Mack Grenfell on Byword.ai and AI Content Generation

[00:00] Matt Hepburn: In this episode, Mike Goldstein from Kitchen Remodeling SEO joins me as a guest co host as we talk with Mac Grenfell, the co founder of Byword AI, which is a platform where you can generate articles through AI. Please join us. Welcome to the Entrepreneurs Marketing Journey Podcast where we help experienced coaches, course creators and consultants who are motivated to increase their revenue by implementing market tips and strategies. Hey, it's Matt Hepburn. I'm a digital marketing professional with 14 years of experience working as a consultant, working agencies both large and small, and for the past eight years in the enterprise sector for some of the biggest brands out there. We provide the latest digital marketing tips for coaches, course creators and consultants so that they can grow their businesses bottom line across marketing channels. Hey there, Mac. Welcome to the show.

[00:51] Mack Grenfell: Hey, thanks for having me, Matt.

[00:53] Matt Hepburn: Absolutely. Really excited to hear more about AI and content generation. I was hoping you could tell the audience a little bit about yourself and your company and then we could go into the questions after that.

[01:05] Mack Grenfell: Yeah, sure thing. So I'm Mac. I've worked for the past couple of years now, building initially all sorts of kind of custom bits of tech around AI and SEO. So in the first place, this was kind of building one off pieces of tech for brands to kind of generate lots of SEO pages at scale. And then later on last year transitioned into a productized version of this, which now exists as a product called Byword AI and is a platform that kind of grew out of what I was doing as a consultant. So lets brands come on and generate lots of SEO ready content with kind of all the bells and whistles and integrated with your site so you can kind of take advantage of what AI offers in terms of being able to build SEO content at awesome.

[01:47] Matt Hepburn: Well, I'm going to also talk to Mike here for a second. Mike Goldstein from. Mike and I work on our main gig together and then he also has a side company. Mike, could you talk a little bit about your company because you're the guest host today, and then I'll let you start with the questions. How about that?

[02:04] Mike Goldstein: Sure. Absolutely. So my name is Mike Goldstein. I'm the founder of a digital marketing agency that specializes in working with Kitchen bath cabinetry providers, essentially people in the home services space. And we provide local SEO use of AI and automation chat, GPT, go, high level integration things to that extent, as well as running Facebook and Google Ads. So I've been doing that for about over 20 years now. I like to tell people I've been doing SEO longer than Google has because we started back in 1998, and very first client was this little unknown company called Ink Tommy that no one's ever heard of, but everyone knows who they became. About four months after we worked with them, they were acquired by this little unknown startup out in Silicon Valley called Yahoo. So when I say we've been doing SEO for a while, we've been doing it for a while. Fantastic. Thank you for having me on the show, Matt. This is fantastic.

[03:02] Matt Hepburn: Absolutely. Did you want to ask Mac some questions? You want to start with the first question?

[03:07] Mike Goldstein: I do. One of the things that is really big in SEO right now and has been so for the better part of the last few years is this thing called Eat E-A-T. Right. And expertise and experience are big factors of that. One of the things that I know that AI is really great at is creating a ton of content and demonstrating an expertise in that content. But the one thing that I've always found that AI kind of lacks is how do you add that personal touch? How do you add the experience that Google so recently has asked us to kind of add into the profile? So what I want to kind of find out is really how does your software go about adding that personal touch to the content creation to avoid the whole thin content issue?

[03:53] Mack Grenfell: Yeah, sure thing. I think there's a few different ways to take this. So in one sense, I think it's fair to say that there's no kind of AI writer that really has this solved and if someone tells you otherwise, they're probably not being entirely truthful. So the way that I've kind of advocated for using this in the past, particularly in the sort of yMYL content areas where it is more important is not really seeing AI as the kind of finished product, but using it as more sowing the seeds in a way. So, for example, I've kind of worked with a lot of brands where we'll say, okay, we'll kind of generate these thousand articles or so, and this isn't the end of the road. We're just going to put this on the site, see how things rank. And then rather than starting from scratch and writing 1000 articles by hand, or if you're human, we can go back in three, six months and say, okay, these terms or these pages are ranking. This is the time where we go back and kind of use the ranking so far as a signal for how we prioritize the human editing, which is where you come in with that expertise and authority and so on. So that lets you get it sort of lets you take advantage of AI as a starting point, but you kind of layer on that human element once you can have the signal of, okay, this is where we're actually going to get bang for our buck if we invest the human resource in. So I think that's kind of the biggest and most honest answer, really.

