So much content: who's going to write it all?

The customer journey is getting longer and longer. And content is needed at every step. How can artificial intelligence help produce all these texts?

On which parts of the customer journey can you use large language models like GPT-3? Where are you better off using data-to-text solutions like AX Semantics? And how much human work do you still need, even if AI helps you write?

You'll get the answers to these questions in my keynote – and in the subsequent discussion with Manuela Frenzel and Saim Alkan. It starts at 17:22.

Highlights from input and discussion

24:06 · Make the subject-matter experts part of content creation. You have to tease their experiential knowledge out of them, they can't all write.

36:02 · With GPT-3, you can sit back as a copywriter and say: I'm a creative director and ten bad junior copywriters work for me.

36:02 · My role is shifting. From: I have to think of something on the white sheet. To: Someone fills the white sheet for me, and I just have to choose.

38:42 · Content-wise, GPT-3's output is often bullshit. This machine can text, but it can't think.

41:02 · GPT-3 is a wonderful sparring partner for subject-matter experts to evaluate: What is bullshit here and what is actually good?

The transcript – my contribution and the whole discussion

Arne [00:17:22] So much content. Who's going to write it all? I introduced myself as the copywriter who does away with himself. I don't want to write it. That's why happy to write it with you guys and happy to write it with machines, but actually happy to write it for years with other people together in companies that actually write it.

[00:17:55] "When humans and machines play well together, great opportunities emerge." When machines and humans play well together, great opportunities emerge, says Abraham Lincoln, right? Wait a minute, he probably doesn't say that. That's totally wrong, but somehow still believable. If I give a machine a brief, talk about how humans and machines interact, it might say something like that. If it's a statistical machine, like GPT-3 and other language models are, then the machine says what it learned in machine learning. That was the whole internet and Abraham Lincoln I think is the most quoted person. So who else is a quote supposed to come from if a machine mixes it together?

[00:18:41] So much for the machines in the team. But let's go back to a time when the copywriting team was very, very small. There's Don Draper, famous from Mad Men, sitting there thinking and thinking and thinking. He probably gets there relatively quickly and has about an idea for an ad like this. There's a headline needs and there's a picture idea on there. Besides, he's not just a copywriter, he's a creative director. The car is sexy, then there are four lines of copy and that's it. Everything Manuela talked about: the objections, the advantages, all that doesn't take place in the text at all, that takes place in a personal conversation. That's what the salesperson does with the customer. Notwithstanding the fact that there were probably one or two junior copywriters at Don Draper's agency who had to write a brochure, for example about the cargo space concept of this Buick. Don Draper himself would probably have found: I'm a diva, I don't do that.

[00:19:56] So we find, High Involvement, a car, which you don't just buy in passing, has always needed more information than just this little one-page ad. A couple of products have just recently gotten that need for information. In the past, a ketchup was just a ketchup. Today, you have to explain how much sugar isn't in it and whether it's really vegan. You all know the digital world has more touchpoints than ever for the customer journey.

[00:20:32] So who's writing this? I try not to write everything. If someone has to do it, they've probably had burnout for a long time. As a brand, you need more than one creative person. So let's build a team. I think that's very nice: today's survey, I didn't know this, but where do you guys belong? In marketing, in SEO, product data management, in the content team? These are exactly all the people where we say: where is the potential, how can you all become copywriters and where do you still need a copywriter?

[00:21:08] To do this, we first go into the customer journey. What can we expect there? There is a very, very simple model from the past. It's called Attention, Interest, Desire and Action, or AIDA for short. I've never really been clear about what the Interest and Desire part of it actually is. Attention is clear. That's what Manuela described as advertising text. And Action, that's what used to take place exclusively in the showroom in the Don Draper days. Today we do parts of it online and with sales texts. That's why it's worth playing with the model a little bit. To say okay, let's assume on white here, Attention, Interest, Dilligence, Action, AIDA still as an acronym. And assume we have rational customers here who are really looking closely at what the offer is. Or maybe not so rational Attention, Desire, Information, Action. In between, there's a lot of explaining to do and a lot of objections to avert, to discuss, to discuss benefits, as you heard Manuela say.

