you upload your tone of voice to AI to check content for brand alignment?

“With ChatGPT, you can upload a document for your tone of voice, brand, or style guidelines. Then, it essentially becomes your chatbot with that tone of voice in mind.

It doesn’t need to be in a certain format, as long as it’s in text format in some way. It can be a Google Doc, a Word doc, or a PDF – it doesn’t really matter. It can interpret that. We’ve also uploaded 50 pages of a slide deck and had it read that as well.

Once it’s interpreted all that information, it will be like the arbiter of those brand guidelines. Then, when someone puts together content for you, that model can read the new content and judge whether it meets those guidelines. It’s a nice way of automatically giving you a belt and braces approach to QAing your own content.”

Is more content always better when it comes to brand guidelines, or can too much information be confusing?

“There’s got to be an element of keeping it simple and understandable. I sometimes see brand guidelines or tone-of-voice documents that are quite fluffy and dense, so it might need a bit of refinement before you upload it.

Just make sure it’s understandable. Don’t go overboard.”

What key elements should you include in a tone-of-voice guidelines document?

“Words to use and words not to use. Basically, words you like and words you don’t like – or the client likes and doesn’t like. That’s especially helpful if you’re producing a lot of content for a particular site.

If you’re moving across different copywriters or you have a big content writing team, you’ll sometimes have a slip-up where a word is used that you know from experience the client doesn’t like, but that copywriter might be in their first week and hasn’t got that experience yet.

You can use that AI model so that, if it spots this word, it will flag that. Obviously, you can do a Ctrl+F in your own manual checking, but this is an automated way of checking whether you have used any of the words you don’t want.”

Is ChatGPT your favourite AI software to use at the moment?

“I prefer it mainly because of the integrations that are available. It’s the most popular, so there are a lot of good extensions and Google Sheets plugins for it.

It’s easier to find integrations for ChatGPT whereas, if you go more off-piste, there might be other models that are just as good, but they don’t have the same integrations.”

Are there any tasks where using AI is not a good idea?

“In content production, which is pretty obvious for most people. We had the idea of using AI for e-commerce website category content that only needed 150-200 words. It was just boilerplate introduction content about what’s on the page.

However, by the time you’d got it all in place and had it proofed and fact-checked, somebody could have just written it. It was quicker with a human. We also found that it ends up using curious words that a human wouldn’t use when describing certain products, which sounded unnatural.

You can give it a document to try and make sure it doesn’t use those words, which would be a second version of trying to create that automation. However, you still need to have that fact-checking in place. You can’t send something out that hasn’t been fact-checked. We found that, for that 150–200-word content, it’s actually quicker to get somebody to do it with their proper knowledge of the site.”

Is this something that you will continue to test, to see if AI improves enough to take over that task?

“If we can get it to a point where you can remove those oddities, I suppose it would become a time saver.

At Reflect Digital, we are part of the Human First Collective, so we always want to fact-check things. We never send something out the door without it being fact-checked, so it would have to be pretty bulletproof for it to not be subject to quite a bit of human intervention from our side.

The Human First Collective brings together Reflect Digital (our SEO, paid, and digital PR agency) and Lab, which is a design and development agency. We are able to do everything effectively as two separate agencies, whether you need a new website or marketing for that site – and everything in between.

We aspire to approach everything that we do by putting humans first. Our USP is that we are using behavioural messaging with all of our work. When Lab is building a new website, they’re using behavioural insights before putting a design or navigation together. Everything has that human element in mind.”

If an SEO is struggling for time, what should they stop doing right now so they can spend more time doing what you suggest in 2025?

“Focus on the time sinks; the things that are costing you time right now. Don’t push them to one side. Really consider what’s worth the investment of your time.

If you’re finding yourself doing repetitive tasks where, every month, you do the same report and it takes you 3 hours, how can you automate that process and get that time back? Invest in efficiencies and automation because you have the technology available to you now – whether it’s ChatGPT or Sheets connected to whatever AI you want to use.

You’ve got the technology available to you to automate those tasks and make your life easier in the long run. Prioritise those efficiencies.

To determine whether a task is a good use of time, firstly, is it boring? Is it repetitive? Is it something that you can automate by creating something that will do the job for you? If there are rules or principles that can be repeated and given to a computer or any sort of script, can you automate that process and get something else to do the dirty work for you?”

Andy Mollison is Head of SEO at Reflect Digital, and you can find him over at ReflectDigital.co.uk.