Dhaval:
This founder built an AI copywriting product and got over a hundred thousand users in two years. In this episode, we talk about his approach to deciding where to invest your time, money, and energy, as an early stage AI founder and I learn a lot about his approach towards prompt engineering. Yaniv is the CEO and co-founder of Anyword leading AI copywriting solution designed for marketing performance. He has a Master’s in Computer Science and Information Systems, and he has conducted extensive research in the fields of machine learning and natural language processing. His work optimizing ad and content channels for some of the largest publishers like New York Times, Lead him to found Anyword generative AI platform. Yaniv oversees operations across Anyword’s New York, Aviv and satellite offices across the world.
Dhaval:
Welcome to the show. Yaniv. Tell us about your product.
Yaniv Makover:
Hi. Thank you for having me. Yeah, I’m co-founder & CEO of Anyword. We are in the coparating AI space. Our product primarily focus design on making copy more effective and converting more and engaging more.
Dhaval:
Wow. Okay. So Anyword and. how is it different than the plethora of other copywriting tools, AI-enhanced copywriting tools that are out there, at this time, or they’re probably gonna come out now that there’s a lot of those capabilities? Yeah. Tell me about how is it different. What’s the differentiation there?
Yaniv Makover:
So we started out from becoming performance mindset. And I think there’s one. Big problem that generative models can solve, or language, large language models can solve is basically helping you just get more content out there and high quality content removing solving for writer’s block. And I think that’s an interesting, huge problem to solve. For us, it’s kinda like not our DNA. Our DNA is more about in our products DNA is how to make your. Copy better. So you already are a marketer. You already know what you’re doing. You have a strategy. Yes, you could get ideas from ai, but this is how Anyword will work better for you for a specific audience. Or you’re selling, I don’t know, a sweater to somebody in the US versus other countries or a different occupation or different age or different gender. Then how, what are the best words to use for every use case and I think we use large language models to actually empower those insights or actually leverage those insights, to create ROI for our customers. I also think that when you’re just a click of a button away from creating hundreds of variations of copy we thought it was a really big problem to solve. Which one are you, you’re gonna publish, you can’t really A/B test 1000 tweets you have to send one. And so we thought that was like the biggest problem solved. So we, early on, we focused on that. And our product, we pretty much tell you if with every copy variation that the AI generates, how will it will work and for whom and why. And if you wanna make improvements, how to do them.
Dhaval:
Cool. Okay. So you’re taking the existing. LLMs and you are not only generating content, that’s an easy problem to solve, but what you’re really doing is you are helping fine-tune that content to resonate with the specific audience that your customers may have, and then fine tune the copy to increase engagement or retention with that audience is what is the primary metric that you aim for with your customers to improve? What is their North Star metric that you’re helping them?
Yaniv Makover:
So typically it depends on the use cases of the marketing, but they’ll, they’ll measure lift in conversion rate or lift in engagement, and then they’ll measure just ROI so if they’re running ads, they’ll, they should be able to see a lift in their ROI or if they’re, the conversion rate on the lending page or open rates and emails. And it’s pretty easy to measure easily. Just copy and see if see if it works for you.
Dhaval:
So do you, how do you get the information on their audience, like to fine tune the output with that highly fine-tuned output? Yeah,
Yaniv Makover:
so Anyword collected its own data, and basically, we have our pretty large corpus of data, performance data. And also when we, we partner with our customers, our partners, basically they have their own data sets, and then they upload in them into Anyword, and then we have, we fine-tune what we call custom models to help them predict better how their copy will do. So, for instance, just based on our data, we have an accuracy measurement of how well our model predicts performance of like copy that we already knew how well work. Somewhere around 76% depending on, on what we’re testing. But if you, bring in your own data and you’ve actually A/B tested or just ran a copy in the past, then it goes up to 85. And that’s just because you have your own audience, kinda like your own topics. I think for me, one of the most interesting parts of the space of large language models, not only they can write really well, they also understand text really well. So like five years ago, if you train a. Just lots of text and tell it, this text is good, this text is bad. We’ll probably figure out that one has an emoji or an exclamation mark. But now there’s a deep understanding of why text works. Like are you using, if you’re missing out is that even relevant for some audiences and for some products or industries? And I think that’s super exciting. So I think it wasn’t possible a few years ago. And it’s possible now. And I see this as kind like a. Booster and performance for marketing.
Dhaval:
Hundred percent. There has been a plethora of content generators using chatGPT and tools like that to create content.
Dhaval:
But what is still missing is the ability to fine tune that output to the specific segment of your audience and then be able to create content based on that readily, readily, as in with a click off a button. I’m sure you can stitch together a few data pipelines. Do that with existing tool suites. But what I don’t see happening is the ability to just readily click of a button, say, this is my audience, this is my content. Generate some copies. Is that, am I understanding your pro product correctly? You offer that?