With the model, I'm now going on a journey to the text that has to be created for this. And then I can say, it's already become a hell of a lot. In the past, everything was done by the salesperson. And this moment of shock is something that everyone has at some point on their journey as a copywriter. And there's just the question: do I have it before and not even start because of the fright. Or do I think I've got it all figured out and realize it's just an incredible amount. So let's see, what do we have here?

[00:22:54] Attention: the advertisement, the poster. Known back in the Don Draper days. But today also a post from an influencer or lookalike ad on Facebook or Insta where someone doesn't even know they want us yet. But we times show: Hey, we also have something like what you saw at the competition.

[00:23:20] Desire: People have already expressed that they want something of that nature, they're looking for it. Then SEO content or even paid advertising and all the retargeting helps us. People have been with us before, but they've left. But maybe they do want to.

[00:23:41] Information: that's everything that took place in the showroom before. But also things that hadn't even taken place in the showroom, they just took place afterwards just in customer service.

[00:23:52] Action: used to be in the showroom with the vendor and now for us check out, add to cart, newsletter sign-ups. Those are the things where we really want action.

[00:24:06] Yeah, who writes all that? Everyone, if possible. My thesis would be: make the subject matter experts part of the content creation. The subject matter experts know their field. You have to tickle their experiential knowledge out of them, they can't all write. That's where the copywriter can help. It helps to make them feel that you own their content. You know it's common in content governance online to have them come back to you later to update their content, but you can actually do that from the first moment. From the first moment, make them part of your content. The copywriter is there, but they're really just overseeing the headlines and the calls-to-action, so attention at the beginning and action at the end. For that, you help prototype times and build blueprints, templates. You shouldn't actually write everything yourself, but the subject matter experts should be able to help out a lot. You can do that in part. In part, machines can help them. Then the landscape looks something like this. At Attention, it's still Don Draper, the diva or somehow maybe a little bit more service-minded copywriter, but maybe the influencer themselves. At Desire: do you know in the team SEO experts, paid experts so far already and still the copywriter. With Information, the Copywriter, who should be allowed to lean on, a little bit in the role of a journalist, what's happening with customer service reps, what sales reps know and the stories we can tell. But also what product managers and category managers know and say about their products. At the end, when it comes to action, again it's the copyright that gets the psychology it needs from the conversion rate optimizer to make it all really work.

[00:26:33] Now on the subject of machines, software, AI. This is to help us. I'm using three different genera here essentially. We'll talk about those today and a few more. Essentially content optimizers, like Frase or Market Muse, Clear Scope, Surfer. A couple of those also tout themselves as SEO tools. Then Language Models like the big, big hype topic GPT-3, made accessible with commercially viable variants like Jasper or Neuroflash. And rule-based AI, which is why we're here, like AX Semantics, which helps us scale. And lots more of transcription, so that we can get the stuff that vendors are saying verbally into writing right away, to proofreading, we're not talking about that today.

[00:27:31] Let's look at the three groups in the landscape. That's where Attention, where it's copywriting, GPT-3 with Jasper and Neuroflash can help them. In the short form, things come out quickly. A lot of junk comes out. Among other things, quotes like Abraham Lincoln come out. As a copywriter, you have to throw that away afterwards. In the area of what people are looking for, SEO experts will use keyword research, including possibly something like Ahrefs, but also Semrush or Sistrix or whatever you all use there as SEOs. There's like a new class on top of that, Frase, that already put that on there. Let's call that plagiarism engine, I'll show use cases later. In the information area, we also have the chance again to build things with Frase, especially in the FAQ area, to do headlines with Jasper. AX Semantics actually, as you know here, from the many, many posts that you see here and from the training materials, how to scale then reliably. On the action part of the track again GPT-3, so Jasper or Neuroflash, just to make very, very much stuff that you can then test with afterwards. The main thing is to see what works. And in order for you to know what works, you have to have tried it out. I'll show a few use cases from the landscape even before we're done.