[05:08] Mike Goldstein: I love that, because the one thing you can always say about AI is it might be smarter than any human being out there, but it can't have an experience that a human can. So taking it as the starting point and putting it out there and then adding in your personal touch really does make a big difference. And the next thing that comes to my mind is, as someone who owns a digital marketing agency, I'm in a niche and I work with the same type of contractors, kitchen remodelers, bathroom modelers, cabinet builders, people like that. Oftentimes we're writing about the same topics over and over and over again. And although we have a little bit of an expertise in that area, how do you go about, or how does your software go about really taking a topic that you may be writing for five, six, seven different people and kind of differentiating it up enough so that it's fresh and it's not really a duplicate, it's not just cookie cutter content?

[06:06] Mack Grenfell: Yeah, that's a good question. And again, I think there's a few different ways to look at this. So one, on a kind of purely mechanistic, technical, mechanical level, there's a lot of different stages to kind of what happens behind the scenes of my word when you generate an article. And each one of those stages has a lot of randomness built in. So even if you kind of provide the same inputs and all the same kind of article settings, you'll get quite different articles coming out. And at a more fundamental level, if you look at how language models work, you're kind of breaking down an article, let's say a 2000 word article, into three and a half thousand tokens or something. And each token as it's being generated, the kind of next token or word that's being generated is really a function of all the ones that have come before. And so even by the time you've generated your 1st 50 tokens, your 1st 40 or so words in the article, you've run a number of quite highly random calculations where the chance of two articles diverging or deviating is very high. And as soon as those deviate within the first 50 25 tokens or so on, because each next token is a function of the ones that come before, they'll tend to diverge more and more. And there's another half of this answer, but even at a purely kind of mechanistic level, there's no real risk of duplicate content even from the same tool, or at least in BioWord's case. But the other side to the answer is kind of in terms of the roadmap and product development and features that we've put in. And I think a big trend over since Biodiver launched has been given more control to users and giving them more options for how they tailor their articles. So it's not just kind of basic things like tone of voice, but also giving the users the ability to kind of inject structure and stuff into the prompts, to direct bywords writing in a certain way. There's a few things that we're looking at doing in future. There's actually a sort of. I can't even call this a beta feature. It's sort of a hidden feature that we have that all the kind of infrastructure is built and we're just waiting for the model to come, where it will take either a kind of scan of your site or a few pre uploaded sort of content samples or writing style samples, and then be able to inject those into future content. And it's a kind of funny one, because we've seen in chat GBT and the user interface, we've had really good results with the sort of same prompt structure that we're trying to build in the back end. There's a little bit of a difference in terms of the models that are available in chat GPT and what we can build in production through the API. But as soon as there's parity on those, we know it's going to be a really good feature where you can almost perfectly kind of mirror the sort of style that someone's taken in their article so far and be able to transfer that through to the articles they're generating and give it even more diversity.

[08:38] Matt Hepburn: Yeah, I'm going to hop in there on that. That's exciting. So I think what we're starting to see is with some software. I think one of the recent guests I had was talking about lately, AI, how it uses all your own content, versus Google's 2021 scrape. That chat GPT did.

[09:03] Mack Grenfell: Right.

[09:03] Matt Hepburn: So it sounds like what you guys are moving towards, and this would be fantastic, a model where we could pull all the content from our site or uploads. Maybe it's something that's not publicly available, like PDFs and things like that, where we can get deeper content. Is that the direction it sounds like that's where it's going. I'm hoping that's where it's going.

[09:27] Mack Grenfell: Yeah, I've seen a few people implement similar sorts of things. So far, I haven't seen one that does it kind of in a way that I would feel is high enough quality to put into production on Byword. Sure. I think we're pretty close to the point where the models are there and we can have that as a future. We have some similar features right now, which, well, they look very different to a user, but infrastructure wise they're actually very similar. So there's one where a user goes and uploads a sitemap of their site and byword essentially crawls every single page and indexes and understands it. And we use this currently for interlinking to be able to kind of insert links through to the user's site in new articles. But there's also a very similar application where we just use the same sitemaps and the same vector embeddings, which is a fancy way of saying by words understanding of the articles to allow us every time a user generates an article to look for similar ones, to look how the user's written that before, and to be able to not just copy, but be able to take certain stylistic elements and tone of voice from the article.

[10:23] Mike Goldstein: That's really interesting when you're talking about kind of looking at sitemaps and looking, one of the things that I love to do is looking at sitemaps and using AI to find internal link opportunities, things of that, you know, it sounds great in principle for smaller sites, for local SEO. How does that come into play though, when you're dealing with an enterprise organization? Let's know, you have an international bank or you've got like a Home Depot level or somebody at that know where the map might have thousands and thousands of pages. Is there a processing delay? How does that all work together?

