#2286 Pepper: AI + people = > $10 million

0:00Hey there, Freedom Fighters. I've got a story about a guy who's creating a content business that's exciting. Check it out. I read about a guy who uses AI and people to create over three quarters of a million pieces of content. So I asked him to do an interview about how he does it. Let's watch. Anirud Singla is the founder of Pepper, which uses AI and humans to create content for enterprise clients.

0:27Anirud Singla, what's your revenue right now at Pepper? so we just watched 10 million in arr and we're now in the 10 to 25 to 50 journey so yeah being uh it took us some time to get us here but uh yeah very excited to finally cross this when you were working on upwork as a writer how much money were you making a year then you know it's funny uh so i just for reference i uh i was trying to figure out a way to finance my education and the goal was to earn$2 ,500 in two months and I've never earned a single penny in my life and I slogged about 17-18 hours on Upwork, Fiverr, all these Chegg and all these other tutoring platforms and made that happen and you know that actually gave me this one very unique insight that you know hey I put in the hustle and the slog to get this done but there's a real need out there where there's like like i discovered these facebook groups which are like hundreds and thousands of writers at that time and i figured someone needs to aggregate this market because huge value and uh if you just recently saw mercer which went crazily to i don't know 250 there is 250 300 million like uh 250 300 million where they're getting freelancers to train on ai data sets.

1:54So just see how the journey moves from, you know, people going from content production to not training air data sets. And the story as I understand it is, you said this is way too difficult to find another writing gig. Every time I have to look, I end up spending more time looking and dealing with the business of getting writing work than actually doing the writing, which is what people are paying me for. There's got to be a better way, a better model. And you created it, you called it Pepper Content. And it was a marketplace where you personally at first were matching customers with writers. The writers were in India, so the price was lower than it might be where you are today, which is San Francisco, right?

2:29That's the model that you had. Yeah. And the model evolved quite heavily where we now have a global talent network. So we have about close to 150 ,000 freelancers who've applied to write for Pepper. And this is not just now writing as well. So just a minor correction, we also do design, video production. And so it's a multimodal in terms of what we do. And the talent pool is now global. So 50 % of our user base on the talent marketplace side sits out of the US. So and we work only with the top 3%. So our value proposition to an enterprise today is we have subject matter experts in retail, cybersecurity, crypto, who are US native experts.

3:19And we're also working with global organizations who need regional experts. So Pepper also produces content in 45 languages. So we right now are working with customers where someone needs Arabic strategy for search or someone needs a native Chinese transcription. So we have that level of global imprint. So far from, you know, where we started on. And also far from just humans. AI is the next thing. How does AI and humans, how do AI and humans work together with you? Yes. So I'll give you a quick evolution of how Pepper also evolved. So we started off with this managed services marketplace for content production, where we were matching customers and these writers.

4:07and then we tumbled upon this thing called gp3 and this is two years before chat gpt this is 2019 we're talking about and we were the first 50 users to get access to this i had cold email sam altman and greg brockman to get access they were happy to give it to us and we realized very soon that you know if you put content on a number line of one to ten uh there is probably and you know this is my prediction that time and obviously the strength changed one to four which is high volume low value content will get automated and everything above that five to ten will need experts enabled by ai i can say safe to say that with ai now it's one to seven or one to eight and they keep getting better and better and that's when we realize that ai needs to be natively embedded in workflows to produce content at scale and do effectively.

5:05So we built a lot of tools. We built this tool called PepperType.ai that was an AI writing tool that scaled from zero to half a million users while we were running the marketplace. And we realized that, hey, there needs to be this marriage which embeds both. So we started building in this agentic technology called Nimbus, which combines how AI can enhance humans to produce content at unparalleled scale, but while still having a human review layer or editorial layer in picture. So it's still a human engine, but it's heavily propelled by AI. And just for reference, we have now about 500 custom apps, which are powering, you know, this human plus AI content engine for the world's largest enterprise brands, while ensuring there is human level quality checks.

5:57But the first layer is AI starts the work, creates the first draft always, then a human being takes over from there. Do you then take what you've learned from how the human being adjusted the AI content and feed it back in for that? You do. Talk to me about that. Yes. So we have what we call as a reinforcement loop where it's not just, Andrew, about AI generating the first draft and the human reviewing it later with as part of the platform technology we built called Nimbus and we're publishing a lot of blogs on our website now which talk about how Nimbus works we are able to create unique insights and fetch data from a lot of sources and throw that into the LLM and it's not like I'm going to charge GBD and putting in a topic saying hey help me create content on this and it gives me something and i use that we chain multiple prompts and these are very complex prompts that we write for enterprise customers and we chain those responses and each response's output goes as an input to the next response so for example i want to create an article on something i'll pick up that topic uh my platform technology called Nimbus.

