Staying Visible in the World of AI Search and Invisible Clicks with Quentin de Quelen

How do majority of people search? They go to Google, they put in a key phrase, they see what they like, they click into it. That behavior is rapidly changing. Today, 46% of Gen Z prefer Instagram and TikTok over Google. YouTube is their go-to for tutorials and reviews. And then you add tools like ChatGPT, Perplexity in the mix, voice assistants Siri, Amazon Echo. All of that gives you no links to click into it. You just get the, the answer, the context, right? So here's the big question: when that behavior becomes the norm, will they still find you? Hey everybody, I'm Mike Sirius and this is Mastering Tech Growth. And Every week we sit down with industry leaders to bring you no-fluff tips on how to grow. If you enjoyed the show, please make sure to subscribe and send me your biggest growth challenge at masteringthegrowth.com. I really wanna know what you're working on and what keeps you up at night. Joining me today is Quentin. He's a CEO and a founder of Mailyssearch, one of the fastest growing open source search engines on the planet. He's been building high performance search tools for years now, helping companies create lightning-fast, typo-tolerant, and fully customizable search experiences. And he's been sharing that journey on a global stage, conferences like Web Summit and Slush, shaping how we think about discovery in the AI era. Quentin, thanks for finding time for us today.

Thank you for having me.

So future of search is a big, big topic, right? And we only have 60 minutes with you, so we'll concentrate on what actually happens when the search leaves the search bar and how companies and individuals can still stay visible in the world of AI and effectively invisible clicks, right? So we'd love to open the show with this question, Quentin. When I say mastering tech growth, what comes to mind?

When you say mastering tech growth for search?

In general.

In general. Yeah, so because I'm in the search topic, I think the, uh, okay, most important for me is mastering and not being passive. And it's where it's super important to be able to, to use the right tool, to be able to not only be passive about what others are doing, but really being able to mastering how people will interact with your product, your tool, and how you can master your growth overall.

Okay, so did you have how you approach a path to mastery? Like, what's Quentin's way of making sure that you stay relevant, that, you know, you're getting all the knowledge and you just, you know, you're continuously building towards something which you can call mastery?

Yeah, so basically I'm using most of the tools I can daily. I try to stay up to date on every tool that are released, try to try most of them and to get my own knowledge about them, not only following what has been said on Twitter or on Instagram, etc., but looking really like how it can serve my business and how I can use them. And then I try to adapt what I see, what I learn to actually manage such businesses.

I love that. So you're playing not only with your own tools, you're actually playing as well with the with competitor tools and just a broader range of all the tools when it comes to search?

Yeah, search today is moving quite fast. It has been something for the last 20 years. It changed the last 8 years slightly, but now it's kind of changing every month. It's going super fast.

Every month?

Yeah, it's just a few months ago, AdSense was not existing. And the position of search in this agent world could be very strategic, but it's not really defined yet. What will be looking the world of agents, the world of search in the next 6 months, 1 year, 2 years is not defined yet. So we have to be always on the starting block to look at what are the news, listening a lot of customers, understanding what are the actual pains, what would be the future pains, what they could be afraid not only about search, but in general with this AI evolving stuff, because most of the time it could have some impact on search.

Okay. I love that. I mean, you're not the first person saying that the landscape of AI tooling and/or AI impacting behavior and processes has such a huge impact that things just is evolving so, so quickly that it's almost impossible to keep with that pace. So I want to set the scene with and kick off the episode with this one, right? So majority of the teams design the search experiences or search tooling for the screen. And as we implied in the beginning of the episode, that behavior is really, you know, changing, right? We have voice activation. We go to ChatGPT, we just put in, you know, we get the answer. So there's effectively, there's no clicks anymore, right? There's no going to the website and According to McKinsey, nearly 70% of customers now expect to interact with companies via voice or chat interfaces, right? So the companies building for these formats already seeing better engagement and reach. So Quentin, what's the biggest mindset shift companies need to make when their content isn't being clicked anymore? It's being spoken, it's being summarized, it's being embedded in someone else's answers.