[10:58] Mack Grenfell: I guess, yeah, it's a good question. I don't sort of keep huge tabs on the size of sitemaps that you upload within BiWord. It currently works up to a limit of 5000 URLs on sitemap currently. So if you submit 10,000 it will randomly sample 5000. Aside from sort of some cost involved, there's no technical limitation why it can't be 50 or 100,000 articles. The system that it's built on, the vector database where it kind of stores by words. Understanding of all these articles is super scalable, so there's no reason that it couldn't be pulling from a sitemap of six or maybe even seven figures of hop.

[11:37] Matt Hepburn: Let me hop in here real quick, Mike, then I'll let you go back to that. So I just had actually Ryan Brock on from demand jump on pillar based marketing. So it sounds interesting where you're going with the internal linking. And part of that pillar based marketing philosophy is going to be they're going to do the keyword research or topical research, and you're going to be able to have these top level pillars, sub pillars, and question queries, and they all kind of link together. If we were able to have a hierarchical, maybe in a CSV or something that was uploaded and saying here's how we want to write our content and here's how we want to internally link it, is something like that on the roadmap for the future?

[12:20] Mack Grenfell: Yeah, we had some thoughts about building something similar to this before. I think just from a kind of purely prioritization and technical level, we've tended to go for things that are sort of implementation agnostic. So if we built something like this, we'd have to kind of say we're just building it for WordPress, for example, and build it in a way that kind of plays nicely with the WordPress API. And we've been a little bit cautious of doing stuff that really just works for one CMS because it's a lot of work and kind of limits the applicable audience, which I think is why we tried away from it a little bit.

[12:53] Matt Hepburn: Yeah, that makes perfect sense. Mike, did you want to jump onto.

[12:57] Mike Goldstein: Your thoughts just to kind of go back to what you were saying about uploading a sitemap and all that? Do you have the ability to kind of look at live data as well? And one of the things I would want to look at is Google search console often gives me an idea of what a site is already known for, kind of what the topical relevancy might be. I can look at the queries, things like that. Is there a way to tie into Google's API and to kind of figure out what content the site already has, kind of, or where Google at least thinks what it's about so that you can kind of support that and kind of, I guess, boost the existing relevancy to.

[13:37] Mack Grenfell: SO I'm working on a future at the moment before this call and probably will be after, which is based around the integration of Google Search console works a little bit differently in that we haven't been looking at it from the angle of trying to feed this into the generation process itself. I think there's a lot of complexities there. And it's difficult to kind of build this in a completely agnostic way that works for all sites. So we're taking, as a little bit of an experiment, taking a slightly different approach. So instead of generating new articles with it, we're using it as a way to sort of generate sections or kind of snippets for existing articles. So the way it would work is you go to the page integrated Google Search console, it pulls initially a lot of the same data, that search console pulls itself in terms of page and query rankings, but you're essentially looking for, hey, I have this page. It's ranking position one on a few keywords. That's great. No work needed. There's a few keywords where I'm ranking maybe positions three to ten. So there's a chance of getting on the top spots, chance of getting more traffic, but we're not there currently. And then the way we're currently thinking about it is having a few different modes of the sort of content you can generate. So if you just want a few sentences that could fit into this article and are optimized for a particular query where you're not quite ranking at the top so far, all the way to, if you want to generate a whole sort of H two section with nested H three s, to really go after one particular query and then be able to sort of copy and paste that or export that as a CSV, and then likely, at least in the initial version, I'd be the kind of bit of manual work in terms of copy and pasting that into your CMS. But you kind of get the idea of using some of that functionality to see where you're not ranking maybe as well as you could, and where you could add or modify content to increase ranking on certain terms.

[15:20] Mike Goldstein: That makes a lot of sense. And just kind of one of the things I'd be interested is because I've used a lot of different content AI content generators, whether it was Chattix, whether it's surfer. There's a whole host of them out there. What are some of the kind of the standout perks that users can expect when they kind of plug your content into their creation process?