Yaniv Makover:
Yes. For every, like, while you’re typing, you can even not, you can use your own copy, not even generate with ai. You’ll have insights, analytics about that. Copy how well we’ll do for your target. The way you defined it, what talking points work better and maybe replace them with others? And you can use your talking points that work for you while you’re running ads in your emails and the talking points that work in your emails or the insights you gain from that and your landing page. And I feel like that is kind the future where there’s just so much content that’s gonna come out. You really know you have to know what, what resonates with your audience.
Dhaval:
Yeah, I imagine a world where the specific copy will be highly personalized to the reader that is consuming that information. And it could be hundreds and millions of variations of that depending on who’s the reader, right? So what I am curious about is when did you. Now that I have the context of your product, let’s talk a little bit about your business. Tell me about when you founded the company, and, tell me a little bit about the number of users you have amount of your revenue, et cetera. Anything you can share to provide us context on your business?
Yaniv Makover:
So Anyword launched March 21. And basically, it was a spinoff of the first company we founded, which called QE QE it’s a SaaS platform for publishers, New York Times, CNN Washington Post helps them. Distribute their articles and their content on, on, on social platforms like Facebook. And then what we figured out there and QE works with 70% of the top media companies in the US What we saw is that some of our customers did way better than others in engagement and conversion rate, just based on their copy. And we thought, okay, we can help our other customers to, to write better copy. And then that was kind of the. What made us think about Anyword? And then the problem was much bigger. Not just copy for social posts, but you can even rewrite today a whole article for five different audiences. So you might be reading a different version of that same article, then me and the author. It’s the same the same author with the same idea or message. But we’ll be reading different articles that talk to us based on our. Words and familiarity and language. And I feel like that’s super cool. I think I always thought that was like, a future and I think it’s very possible today. I think something that, pretty profound.
Dhaval:
Yeah. Yeah. That, I think that’s pretty cool, actually, to be able to have such a level of personalization for all the segments of your audience. What I’m curious about is, Do you have, is this product out already? Do you have any customers using? I know your previous product was mega successful. 70% of the media companies using a product. That’s amazing. Congrats on that. Success, QE QE, right? Now with this that you launched in March, 2021, Anyword that was launched in March 2021, what are some of the business results that you could share with us.
Yaniv Makover:
So we have almost a hundred thousand, users using the product. And we have a few thousand paying customers. And we’ve the product is like the our top line is is now at an like, exponential rate growth. So chatGPT was a big help Introducing the whole economy, this space. And I think, as far as product market fit, I think there’s a hunger for helping, in copywriting, specifically for performance copywriting. And super happy to be in this space and yeah, to grow with it.
Dhaval:
Yeah. How many to? What’s your pricing point? Are you are you enterprise? Do you have enterprise customers, individual copywriters, or freelancers? What is the market segment that you aim to grow and expand?
Yaniv Makover:
So we have the comic starter offering. It starts at $29 a month for consumers pursuers. And then, like half our revenue comes from enterprise, where we are API powers ad vendors, agencies, big brands that have a lot of data and, trying to leverage that data to create performance. So I wouldn’t say their biggest issue is to create more content or copy. Their biggest issue is what we’re basically helping them do is this is the easiest lift they have in their, their easiest 10 or 15% lift in their marketing funnel that they’ll get just by improving the copy. I feel like it’s a must happen. I think every marketing org will have a box like this in their stack, to improve their copy.
Dhaval:
Hundred percent. Yeah. It’s gonna be part, it’s gonna be part of every single stack, all the way from creation to publishing. Right. I totally see it like becoming an ingrained part of copy creation and copy publishing and all of that. Yeah, now tell me a little bit about how you stumbled on the product idea. I know you. Talking about how you had media company and a company that was used by a lot of a product that was used by a lot of media companies, 70% of the media companies, and they had challenges with the conversion rate on you noticed that you could improve that? Is that how? Tell us a little bit about your origin story and tell me also about whether you are a technical founder if you have another technical co-founder.
Yaniv Makover:
Yeah, so I’m a technical founder, and I also did my master’s degree in national Honors processing. And my co-founder is also technical. So we’re both of us are technical in that aspect then, and even before large language models like that makes I really believed in in, in our ability to improve performance of copy. So when I saw that, some publishers were doing better than others and not necessarily the ones that have. Had a bigger budget or bigger teams or adjusting what they were doing, and they were A/B testing more copy they’re doing way better as far as engagement and conversion subscriptions and the, and sales than the other ones. I thought this was, like, something we can help with AI early on we are just struggling to take like an existing social post or a tweet from a publisher and just replace two. To have it make sense and perform better. And as models became larger and foundational models became larger, we could leverage that to solve all those contexts issues and, that kind of really played in our favor.
Dhaval:
Wonderful. Tell me about how you created the product. What is your stack like? What how long it took you to create the product? What is the size of your engineering team? The reason I’m asking that is because our audience are people who are either interested in creating an AI product or have. How to Infuse AI in their existing stack. They work for an enterprise, and they want to do that. So give us an estimate of like the budget or rather, the size of your team and the effort it took you to build this product. Yeah.