7:13What it'll do is we create workflows where we put the topic on SERP. We will see the top 10 URLs that come in. Our platform will scrape all those top 10 URLs and see what all have they covered. What are the H1s, H2s, all of those things. We combine all of that intelligence into one flow and then identify what's the most optimized SEO focused article. And obviously now it's GEO. So we are also inculcating workflows you need to optimize for AI search as well into the brief when the article is getting produced. So it's not just about producing an article. It's about producing an article which can optimize and rank.

7:53And we are taking insights from what's already ranking from competitors. And what we also end up getting is a human who used to end up doing comparisons across 10 competitors, seeing, hey, they've added this, they've added this, all of that is already being done as jobs to be done by the AI just in part of the research. And then we get the human, we get the draft generated, we have a human review it, it goes live, then we do another two, three things. We also look at data from Google Search Console and also have figured out workflows which can re-optimize that content six months down the line if it's not performing.

8:33And that's something that customers are loving Pepper's technology for, where people have websites which are like thousands and 10 ,000 pages of content, which they produce sometime or the other, but it's probably not doing anything to their business today. How can we re-optimize that content automatically to start driving more hits and traffic? And that also helped in Pepper's evolution as a company to now a content-led growth company. Okay, you and I were looking at a Humana article earlier today where you said this is an indication of how content can be written. Let me bring it up right now.

9:10Tell me why this is good. What am I looking at here that says that is leading it to perform especially well in results for AI searches? Yeah. So I'll give you a quick analysis on this. So if you start at the top, you'll see this one thing called published date and updated date. Right? Now, this is very powerful. You'd be surprised. So just for reference, all LLMs give, there's a 25 % probability that your content will get surfaced up better if you map out freshness of content. So this is actually a signal to the LLM that this content was recently updated. So LLMs love freshness. So that's one very small insight.

10:02By the way, small note, the publish date was April 21st. The update date is April 16th. So this seems like a bug or some kind, right? But okay. Yeah, that's probably a bug. But if you get, yeah, that's a bug right now on this. But just to get you a sense that notifying the LLM when this was last updated is like the, you know, one of the school hacks that is driving a lot of search. Okay, so that's one. If you go down further. Keep going further, okay. That's one. If you go down further, you'll see this key points, you know, summarization. It's like a TLDR summary of the article. Now, if you normally have seen, we've all been, you know, brought up with this logic of throw everything you have on that topic in that article and make this like a 3 ,000, 5 ,000 word long form article.

10:50Guess what? The user doesn't care. They want to get their answers in a very simple manner. And the LLMs want to summarize this in a microsecond split decision. So if you can do the job for the LLM, if you can already summarize your existing content in a TLDR summary, that is a huge advantage. I see. So I've taken one more step away from the LLM. It doesn't have to summarize it to analyze whether it should provide it. Okay, so that's the next thing. I like that. That's the key point section. now if you go down further if you see this compare medicare supplement plan costs now see how it's giving an example uh if you see how llms respond they're great at giving these example structures because people are going in and putting in comparison queries they want to know differences between different plans and you'll see llms are very great at whipping up tables.

11:46So if you go down further below, you'll see this table summary, which is almost like a comparison analysis, which gives you a sharp understanding of what's working, what's not. Companies starting to do that is super interesting. So you'll start seeing that these are things that drive a lot of LLMs to suddenly start pushing out this data because they want to give a user a holistic sense and try to appear unbiased. Okay, I get it. I've heard that too, that this kind of data is what the user wants. And so you might as well put it in a table and make it easy for the LLM to give it to them. Anything else I should see here?

12:28I have a few factors that they could use to even further optimize this better. If you see the H1s and the H2s, the headings, right, or the subheadings, if you convert those subheadings into interrogative questions. So instead of sentences, if you would have written what are the factors that affect your Medicare supplement plan? Instead of what's written here, which is factors that may affect your Medicare plan costs. That's the title. You're saying turn it into a question would have been even better because? Interrogative question. Because FAQ schemas are and FAQs are driving a lot of this growth in terms of user discovery.

13:08So people are seeing like we've worked with customers where they've seen massive traffic uptick just on converting the H1s and H2s into interrogative questions. And that's something that can be done across the website because suddenly now your website's already FAQ optimized. and it's being built and showing trends based on that. Okay, I'm with you now. I see this. And this is what you do at scale. How many articles would you say that Pepper is responsible for creating a day? So just for reference, we've done about 750 ,000 articles till date. There have been months where we've done 30 ,000 articles a month.