[15:45] Mack Grenfell: Yeah, sure thing. So I think you're right to call out that it is a somewhat commoditized market. There's a lot of product products that do various similar sorts of things. I'd say. Yeah, a lot of the work product wise so far has been getting, or at least up until sort of the end of summer, has been getting to a point where we're kind of on par with a lot of these tools and a lot of the features that we're looking at now, I think are really taking us beyond those tools. So interlinking is a big one. I haven't really seen that many products do this in a kind of intelligent, natural way. Interlinking was one, this product I talked about earlier, that we were sort of waiting for the model to come out, the custom styles one. Again, I know there's a few products that offer similar sorts of ideas, but the sort of results that I've seen not in production, I think give me a lot of confidence this will be much better than what those have put into production so far. And then I think a big thing is that, as I mentioned right at the start by word came from my experience of doing this sort of stuff as a consultant with brands, doing it at scale, 1000 or 10,000 articles at a time. And so it's always been a very core part of the philosophy of byword that everything has to be done or achievable at scale. There's no features that sort of work for one article at a time, but don't work for 10,000. And I think that's really where Byword shines. We had good case studies of brands that have used it at really quite large scale and it's just the way that it works in terms of generating stuff asynchronously in the background. I have users on the limited plan who will go in at the start of the month, spend a few hours, not touch it for a month, and then they'll just have tens of thousands of articles with all of their kind of custom settings and configurations linked through to various WordPress sites, just generating over the course of the month and being submitted and uploaded to those sites. So I think, yeah, it's some of those custom features and then being able to do all of that really seamlessly at scale is, I think, what sets it apart.

[17:34] Matt Hepburn: Did I just hear you saying that it's submitting on the CMS by itself?

[17:40] Mack Grenfell: Did I hear. Yes. So there's a few integrations. So WordPress Webflow, there's also Zapier, and there's an API as well. So if users who build essentially build their own kind of custom integrations on top of the API.

[17:51] Mike Goldstein: Interesting, one thing you just mentioned is the ability to create content at scale. And something that I've kind of been wondering about for a while now is, is there really a best practice in terms of what is that threshold, how much content can you create? At what point does Google really say, hey, this is clearly not a human being writing this. Is it even helpful content anymore? Do you have any kind of recommendations for how quickly you can actually grow a site? Because I know a lot of affiliate marketers, for example, they're pumping out hundreds of pages a day with the goal to get, if I can get ten people to this page and ten people to this page, they're going to make money. But what is kind of the, at least in your opinion, what's that best practice? Where is the cut off? Or how much can you create safely?

[18:38] Mack Grenfell: Yes, it's a great question. I want to get asked a lot. And the truth is I don't have a kind of clear kind of graph in my mind of like at this stage it's acceptable to post this much and so on and so on, but I do have a few data points. So I've done a fair bit of experimenting myself with kind of random domains and sites with Da Zero or 0.1 or something like this and found it pretty difficult, unsurprisingly to get traction. But I've also then worked with brands. A lot of the sort of brands that I was doing this consulting with before Byword were DA 40, 45, something like this. And in a lot of those cases we'd push it quite hard. So in some cases we're talking 4000 articles a day. And in all those cases really I saw pretty good results. We get stuff indexed within 24 hours and see no sorts of penalties and kind of good consistently growing traffic. So it's kind of shaped my views to the point where if you have a decent amount of preexisting TA maybe not necessarily 40 but maybe let's say sort of 2030 or so, then you can probably afford to be fairly aggressive with the content that you put out. The cases where I've seen it go wrong are people, like I mentioned earlier, taking completely fresh sites without any sort of resource or time spent on link building and just putting obscene amounts of content, especially if it's. I've seen this go well in cases where people do push lots of content, but in a very tightly focused area. It's when people put huge spreads of content on all sorts of different topics that aren't really related to one another on low DA sites that it really starts to go wrong.

[20:16] Matt Hepburn: So I'm going to hop in there because we're getting into Google's EEAT framework again. And so one of the things that Google in this eEaT framework is kind of looking into or leaning into is wanting to have an author associated with a page or a post. And if it's created by a brand, let's say, and there's no real author associated by it, it wants to have it at least reviewed by. Right. Associated with that and that being a person. So for me, as a smaller to midsize sites that are building up with this method, it sounds like to me that that's kind of what is needed. I'm a little surprised to hear with the larger scaling that you were just talking about before that there weren't any issues with that, but I think that's good. Are you familiar with this? Can you comment on this? And Mike, do you want to tag. I'm going to tag you in on this too, because this is something we talk about all the time.

[21:18] Mike Goldstein: Yeah, I think one thing that with EEAt, again, it goes back to it used to be eat. Why do we have that extra E? Because of the experience. And the one thing that every expert out there has said is the one thing AI can't do is it can't have experiences. But I think when you add in authorship and if you can take a piece of content that in AI wrote for you, but it was a ghostwriter and you attribute it to a human being who has a LinkedIn profile, who is licensed YMYL type of stuff, if they're a licensed accountant, lawyer, doctor, whatever it might be, and you tag to them, I think that all of a sudden that really great AI generated content now gets the human boost, the experience connection, and it can be a whole game changer at that level. But Mac, I'm curious as to your take on how you see that.