[00:29:14] That's first on the line. Desire: Frase. Yes, to put it hardcore the fastest way to plagiarize the competition. Here's an example of me in German in the right column of this software you can see how many results have been searched here. I go in with one keyword. Normally get 20 results, here I have already hidden all shopping results, because I want to write long form at this moment. Then I just go paragraph by paragraph with an outline together. This is a finished text. It just doesn't belong yet. It's just not unique to me yet. Fortunately, the machine tells me where it comes from before I rewrite it. I can do that afterwards in Jasper, rewrite. Next case, Frase again. Same topic for the same keyword. FAQ. All questions actually written that others answer, I should answer. I can also evaluate Quora and Reddit. I personally have never gotten to FAQ faster than that. All SEOs among you know how much this can help you to be found. These are two use cases with Frase.

[00:30:40] I'm going to go to Jasper and/or Neuroflash for a moment, the GPT-3 based language models. That can't really do anything. That can, that just helps. Don Draper to come over the writer's block. Likes to be called, I've also heard Saim say, the most expensive parrot in the world. Can be very helpful in getting faster though. When I remember what we used to do in teams as junior copywriters on lists and lists and lists and afterwards the creative director would sort out. I can put a brief in here on the left, the block on AI writing and give a little brief with it, give five words. Jasper just writes me a list and every time I press Compose, there are five or ten new things. Afterwards I have a bunch of junk. But also one thing I really liked. Robocopy, that's what it's called on my website now., you'll find all the contact details there afterwards.

[00:31:53] AX Semantics, I don't need to tell you much about your garden. Saim and others have already done that thoroughly. For reliable scaling, where you just put brainpower into it as a copywriter beforehand, so that it works well in many products and in many languages afterwards.

[00:32:15] If you're ready to team up with machines, follow me for news, for deep dives into the individual use cases. There's already a little bit happening on Twitter and LinkedIn right now. YouTube, nothing is happening yet, but it's coming soon. If you want to know when, sign up for the newsletter, all at That was the copywriter doing away with himself and ending this presentation. Thanks.

Saim [00:32:50] Great! Very nice. Thank you so much. Yes, two copywriters approaching AI and communicating quite openly about where they use it. I just loved the examples. By the way, the Abraham Lincoln example moved me the most today. Imagine if you call in a quote and the machine, because Abraham Lincoln is quoted so often, assumes that virtually every quote that is new to the world in the world must automatically come from Abraham Lincoln because the machine evaluation says: statistically he has spoken the most quotes in the world. I find that quite great, this example, because it makes clear what the most expensive parrot in the world achieves. But I don't want him to go. I think it's a cool idea to be inspired, Arne. So to say, I'll see what the competition is doing. Or how do you see it? The other tools that you just mentioned, how do you classify that? Again, to elaborate a bit on the customer journey.

Arne [00:33:50] Yeah, you can go from the Customer Journey maybe into the Copywriter Journey or into the Content Creator or Writers Journey. When you're creating a lot of long-form, which is what you're doing with customers to some extent, the question is always, do we do an editorial plan? Then you can brainstorm what we actually want to talk about. So, what do we want? What are we good at? What do we want our customers to know? And the other thing is clearly SEO as market research. What keywords are being searched for? Which keywords have search volume? What is it worthwhile for us to create content about? Because we assume that it will be read afterwards. In my opinion, Frase is a wonderful method to make distance very quickly. Very quickly assembling longform articles that don't even belong to you, because that's pure plagiarism. But then you can go here as a copywriter and talk to the experts about whether that's actually true. They are much quicker to come up with the idea: No, say this instead. Instead of asking a product expert in a vacuum about what the most important things are. I think Frase is a great tool to have a bunch of questions yourself that you can ask the product expert. So I don't have to post them all as a FAQ. I can also just take them as questions into an interview and pester my service rep with them, so to speak. What would be his answer to this question? And then it becomes unique. Then I record the service representative in the interview, ask him these questions, transcribe them, then the answers, and the FAQ is ready.