13:52So we've seen scale of all kinds. Right now, what we're focused more on is not just, you know, huge volume content production, but content which will actually drive meaningful outcomes. So Pepper has now evolved into a content-led growth engine where we're helping companies show up on AI search through our own platform. And the interesting part is we have workflows built in. These are autonomous workflows which can accelerate content velocity. which can accelerate things like content refresh, which can give you a sense of what are low-hanginged suits that can help pick up traffic at scale. Because if you see 80 % of the websites in the world, traffic's been going down.

14:41People don't know how to respond. So we're working with CMOs. So we do a lot of these CMO dinners. So we've done like 40 CMO dinners over the last one year. And you'd be surprised. everyone's still not able to get that you know it's not about seo being dead it's now seo plus geo we call it geo generative engine optimization which is uh it's creating new real estate in the internet and suddenly it's not just about optimizing your own website you need to think about platforms like reddit quora uh having a wikipedia page you'd be surprised we've literally created 15 Wikipedia pages or helped create that over the last few weeks.

15:25So there is this huge value in now starting to analyze what factors do LLMs spend a lot of time giving weights to? And that's how you have to now think innovatively about organic content and search. And what you're doing is automating a lot of it. And so, for example, if a page has been published and it needs to be refreshed, your software will automatically refresh it and add the new updated time to the top of the post. If you're seeing that something's not ranking, you're looking at Google Search Console, understanding what's going on and editing the page for your client in real time. Updating the pages and Andrew, it's also giving them visibility on where they get shown up on LMs.

16:11So we've built this platform called Atlas, which is our own proprietary tool, which where we define a universe of prompts that could potentially get you, that probably users are using for, and we have these volumes attributed to those prompts. And we throw those prompts on these five different LLMs like Cloud, Perplexity, Charge, GVD, Gemini, and all of that. And we see where all is your brand getting mentioned. And we start seeing, is your brand getting mentioned? Do you have citations? are these LLMs picking up content from your website or not? And once we get that analysis and most companies are doing, don't get a great analysis because of them not being LLM ready, we then give them a comprehensive checklist of, here are the 10 things you need to fix.

17:01These are the low-hanging fruits. And then we have technology which can automate a lot of those jobs to be done that they would have probably needed like 10 people to do in a far effective manner. But like I said, our differentiation is we're a platform plus service. So it's a solution versus it just being like a tool that a company can use and be happy about. Because fundamentally, the lift needed to pick up things like this is huge. Like we work today with Ecclasean, Sprinklr, ClickUp, Instacart. These are big enterprise companies which care a lot about organic search driving a big part of their business.

17:42and for them, impact is directly linked to revenue. So we now think a lot about speed in terms of how can we get organic transformation in these enterprises. And one big learning, and what I want to share also is video search is going to be huge. We're helping companies show up and start producing a lot of content on YouTube because soon LLMs will actually push up videos a lot more. How are you doing this now? articles. It's not, you and I talked before we got started about how it's not just human beings creating videos. You're using Sora and other tools to create videos using AI. When you're using Sora, you're not using the Sora like social app, are you?

18:27No. What are you using? No. So what we have is obviously all of these models and the API is integrated. And like I told you, we have built an agent technology, which allows us to create custom apps and custom workflows for customers so imagine we're working with a travel tech company where we're helping them create hundreds of videos which are performance marketing ads on instagram where we're putting in the text almost converting a top 10 places to visit say in west coast into like almost a reel which can be then run as a performance ad to a user with colorful creative now give me this whole workflow.

19:09So you're using AI to create videos that you know will perform and when they do well, you buy ads for them. So where do you find the content? How do you know what content to turn into videos on Instagram and YouTube? Yes. Yeah. So if you see most of Instagram and TikTok, they've just become indexable on search, Instagram specifically. So which means LLMs can now actually go through your video and see all the text you have on your video and then prop that video up specifically. So it's become machine readable metadata. So that's why what we do is we pick up all these listicle kind of topics and queries like top 10 places to visit in West Coast or here are the five things you're not doing.

19:52All those article content which no one now reads but wants to love seeing good videos and visuals about. We create summaries through a normal article builder. We then convert it into a video script and the video script is made punchier with respect to audience cohorts. So we have like, we work with customers to define either 10 audience cohorts you might want to build on. We create 10 scripts. We basis each script. We then add in the video elements and the brand elements that they care about. Get a video popped out. Play those videos organically. See which ones perform. Do A-B testing and then run ads on the ones which perform the best.