[22:17] Matt Hepburn: So I'm going to go back to. Right, I agree with that. And I want to tie that back to what Mac was just saying about sites that went really broad. Right. So they don't have experience in all those different things. A person would, they would have some experience in a very specific area. So that makes more sense why the ones that were narrow, even that were starting out did all right with that, but the ones that were broad were more kind of like, I guess sounded like a smashing magazine type of site trying to cover as many topics as possible, probably didn't have this authorship associated with it. And that may be one of the reasons that I'm just guessing, but that would be what I would suggest why that had a problem.

[23:00] Mack Grenfell: Yeah, just going off the data points I have, but certainly after the HCU most recent helpful content update, I know of two sites who are going to be aggressive on content and I don't know off my head what they were doing in terms of authorship. So I'd be a bit cautious, at least on just the basis of my experience of drawing conclusions there. Yeah, the only thing we kind of noticed was the kind of breadth of content and how they're approaching things there. That seems to be a big focus of Google in that last update, right?

[23:29] Matt Hepburn: Yeah. And maybe it's Google narrowing down to this authorship, but the breadth, that's just telling it right there. What's your expertise of the site, the overarching topic of the site, and if you go too, really, really interesting. So, Mike, did you want to talk about clickbait titles and SEO Best practices? Because we know you have both a page title generator and meta description generator as part of this as well.

[23:58] Mike Goldstein: Yeah, I think it'd be interesting to kind of find out how you're striking a balance, really, between those catchy headlines. One of my favorite prompts in Chat GBT was always, I want you to write me a clickbait style headline that does not stuff keywords, something that. How do you balance kind of the conversion based type of language with the SEO best practices that Google's looking?

[24:26] Mack Grenfell: Yeah, definitely. So we don't necessarily force this in users in a way. So there's two kind of core modes within Byword. One is title mode and one is keyword mode. So if at any point users want to just kind of go off pre written titles, they use title mode, and there's no reliance on Byword to make a choice in terms of their title. But in keyword mode, it essentially adds an extra step to the side of the writing process, which is generating that title. And we've taken a fairly agnostic approach. There's not a sort of strong focus on going kind of either clickbaity or keyword stuff or anything like this. I think on a philosophical level, my perspective, a lot of the best use cases for Byword and kind of AI content scale are use cases where the title isn't that critical. You're not competing on content where there's 100 or 1000 articles out there that are going for that one specific keyword. The best use cases are generally going after long tail content where the content out there is already quite thin. And so just having being that one person making the relevant article on how to start a lawn mowing business in New Hampshire or whatever it is, is the most critical factor, rather than the kind of minutiae of whether it's SEO sort of clickbaity, your keyword stuffing, et cetera.

[25:46] Mike Goldstein: Let me ask you in terms of Byword getting the keywords and getting it right and all that. Let's just talk a little bit about Bywords research feature for keywords. Where is it getting all the data from? How does it know kind of what to target?

[26:02] Mack Grenfell: Yeah, so we have this research feature which we launched about a month ago. We launched it mainly as a sort of answer to the question that a lot of people have when they come to pyword, which is, hey, this looks cool. I'm kind of sold on the idea of AI content, but maybe I don't have a marketing person, maybe I'm a solo founder and I don't really know where to get started. And especially if you look at the sort of state of play of a lot of the research tools out there, like Ahrefs or smrush, they're great tools and Bywords obviously not going to compete with them on research functionality, but aside from the sort of one dollars for seven day trial or whatever, they're generally pretty expensive and a bit sort of, maybe a bit too powerful or a bit too complex for any user. So we wanted this to sort of address a bit of a gap in the market for these people who are kind of sold on the idea of AI SeO, but don't really, or maybe need like a bit of a helping hand. Sure. So the research feature is essentially pulling from external like a third party API. So it isn't doing anything kind of too crazy in terms of getting the data. Bioid hasn't built its own scraper or anything like this, are building a few things on top of it. So there's a kind of functionality where you can search for programmatic structures within keywords. So it's like can asterisk eat grapes? Where you get, can dogs eat grapes, can cats eat grapes? Et cetera, which kind of plays into, as I mentioned earlier, I think some of the best use cases of Byword, which is generating this long tail content at scale. And then on top of that, we also, so that maybe focuses on slightly smaller scale, kind of getting people started with a bit of research. We also have somE, quite an interesting feature, which I haven't seen anyone else do, which doesn't pull keyword data directly, but it generates ideas for programmatic structures. So I kind of implicitly gave an idea for programmatic structure earlier when I said to consider a search term like how to start a gardening business in New Hampshire. You could consider that a structure of how to start X business in Y state, and then you build up your list of businesses and states and you times those together and you get lots of permutations. We have a sort of idea generator for that as well, which if anything was a sort of personal win for me because when I used to be as a contractor, the hard bit was never generating the articles, it was never writing the code. It was just coming up with the ideas for these structures that would then build the pages in the first place. I hold my hands up and say it doesn't kind of validate these ideas in terms of search volume, but a lot of these are going after quite low search volume terms in the first place, where the data available just isn't reliable, which I don't think is a bad thing. I think there are a lot of SEO opportunities that are overlooked just on the basis that everyone's relying on Ahrefs or Semrush. And if those tools, or one of those tools says there's no traffic here, then people ignore that, even when in a lot of low search volume cases there's actually a decent bit of traffic, especially if you're going after a thousand of them at, you know, I'm going.