[00:36:02] They all don't make a finished text. So I have to say now, AX Semantics also does the finished text only after you put a lot of brainpower into it beforehand and scales it up for you afterwards. These other approaches like Frase, are basically a great market research tool for interviewing experts in-house. Questions to customer service reps, outlines discuss with product experts. It's quite enchanting. With Jasper or anything else based on GPT-3, you can then sit back as a copywriter and say, I'm a creative director and ten bad junior copywriters are working for me. That means my role shifts. From: I have to come up with something on the white sheet, to, someone fills in the white sheet for me and I just have to pick. In the beginning, it felt very much like overwhelm to me. I once heard it was bad psychology to say in a brainstorming session: we need a good idea. You should say we need an idea and you should create a fault-tolerant culture where no one is too embarrassed to have a bunch of bad ideas so that afterwards we find the three for the shortlist out of the 200 bad ideas. In my opinion and experience, GPT-3 does that quite well. A lot of it is completely irrelevant. I think making subheadings in long form is what this thing does well. It somehow makes absolutely expectably boring subheadings, as you know it from machine translation. If you put in a good German text, you get a working English text, but it's boring because it's just been mainstreamed.

Saim [00:38:12] We have a question in the chat. Sorry Arne, I'm straddling it right now. As always, we don't name names from the chat, but I'll read out specifically, then you can classify it.

[00:38:21] Can you give examples where and how you use GPT-3 and how you assess the current quality? I think that reflects very, very strongly what many customers are currently asking. And I think we should also answer that very specifically. Where and how we use it was the question and how we assess the quality.

Arne [00:38:42] I'll start the question from the back, with the quality assessment. GPT-3 formulates knackered texts in English, on the basis of word choice and grammar considered. This is grammatically high-quality writing. In German, sometimes with unpredictable dropouts, but actually largely good too. I can't judge other languages. I suppose the more exotic the language, the worse it gets. That should follow from the nature of machine learning, that if the training was smaller, the quality decreases. That's the quality. In terms of content, that's bullshit. So this machine can text, but it can't think. There's a legendary example about this, GPT-3 is questioned about how to get a table that's too big through a door that's too small. And GPT-3 has, as one of many answers you have to credit the system, it knows that doesn't know anything, that's why there's a lot of choice, one of the possibilities is completely absurd. Turn the table top on edge and saw off the legs. This fulfills the requirement: the table should fit through the door that is too small. Completely pointless, but fun. And then you can think: How did the machine learn? It read the whole Internet. But possibly no one has ever covered this topic in text. Possibly this topic exists only videos or schematic drawings. Ikea instructions, little people spinning graphics around. How would this poor machine know? Possibly the only thing it has learned is an East Frisian joke. How many East Frisians does it take to get a table that's too big through a door that's too small? Five. One turns the table top on end and the other four saw off the legs. That's how GPT-3 might have learned it. Now we're getting right down to it. Pointless, but fun. That's why I think GPT-3 is a great – a miserable text tool. Inspiring in the better case because sometimes absurd, but a wonderful sparring partner for subject-matter experts who can evaluate what's bullshit here and what's actually good? If you remember the journalistic presentation form of an interview like this, you can conduct quite nice interviews with GPT-3. According to the motto: Dear GPT-3, what do you think about high-fat nutrition? Then comes a paragraph. Then you go and say: Okay, that sentence fits me. No, it doesn't fit me anymore, I'll throw it out. And then you ask the next question. You interview the machine, and sometimes you get good ideas.

Saim [00:42:11] I think you said it beautifully, this white sheet syndrome. That's a great inspiration. You can really throw it in there. A lot of our clients, yes, do product copy and you can put in some hard keywords and see what comes up as a result and then develop and derive anything reasonable from that. But I'm totally with you.

[00:42:29] By the way, in the chat I read also just here again very nice: pointless, but funny. So your quote just now was again taken here and confirmed. I thought that was a nice idea. Manuela, I would like to talk to you again about the topic of sales texts. And we've already discussed this very much on a meta-level in your presentation. What makes up the advertising text and what actually makes up the sales text? But if we break this down to product texts, what are your two or three concrete tips? So not on a meta level, but really quite concretely, what belongs in a good product text? What can we give our main users, product copywriters, that absolutely belongs on a product page in a scaling model like AX Semantics and thus in the product text and the sales text at the end, because they actually want to sell.