20:34so this is like agentic workflows and and telling you these videos very soon are going to start popping up on search instead of that article you were reading from you know right maybe booking.com okay and who's doing the voice what tool are you using for voice what tool are you using to get the writing to not feel like ai slop or is it ai slop and that's okay right now yeah at least in the short form video it is ai slop uh and it is getting better uh as you know these models become more you're saying it's a i slop but it that you're producing but it's still doing well it's still doing well but i'll tell you it's not sustainable long term so one has to eventually create that workflow with script writers and you know basically smarter people have to use these tools it's not like you know a normal person can just go in and write in this and get this out you need person with creative input creative output understanding to get this workflow in place okay uh we use tools like 11 labs which help us with voice and we're using we're also uh you know building a lot of synthetic ai models where we can have a human be generated completely from ai so imagine i and you are speaking and we have a platform which understands how we're speaking uh on a live basis and then starts mimicking us and then starts the next time i you don't need to get on a podcast.

22:01You can just speak to Anirudh's AI avatar, who's going to make the same hand movements, hand gestures, facial expressions, and probably speak to you in a much better condensed manner than I am doing. You know what? I've used 11 Labs for audio, and it works so well that when I do it for people using their voice, they can't tell that they didn't talk. They think it's them and that I captured a video of them somewhere. I understand also how to use AI to write in a certain voice. Like you can kind of write using my voice. There's enough content out there. What about video? Are you saying now that you can even have talking head or do you need to have images only?

22:40Can you do a talking head like me and move and gesture and everything? You can. What are you using for that that works? Yes. So there are a lot of proprietary tools right now. So these are not like, you know, mass tools at this point. we're experimenting with a lot of them nothing where I've seen huge success where I can tell you that you know this definitely works but I can tell you this entire concept of Synthesia I don't know if you've heard of them they've now obviously expanded into a much larger company I used to you know I had a friend who used to run this company called Rephrase.ai which used to do this which was basically synthetic voice cloning and And they got acquired by Adobe eventually because Adobe wanted to introduce them in their GenEI platform features.

23:31But there's hundreds and thousands of experiments being built and run on these things. Are these good enough right now that Pepper is using any in their videos? Are you using Synthesia? You are. You're using Synthesia. So here's what I understand that you're doing. You are finding articles that work already in a format that people just aren't going back to, which is text. You're turning them into scripts. You now have a script that works. You're asking for, based on the user that this script needs to be customized for, you're customizing the script. You're then having a talking head using Synthesia or some other tools.

24:08You're still playing around with what works best. you have a human being plus some images that go on top plus an audio an ai generated voice all creating a video that's going on instagram and youtube we're not talking about stuff that's competing with mr beast or mark rober but we're talking about stuff that come that competes with text which people aren't reading anyway got it and you're saying we're going to keep iterating and improving it i see what you're doing and the tools that right now that i'm sorry go ahead and you know the vision with it is not just the perfecting workflow it is to engineer growth so we're thinking about content-led growth where we're saying can i engineer growth where i create 100 videos which can get you potentially this engagement rate or half a million views and it's imbued with data so we know the search volumes are the keywords there are a lot of tools like grid iq which tell you search volumes on youtube uh you can also have a lot of tools mr b also has one of his own tools around that um and once youtube analytics is getting very good uh and similarly all of these other platforms analytics are getting better if you can first buy those analytics into the workflow process you can re-engineer how you can do content better and that's what we're about i have no idea you're doing all that okay images too you and i were talking before we got started about cling and other tools that you're using because it's not just for text it's not just for video what are you doing with images and what ai tools are enabling you to do that yes so we use a bunch of ai tools all the standard ones including run you know runway for video mid journey for creative uh what we're seeing very interestingly with this is we're doing a lot of always on creative content so imagine you go on doordash and instacart's app you'll see these banners floating around which are offer banners or which are talking about deals and all of that now imagine and you'll be surprised most large enterprises still do it humanly in a manual manner including the push notifications including the creators and all of that we've created these custom engines which can produce banner images, add creatives at scales like 30 ,000 creatives over the last six months.

26:29And we're E-B testing this with data to see which creatives work. And then automatically that's queuing an engine to the LLM engine to refresh that creative, to see how we can write more CTRs up. And with scale at say Instacart or DoorDash level, you'd be surprised that this translates massively into revenue. So we're seeing CRM teams or customer lifecycle marketing teams and all these big consumer apps heavily iterating on AI usage and push notifications and anything customer comms, which could drive more revenue and acceleration off. If you were an individual owner of a content business and you couldn't hire Pepper because Pepper is aimed for a higher level of customer, right?