[28:59] Matt Hepburn: To actually reference Ryan Brock again. He's like, you know, most of these different software companies anyway for search volume. It's all based upon Google's bidding structure. Anyway, for me, this research tool sounds interesting. I think what I would be curious about, and I think would be really great, is if there's the ability to upload, say I have question queries that I already have from another tool. People also ask questions and maybe three other ones or four other ones. I want to have as H two tags further down the page. Can I provide all of that and then have it create an article? So I'm answering all these in questions and have them hierarchically underneath the page topics underneath the H two tags.

[29:47] Mack Grenfell: Yeah, there's a few different ways. So you can either do this, there's a sort of nice easy interface for doing this for a one off article, or there's a kind of way you can stitch together in a Google sheet. For each row is one article and you have your title and all the H two s, and you kind of jam those together and put it in a big string to buy word and get it generated as articles. I have actually, when I was building the research feature played around with some of the people auto asked APIs, not directly from Google but third parties. So it's definitely something that I'm conscious of. I just haven't quite found the sort of best or killer way to integrate this in which I think would be get people really excited. I just haven't quite found the way to integrate it perfectly.

[30:33] Matt Hepburn: Some of the tools that are out there, I won't mention which tool, but also provide an image, kind of like a mind map that gives you a hierarchical breakdown of like here are the top sub pillar queries and then here are the questions that are related. So for me, that's a perfect example of how you could use that for internal linking, saying each one of those sub queries is a post that's going to internally link up to that sub pillar that ultimately links up to the main overarching topic of your website.

[31:06] Mack Grenfell: Yeah, interesting. Yeah. I think I'd approached it previously from the idea of using it to build out article structures, but I think sort of going back to the idea we talked about earlier of building pillars, that is an interesting way of looking at it I haven't really considered before. Yeah.

[31:22] Mike Goldstein: Can I ask just kind of taking that a step further, is your tool is clearly thinking about how best to structure and how best and what to pull in and all that, but how much of that comes back to just the proper prompt engineering and being able to actually ask the questions of the program the right way, rather than just kind of the one liner. Hey, write me an article about skylights in Manchester, New Hampshire.

[31:51] Mack Grenfell: Yeah, absolutely. It's a good point, and I think maybe you'll have experienced this playing with Chat TBT, but there's a, I've never seen someone like the right term for this, but call it like a kind of prompt budget or something, where if you try to stuff too much into a single prompt, you'll find it kind of adheres to less of it the more you stuff in. So I would use as a sort of recursive prompting system where there's not just kind of one big prompt, it does lots of different things at different stages and then sort of goes back and loops through in a way that you're never sort of overloading the prompt at any one point and you're just trying to give it the minimal viable information to get that stage of the article written. And similarly, you see this, we have one feature which we call for the users called custom prompts, where they can kind of inject bits of text to say like, hey, I never want you to give legal advice or something like this. And again, it's tempting to kind of make this limitless and let the users put in as much content information as they want, but I've always been keen to keep the amount they can inject and supply fairly limited, just so that it doesn't kind of detract from the overall prompt that byword kind of passes in through the API and therefore reduce article quality. But it is an interesting one to think about how you especially like some of the new features that we're thinking about and layering on top where exactly those come in. So one thing that is on my mind, and we haven't sort of laid the framework for this yet, but it's being able to handle CTAs and the ends of articles better. Right. So you've had some users sort of MacGyver their own way of doing this, which is inserting it into the custom prompt of kind of explaining what their brand is and saying, hey, include a CTA like this at the end, which is a good start, but isn't totally reliable, because again, you're giving a lot of information to OpenAI and forcing it to try and kind of keep all of this in its mind as it goes through. So I think this is actually a good one where we can kind of, in terms of how I look to build this, we can kind of do the whole article writing process, and then at the end we kind of take a few bits of the article, provide context to the API, say title, introduction and the current conclusion, and then say, hey, based on this and some other auxiliary information that you've given us about their product, their pricing, what CTA they want, then just generates the paragraph that comes after this and provides the CTA. So a nice example of where you're not trying to jam everything into one prompt, but you're taking it kind of slowly and intelligently and following it at different stages of the article process.