Yaniv Makover:
So there’s around 25 people in the economic engineering org. I think there’s a big difference between leveraging. A foundational language model likes GPT 3 or building and fine tuning your own models. So which like GPT 3 or others like it, you can build really quickly. Like it’s,you don’t need data no need for fine tuning and you can just focus on workflows and user experience. If you have your own model, then it’s way slower. You have to there’s a process for obviously training it, but also, you know, getting to the right performance, and that could take months. So I think when you, that’s like a big decision. We invested early on in a big, like a, in comparison to our team, a large AI team. So there’s five people in the AI team, and they, we train models, not not only. Write prompts. So prompts are much easier to write. And I think about decisions like where you need to invest in your own models and when you don’t. I think you can iterate really fast on user experience, product market fit and get initial solutions going with foundational models. And you really need to choose your battles where you absolutely need, if you actually need one your own fine tune model when you want to host it, because I think that’s Much bigger investment. What we found to what worked for us is yeah, you can iterate really fast with just prompt engineering and see and get to initial result and then go deep, so to speak is, to build like your own models to, to solve some specific problem, for instance, and when you have to predict what copy will work better for others or if it’s more. in line with a tone of voice that the customer’s looking for, then we had to really fine tune our own models and that’s an investment but you, when you just want to create a new format for framework for copy, then yeah, just do it with a prompt and you’re good to go.
Dhaval:
Yeah. So for a product creator who’s interested in creating an AI product, there are three areas. She could be investing in. One is learning the foundational models and iterating use cases and workflows. On top of that, essentially a UX and a distribution advantage. This is the first use case. The second use case is the building, fine-tuning and personalizing the foundational model itself. And that could be a lot of investment. That’s not just a distribution or a UX advantage. That’s a technical defensibility of the product itself by iterating and improving the foundational model itself. And then the third, the found, and then the third area that she could invest in is prompt engineering on top of fine-tune models or existing LLMs. And that itself could be that’s a, that’s an advantage in terms of problem discovery.
Dhaval:
How did you improve your prompt engineering skills? Looks like you fine-tuned it to a point where you have hundreds of thousands of users. What advice do you have for someone who is trying to focus on that category of improving their prompts?
Yaniv Makover:
We invested in a feedback loop really early on. So we actually connect to marketing channels. Our customers marketing channels are Google ads, Facebook ads, API their website. And then we really care about performance. If you know what you’re looking for, you can measure and then and then figure out what’s wrong with your prompts. If they’re not creating the right outputs. We also have these. Quality surveys and these mechanisms built into product where they can. You can save like a variation you like. And then we use that as also as part of our feedback loop and, and then you try and make mistakes and fix them. That’s basically what we did
Dhaval:
What was the can you unpack the first option that said you have APIs from Google ads and Facebook ads? How would that improve the prompt?
Yaniv Makover:
If you create a copy with Anyword for your Facebook ads and you run those ads, we will look at the performance of that copy, compare it with your other copy on your ads and then see if it did well or not, and then generalize that into some insights about, okay this is not working well. I’ll give you an example. So in a Google ad, you can talk about your brands depending on keywords and campaigns. And some brands have really well known really well known. It’s it makes total sense for them to actually incorporate way more than that way way more of their brand in their ads and others. It’s just all benefits and features of the product and not necessarily the brands, but that you really have to define, let’s say, the product itself. So for instance, Anyword, if we had to write an ad for Anyword, we’d have to say AI writer at some point because we’re not Nike but for Nike, like I just, the brand itself. So those understandings have to go back into the prompt. And then that’s something, you know, from performance and what people like to use.
Dhaval:
So what I’m gleaning from this is for improving the prompts. Feedback loop is critical. Being able to have that feedback from the market, and from the users, helps you improve those prompts. Now that I say it in the hint side, it sounds very obvious, but thank you for helping us glean that. Yeah. Any other advice you have for product creators? Who are either embarking on this journey now that there is, has been, there has been this new technology or anyone who has been doing it for many years and just wanna make that leap. What advice do you have for product creators in this space?
Yaniv Makover:
I think it’s super exciting, like exciting times about what’s possible at the same time. And I think they’re going to be many winners. And big companies are gonna disrupt like existing workflows and industries, I think thinking about specifically what problem you’re solving and is it can it what is the incumbent gonna do. So this is a feature in another product. I think that’s not easy. I think. I think it I think for different problems, there’s different answers. I’d really focus on that. And also I think. Temptation because of the pace of, how the technology is evolving to, to do more and to cover more ground. Oh, let’s go into images and let’s go into, and you really have to be disciplined because my guess is that if you’re not specific enough and narrow enough in your focus, over time, you’ll have you won’t have the best product. You have kinda like a, and I think it’s always a challenge in product management, but I think now it’s even more with ai. Specifically, you have to get your user experience right. And and it’s very , tempting to just build lots of stuff really fast
Dhaval:
thank you, Yaniv. It was a pleasure having you on the show. Thank you so much for making time for us. Thank you for having me.