Manuela [00:43:24] Honesty.

Saim [00:43:27] But now I'm flattened, in one word slayed everything. So this high-quality piece of furniture just sucks when it's not high-quality. Oh, now I said such a bad word again and we have to cut that out of the video. So honesty I understand. If the piece of furniture isn't high-end, we shouldn't write high-end, right?

Manuela [00:43:50] Yes, for example. That's what I meant with the transparency earlier. I'm just going to go with a women's example. I order a blouse from a store. It's shown in the picture that it's not transparent. I get it delivered, put it on and you can see everything. So it should at least be mentioned in the text. Or I get a refrigerator delivered. It didn't say anything about it not being able to be in a kitchen because it's way too noisy, but maybe rather placed in a basement or should be. That's what I mean by honesty. I miss that in very many stores.

Saim [00:44:38] That's also a bit of the strategy that, for example, uses in their texts. They actually already have a paragraph where they include disadvantages. So they don't necessarily sell the product, but they say: Then you'd better click on two other sites where the product might be better suited to the requirements that you have. But besides honesty, what else is there? I think that's a great one, I also think that the point is to reduce returns. Even if it depresses the conversion, I think that a customer I advise honestly and say: this sofa bed is great for a nap, but nothing for three nights overnight. Or if someone stays with you for one night. But for permanent sleeping it is now rather not so suitable. What would be one or two other points where you say what absolutely belongs in it? Besides this strong statement of honesty.

Manuela [00:45:34] That's all I can think of right now. So maybe more adjectives. You're supposed to avoid adjectives, but there are adjectives that can describe products better, that make you want to buy them. So make the descriptions in such a way that you trigger a buying feeling. I have to have that or I want to be there for a seminar or whatever. But you meant product texts.

Saim [00:46:08] Emotionalizing is a very essential point.

Manuela [00:46:13] But I miss that in stores.

Saim [00:46:16] Absolutely. Adjectives create, they emotionalize, and so the thing goes into the shopping cart. And in case of doubt, my shopping cart is 100 euros bigger than it was before, just because I made an effort with the corresponding product description.

Manuela [00:46:28] All right, it has to fit too. It doesn't fit for information technology products. But for example, you should know how to pick up women and how to pick up men or how to pick up tech nerds? How do you pick up the target group? Be more responsive to the target group. But that's still difficult nowadays.

Saim [00:46:51] Yes, there are quite different ones. I just have another nice question or suggestion here in the chat. Bringing the subject matter experts into the text boat is seen as a great idea here. Their time is short. What's your tip for actually getting that done in a concrete way? And I'll be honest, I've had to do this for years on the intranet, getting subject matter experts on board for the intranet, they never have time to write it. I share the question. But maybe you have a really good idea of what can be done to get the subject matter experts on board.

Arne [00:47:32] Basically, it's the same way. If you set up a website relaunch from the marketing team, you want to have the subject matter experts, then they say, everything is already on the site so far and just do it. But then they are involved in the review process. That means that at some point they have to cut it off – sign off. I've always seen glimmers of hope that certain people, especially when they don't have to write, but when they're simply allowed to talk about it, that they'll then comment on what they've always missed about the previous content. Let them do a review about what was already there or let them get upset about it if they think, why don't we actually do what the competition does? According to the motto: We can do that too. Why don't we say that? Those are two starting points and then call, zoom, record, transcribe. And the text is ready. Of course, the text is not finished, but the copywriter has a briefing and that is already unique. Not like when you plagiarize it together with Frase.

Manuela [00:49:00] Saim, I see a question. Does the tool already exist in German? Yes, there is, but I don't want to advertise for someone else now.

Saim [00:49:08] You may do that at any time. We're happy, but this is not an advertisement for a competing product. I think that GPT-3 just occupies a very different space. But the question here is from Adela, I'm allowed to quote the name because it's an AX employee asking, are you working with the English version of GPT-3 or are there already tools for German? I think Adela just wanted to point out that you can also do it in German. Arne, which one do you use? Manuela, which one do you use?