I think it comes with the fact that there is more and more information online and there is more and more product to sell, there is more and more everything. So the general shift has moved from people are searching for content to people are discovering content. And this shift means technically a lot of changes because people are not exactly typing the right words, are not exactly expressing the right thing. They want to be to have personalized experience. They want to be promoted with exceptional content that fits what they want, what they need, etc. So they are always trying to discover what could be the next thing, and they are not exactly looking for one specific thing anymore. I think the challenge now is that there is different medium, as you mentioned. The good point for us, for Mailsearch, is that whatever the medium, there is always transcription to a final search that will be made. And because search is not about only having a search bar where you type some word and you receive some result, it's mostly about the proxy between data and humans. And now more and more AI agents and stuff, but basically between someone or something that have an intent and the data that could answer this intent. And you can see search only as a proxy between the two of them. The key challenge in that world is that this proxy should be owned by the business owner and not owned by everyone else. Today, when you are using ChatGPT or any other tools, AI tools, they are going to your website, to different websites, are crawling websites, are getting content, etc. But it means that the business owner of the website have no control, no power over what the ChatGPT or what the tool is grabbing and collecting and showing to the users. And that's where we should work with companies because they need to have access to this bridge to make able to personalize, to customize, but also to push, promote some content, to promote some product, etc. There is always a benefit of being at the middle being in control of this bridge.

Okay. So you mentioned people don't search for content now, they discover. What's the difference between search and discover?

Let's try, for example, if you take the example of a basic marketplace, before you were knowing exactly that you want to have this Adidas running shoes exactly because there were only 20 of them, different shoes, and this one was absolutely the best. But So you would basically just type in the name of the shoes, the name of the brand, but also the name of the model, etc. That is basically search. You know exactly what you are looking for and you are expecting to have the perfect fit about what you search. When you are in discovery mode, you are most likely like, I want to find the best shoes that have been, that are trendy or that are for my mood. So what you will type is, I want to have shoes for trail. In the mountains, for example, and you want to have the shoes for sure, but you also want to have metadata about why are they good, why are they promoted, what is exceptional with those ones, etc. You want to have additional content and information. So you are here to not only find something, but also discover new content and having more information about those new content.

Okay. I love how you explained that. That makes, that makes perfect sense. So are we saying that that behavior is changing or is it more like supplemental to the search? So I still thinking that I probably gonna know what I want. So I still gonna be searching in a sense of how you explain it. But at the same time that, that discovery piece, which is new to this, you know, current decade, let's say it's something which is exploding, like people are willing to discover more. Is that the right assumption?

The assumption is that we will always need to search. We will always need to discover new content. The big challenge for most of the industry is to be able to do both at the same time, the same interface, because it's two very different approaches. But you should be able to get both use cases into the same user interface, in the same navigation, with the same tool. That's where Mail Search is really good at also, because we have this capability of performing a search or a discovery search into the same call and to promote the right result at the right time. So basically, Mail Search is capable of understanding if the user is specifically searching or discovering content and promoting the right result at the right time. It's one of the key capabilities of Mail Search. It's something we are pretty proud of, but it's also one of the key challenges of the industry.

Okay, so how do you do that? Is that baked into how you search, the words you use, which sort of gives you some type of intent where you can decide that this is a genuine search or this is more of discovery query?

So basically how we do it is that we have been working on what we call full-text search. So basically search being able to answer very precisely to something that has been asked very precisely. We are the best in the domain, so we can clearly identify if a search is not a discovery intent, but more a very precise search intent. And whatever, if there is not enough, if we are not sure about the fact that it's a pure search because the keywords are not here, the content used or the word used are not super precise, etc., about the results, then we can automatically switch our model to less gripe let's grab everything that look like this, that are in this area. If you are just typing shoes for running or for trails, basically we will look at those words in that order, etc. To find if there is any shoes for trail. But if we don't find any results that have as a name shoes for trail, that will never be the case. We will say, okay, perhaps it's because it's not a search, it's more someone that want to discover something. So we move to the other concept.

So you mentioned as well, Quentin, that effectively the search is proxy, you know, between humans and the data sources, right? So it's just an engine to make sure that people find and discover what they're looking for. And you mentioned as well that companies should have control over it. So what we're saying effectively if you want to install the search engine above your data, that you want to have that unique solution rather than relying on global tools like ChatGPT to index your information and to give the answers. Is that the right assumption where this have control comes from?