27:15Enterprise and middle market, I guess, is the way that you've said it. But if you were smaller and trying to create content using AI today, what would you do? What tools would you use? How would you think about it? Yeah. So I would recommend that don't go in for the staple, put it on chat, get whatever it's throwing out and put it out. You need to make content more intuitive. But if, see, what's happening is with this onslaught of AI stuff being put out on the internet, people need to be smarter about how they create content. They have to, and especially it's becoming harder with showing up on LLMs as well.

27:55So you need to build in authority into whatever you're doing. So here's what I would do. I would actually go and create videos and I'll start putting myself out on LinkedIn. I'll use those videos and repurpose those videos into content and then create articles about me and then start referencing that he say Anirudh Singh spoke about this on ABC and start building more trust signals around it I would create a lot of content on LinkedIn pulse just as an experiment you'll see LinkedIn gets cited almost on 12 % of chat GPT citations so if I start I want to build out a personal brand as a founder, you know, to drive leads for my business and eventually grow.

28:41I start putting out a lot of content on LinkedIn Pulse, which starts to drive this narrative of a UGC growth lever, which an LLM would want to promote and embed my video at the top of the article and then have a summary of that down there. So you got to become intelligent about the way you put out there yourself. Like at Pepper, we have 320 ,000 followers on LinkedIn. I have 80 ,000 followers on LinkedIn. And we've seen that you need to engineer these growth loops into content. So just be smarter about it. Don't just, you know, copy paste from chat CBD. But post on my own LinkedIn account about myself.

29:24Yeah, like you can talk about, you know, hey, these are the things that you're, basically what you need to start building out on is authority and signaling. Like for example, look at you. You focus on bootstrap giants. That's the theme, right? And today when I search about anything around bootstrapping, either you, Jesse, pick up, come up. So it's a lot about owning a niche and building that authority very strongly. And I would do video, you're saying, and post it on there. and you'd also do chat GPT for scripts and for text, but don't just take what comes out of there, add the human element afterwards.

30:02Is that what you said? Yeah. Okay. I like that. By the way, I do see a lot of LinkedIn used in chat GPT and others, which frankly helps me. I looked you up on chat GPT and I liked that it brought up LinkedIn because I would have to go and do it myself if not. There was a period there when they weren't and that was a pain in the neck. You also mentioned earlier Reddit. Reddit is used all the time. I don't think I'm going to quote Sean Puri properly, but he said something like, ChatGPT is a search engine that gives you answers based on what someone said on Reddit five years ago. And in many ways, that is true.

30:36How do you, as a company that works en masse, get Reddit posts created? Yeah. So our Reddit playbook to customers is about don't go out there and try to get a random posts about you getting created because Reddit will ban them. Reddit is a highly, strongly moderating platform. What we would say is pick Reddit as a social media channel. So what Reddit is now undergoing through, obviously, they've obviously got a huge lot of publicity. They're the 11th most visited website in the world. they've started creating and inviting brands to create brand pages on reddit and create communities on reddit so that it's become like your social community platform everyone's been thinking about hey i want to create my own community which platform should i list it on i would say go do it on reddit where engage with users add value do amas and it's not just having you know marketing speak on reddit because like i said reddit typically tends to ban promotional posts you've For show value, like some of the best companies I know are actually getting people from their product teams and engineering teams on Reddit to start posting out on queries, which are helping them in deals.

31:53You'd be surprised. And they're doing it authentically from their handle. It's not advocating they use us. It's advocating this is how you solve a problem. So don't, you know, go the route of influencer marketing and all of that on Reddit. It's great. Won't work. build value, build it as a social media channel and do it credibly. I see. And so Pepper, you don't have any automated AI driven way of doing it. You're just encouraging companies to post on Reddit or are you? So we have an automated way to pull up threads which you should be answering. So what we've done is we have a Reddit module where we can scan through your website and the themes you care about and pull out all subreddits and all threads that need your attention and answering.

32:42So that at least gives you an intelligent way of how you can frame your content strategy on this job. I see. All right. I keep coming back, by the way. I don't want to forget to ask you. Who at OpenAI did you cold email? What did you write that got a response from them? Yeah. So I primarily cold emailed Greg Brockman, who is a co-founder and president. And I cc'd Sam Altman. and I got a reply from Greg in less than an hour from sending an email when they probably had less than 50 users for GBT3. And what did you write that got a response? Was it just, I want to use this software that I've heard about?