[34:19] Mike Goldstein: Interesting. The way you describe is very similar, and I'm not going to mention the software here, but there's an AI video software that I started to use, and one of the things I love that does at the end is put the prompt in after you've created the content and then say, hey, can you change this or that? And kind of that multiple prompt sequence, I think just allows you a lot more control and allows you to really personalize, to kind of dive deeper into whatever content, whatever medium it is, whether it's video, whether it's written content, whether it's infographics, anything. The more you can customize it, I think the better it's going to be. Just for search overall.

[34:58] Mack Grenfell: Yeah, absolutely. It's also an interesting one from an optimization perspective, because if you're prompting a language model or some sort of image model with a fully pre written article or image. It's going to be more expensive because you're passing stuff back. So there's an interesting optimization problem, how you manage the costs here, both in terms of your PNL, but also what's the minimal viable bit that you can pass back in in order to be able to tweak the ask or image just how you want it based on the user's feedback?

[35:25] Matt Hepburn: So I'm going to throw a question in here that wasn't in our list. Sorry about that. And I don't know what the answer is going to be, but my question is, can this be adapted for local companies that are doing local SEO and want to show up in the local organic search results, not the map results so much, but is there a strategy for that, or do you just kind of base it on the larger topics and then have them go optimize things later?

[35:55] Mack Grenfell: Yeah, it's a good question. This is never a sort of primary use case for Byword, especially depending on how local you go. You might sort of reach the limits of other in a certain way, despite never really positioning it in that way. I can think of quite a lot of the sort of enterprise users we have who've gone really aggressive on that local side of things, have a few keyword structures and they're just passing through huge lists of thousands of towns or cities across certain countries. So I've certainly seen good feedback from users in the sense of how they're using it, even if it's not sort of what it was initially intended for.

[36:33] Matt Hepburn: Okay. All right, that's fantastic. Do you have any use cases that are like great stories you can tell us about some of the brands that have been working there or maybe anonymize those, but kind of like the results of what they've had using Byword. That would be.

[36:49] Mack Grenfell: Yeah, yeah. So the big one you talk about a lot was in the sort of early days of Byword where there wasn't a sort of Byword AI, but there was a code base that looks very similar to what's in the product now was working with a startup, a financial modeling startup called Causal, and they positioned themselves as a competitor to Excel, sort of more advanced version of Excel for people who are doing financial modeling. And so this is pretty exciting because Excel, there's loads of content out there. Language models understand Excel well. There's a near infinite list of kind of questions and queries that people have. So we took a few different approaches and sort of broke those down into structured the pages that you want to go after. So some of them were very sort of like structured around an input list. So for example, we thought every Excel function or formula, depending how you're going to call it, every Excel function and every Google sheets function as well because they're pretty similar. We're going to take those, put those into a list and run them through the Byword or soon to be Byword code base to generate articles on those functions. And you'd think this is the sort of thing which a sort of financial modeling SaaS app would never be able to kind of compete on. But all of a sudden within six months or so, we're kind of outranking Excel and Microsoft's own pages and Google's own pages on those terms. And this is just kind of looking at the functions. But we also did a few projects where we're looking at going to Ahrefs and finding the top 1000 keywords that contain the phrase and then Excel or Google sheets and again passing those into the kind of pre Le version of Byword. And this was kind of where we were publishing at peaks of 4000 articles a day with decent DA to start off with. As I came back to what we were talking about earlier and just saw really linear growth. So up to the point where a lot of this was launched at the start of 2022 and in September this year we just passed on their analytics, a million sessions a month on those articles and has proved pretty resilient in terms of HCUs and all sorts of updates as well.

[38:57] Matt Hepburn: Yeah, absolutely. Mike, did you want to ask Mac about the conference? About the OpenAI conference that happened?

[39:08] Mike Goldstein: Yeah, I'd like to hear, Mac, your thoughts about the conference, about how OpenAI seems to be something that now is going to be a lot more available, seems to be something that is going to be able to be featured in just a more product moving forward and kind of what your thoughts are on all of that.