Arne [00:49:38] I had access to a Scientific Beta of GPT-3. This works in German and in English. In English just beta – better. It is what I described in my quality assessment. Jasper as a commercial product that any of you could book has the capabilities in German and in English. They just offer a lot of templates in English. But if you give it enough before, they have a mode where the engine looks back, what did I write before in the paragraphs? And if there is enough German before, then it can continue German quite well. You can sort of force it with the previous input. Now I'm just testing, I can't really say anything about it yet, the guys from Neuroflash have convinced me that they can do German better than others. They actually achieve it the other way around, not by having a better input with more German, but by saying they just let GPT-3 deliver. Just brute force, more output, more output, more output and then put an AI-based filter over it afterwards that is culturally trained on German. I don't know how it works yet, but I'm curious. That's certainly what I'm going to test next.

Heike [00:51:02] I think we need to look at the clock, Saim.

Saim [00:51:05] Yeah, I'm looking. Can I elaborate on one more point? Please, with our participants, if you already have the chance to sit together with two such top-class discussion partners. If I now want to scale product descriptions, then from my point of view this only works with GPT-3 with incredible effort. I actually have to enter the keywords for each product, the key points, then fire up the thing, produce the text and copy it over manually. But now I have 3000 product texts to write, or 20,000, or like one of our biggest customers, 4 million a month. Again, very quickly to differentiate: Where does GPT-3 work from your point of view and where do I actually use AX Semantics? Not as anti-advertising at all. I think there's a huge place for both in these markets. But maybe we need to establish where it works and where it doesn't work.

Arne [00:51:56] You, Manuela?

Manuela [00:52:02] I say AX Semantics very clearly for product texts. Because you guys have an API, an interface. GPT-3 doesn't have that. GPT-3 is for inspiration. But for product copy, I wouldn't use GPT-3.

Arne [00:52:30] I definitely see in scaling. AX Semantics, when it comes to it being reliably scalable. You don't have to look at AX Semantics afterwards to see if it's true if you built good rules. If you've built good templates, then it's true afterwards, and then it's true afterwards not only in German, but also in English, French, whatever. Nevertheless, you can build template sets on the category level. Which you wire afterwards. Then again GPT-3 can take away the white sheet syndrome.

Manuela [00:53:07] Yeah, exactly.

Saim [00:53:08] Dear participants. I sleep much better tonight. You have clarified this so beautifully just now and also just once again set a very important highlight for me in this discussion group. Thank you very much for that. It's great that you were there. Thanks for sharing your presentations, which of course we share in turn. Heike, I'll give back to you. You still need to share the survey results.

Heike [00:53:28] Right now, I'd also like to say a very, very quick thank you to our translator from Simultando, who is still hanging on longer than expected. So we asked here: Where are you anchored organizationally? And most of them are now, oh wonder, anchored in the content team, 53 percent. But 6 percent were even in Product Data Management, 16 percent in Marketing, Content Team, which can sometimes be the same thing. SEO Team was 3 percent and then just others that were not listed here 22 percent. For all those who were now also a product data management team, next time it's about PIM systems, about what has to be considered there, in the application, what has to be considered accordingly for automation and digitization.

Heike [00:54:19] We'll take a nice side trip next time. Using data twice and three times in texts to generate text length by changing perspectives, I think is quite a nice contribution. Robert will grace us with it again. Next week in this respect definitely share the calendar and take and subscribe. Have there in any case all the dates for today. Let's say thank you very much.

Heike [00:54:43] Leave us feedback there.

Saim [00:54:45] We still ask for whole and feedback. Were your expectations met today? Were you satisfied? Were you dissatisfied? We will gladly accept that as well, in order to make it better for the next time. And otherwise, Heike and I say thank you very much for being here. Many thanks to our participants, who have eagerly discussed here. Thanks again to the translation team, Simultando, great job as always. We are glad that you support us. May I say, stay healthy and we will see you next Tuesday at the latest. Until then good bye and thank you, Arne and Manuela.

Similar Posts