[39:26] Mack Grenfell: Yeah, so I think a lot's been said already. I think a lot of the focus has been on the kind of shiny fancy new stuff. So a lot of the kind of vision models, having APIs release those text to speech stuff like this, this whole GPT store of kind of creating customer systems. And I think that's all exciting in a dev conference and it looks good on Twitter threads and stuff like this. But I think the most interesting stuff is the upgrades they've been making to some of the base models. So it's stuff like the release of GPT Four Turbo, which is a sort of lightweight, more economical, faster version of GPT Four, which is not quite in production release yet, but there's a sort of like toy version for developers to play around with. And I think this is quite interesting. So I suspect certainly in Byword's case, and probably for a lot of other apps, there are a lot of features that developers have wanted to build, but either they've just been too unreliable because of the processing speeds or the cost of generating, it has just been too much. And so I think even though it seems like a really boring upgrade, them dropping the price by two thirds or whatever and speeding it up a bit, I think you'll see a lot of features that could have been possible, but people didn't really want to build or release for economic reasons that now will become available. And similarly, there's a similar sort of thing with 3.5 Turbo, where the quality is maybe not quite as good as the existing model, but the speed that it generates content at is just absurd. If you weren't doing anything fancy and you just wanted to build the fastest AI article writer possible, you could get something half decent out in three, 4 seconds and have it be of a length that was reasonable for SEO.

[41:09] Mike Goldstein: That makes sense. It sounds like just open AI is going to be taking us in a whole other direction, or not even a whole new direction, but kind of opening up the direction that we're kind of already going and taking that river and turning into an ocean. There's so much opportunity out there now, I think with AI, although as I like to say, we don't even have AI yet, we have machine learning. AI is even yet to hit us, and I can't wait to see what happens when it does.

[41:35] Matt Hepburn: Absolutely. Hey Mac, did you want to talk to the audience about any type of offers or how do they reach you, or find out more about Byword on the website or demos or anything of that nature?

[41:49] Mack Grenfell: Yeah, sure thing. So on the website there's a few things. I guess a few things are. So firstly, if you sign up by word, you get a bunch of credits. Just play around with background, see what works, see if you like it. There's the case study that I mentioned, if you scroll down the page, the causal case study about how we got to a million sessions a month there. There's also a pretty comprehensive docs or kind of learn section, which talks obviously not just about how to use Byword, but there's also a kind of big section covering quite a few of the topics around broader SEO strategy that we talked about here and is particularly useful, I think, for people who maybe don't have as much of a background in SEO or kind of want to understand how AI changes things. So yeah, a few things for people to look at there.

[42:38] Matt Hepburn: Fantastic. Fantastic. All right, well, this has been absolutely amazing and really educational. It sounds like there's a lot that's coming kind of on the precipice with Byword. Do you have any kind of idea of when your release idea or release dates for these new features are?

[42:56] Mack Grenfell: Good question. I would hope to get some form of search console integration out at least in beta next week. I don't know when this is being released, so maybe I should say this week or last week. The custom writing styles is dependent on OpenAI and what they do with their models, but I'm super excited for that. Hopefully before the end of the year. Yeah, most of my work is just pumping product out, so we're getting stuff shipped pretty quickly.

[43:22] Matt Hepburn: Okay. And then lastly, what do you see for 2024 with AI? And by word, what do you see up on the horizon.

[43:33] Mack Grenfell: In terms of AI generally? Who knows? I don't even know what will happen next week. But in terms of Byword, I think we're sort of always looking at what models are coming out and the best ways, almost intelligent ways we can integrate those, especially integrating with existing sites to do stuff that's kind of above and beyond. Just press a button, get an article. So like, interlinking was a big first step here in terms of understanding the content on your site. This custom writing style I talked about, there's a lot we want to do in terms of lots of users have custom data sets and I think there's some really exciting stuff, all custom catalogs as well, and product lists. I think there's some really exciting stuff to be done in terms of generating content around those because anyone can generate an article on the best time of year to visit Hawaii, but when you get to the best five scarves to wear to Thanksgiving dinner or whatever, that's a little bit harder, integrating it with a product set. So lots of things like that that we're really excited for and will be coming out, I'm sure, at the later this year or next year.

[44:32] Matt Hepburn: Fantastic. Well, thank you so much, both of you, for joining us today, and we'll be talking to you soon.

[44:39] Mike Goldstein: Thank you.

[44:40] Mack Grenfell: Thanks a lot.

[44:41] Matt Hepburn: All right, take care, guys.

[44:44] Mike Goldstein: Bye bye.

[44:45] Matt Hepburn: We hope you enjoyed this episode of the Entrepreneurs Marketing Journey podcast. To get the most value from this episode, make sure to check out the show note resources in the episode on EMJPodcast.com and if you got value from this episode and you would like more marketing tips from us, then feel free to subscribe to the podcast on the podcast platform of your choice so you never miss an episode. This is the Entrepreneurs Marketing Journey podcast with Matt Hepburn, and we'll see you next.