Yes, but perhaps let me rephrase what it means. Search engine didn't always work or didn't always exist. What we were doing before was just to have SQL databases with all the information and website exposing the content. Why not having just given users or people the ability to do direct SQL query inside the database? Because you will have no control about what they will filter, what they will look, what they will retrieve, etc. We created a search engine to be able to control, to say like, this person could have access to those information, but for this person that is looking for this information, I will promote this content because I have some benefits of doing so. So basically, we created also search engine to simplify the access to the data, but also to control how people will have access to the data. And I think it's exactly the same as we did for humans, but we will have to do it for agents. You don't want to let ChatGPT, just crawling your website and finding whatever results they want, they found, and showcasing it to the users. You want to tell ChatGPT, for this intent, your customer want to have information about shoes for trail. I recommend as a brand, because it's my expertise, I recommend to showcase this specific shoes, this specific shoes and this one, because in my opinion, they are the best one for this query. And as a company, Saying Shoes or whatever other company for any other use case, you want to keep the control over what you deliver to the users. You want to have the best experience for your user at the end. So you want to control what you give.

So how that control looks like in real terms, like in more practical terms? So we had ChatGPT, which effectively, let's think it as Google indexing your website. So you still have your content in the public. But you want to somehow let them know that some of the content which you're sharing online should be attached to certain intents and sort of have that control of making sure that ChatGPT uses it well. Like, how does that look like in practice? Like, how do you go about doing that?

So basically, it's where we are going a bit more in the future because today it's only the premises of this. But today ChatGPT is crawling the website and that's all, just crawling. But we can see that there is more and more people complaining about it because it's starting to cost a lot of money for the website that has been crawled like 10 million times per day. And also they have no control on this, so they are not happy. So what we will see in the future is that we will have more and more limitation for crawler on website, on the different website. They will add some layers of protection. To avoid crawler from ChatGPT and all the others accessing the data. And how I see the things evolving is that we will tell ChatGPT or LLMs, all those AI tools, AI agents, we will let them know, like, if you want to have access to this data, use this source of truth that will be a search engine. If you are familiar with MCP servers, basically what we will say, or it could be MCP or everything else, but we will say to ChatGPT, you should use this MCP MCP server to have access to my information. You should not crawl it because this MCP server will be much better to answer your questions. Okay. And that's how I think it will evolve because it's also how it has evolved in the past with human and data.

Can you give us a little bit what is MCP server for the listeners who don't know?

Yeah, of course. So MCP servers are a protocol invented by Anthropic that is running in cloud. To describe how AI and a tool can work together. Basically, when you are building an MCP, you have the client and the server. The client is basically the LLM, ChatGPT, Claude, whatever, or it could be embedded in any product. You have the servers. The servers are often the source of truth of data. It could be a vector database, it could be a search engine, it could be an SQL database. It could be also an API, for example, for having the weather. They will probably have an MCP with weather channel. That allow you to have the weather in real time. And they have created this protocol to make sure that LLM could have access to real-time information all the time. Because as you know, LLM are trained on data, but they are trained every X months, so they don't have fresh and real-time data. To have fresh and real-time data, they need to have access to API, third-party tool, etc. And that's where MCP It's crazy bridge between the LLMs and real-time data.

But when it comes to MCP, which Anthropic created, that's not a globally recognized sort of.org standard yet, right? So ChatGPT will still agent out and crawl it like Google would. But what we're saying is that later on, likely there will be hopefully a unified approach to AI agents. To access your data and the structure of the data will be based on that framework or, or, or I want to say framework, but it's probably going to be something like notation or something like that, which will tell, you know, how the data is structured, right? So that we sort of trying to predict the future. So, but what happens today, they do come in and they crawl your website. Is there any tips you could give to whoever structures the website and the content that would work better for AI agents?

Definitely, it's a hard topic. It's a very good topic. I think also models are changing along the way, so perhaps an advice that will be given today will not stay for future. But I think the most important is to express as many things as you can. And something that works well for me, search, first is open source because models are using a lot of open source a model to train. So being open source for Mail Search has been a huge benefit for being more promoted by LLMs because they already know everything about Mail Search and they have seen a lot of references of Mail Search on the web. But also something that works well is to create end-to-end guides because what people are looking on ChatGPT or other tools is often how to do something, how to explain something, whatever. They have an intention. So having really intentional docs or pages or whatever is really useful, at least for my search. Because when someone asks ChatGPT how to implement search on my media website, for example, my newspaper website, for example, you need to have the specific content that ChatGPT will be able to use to generate an answer. So you should think like this. What will be not only the question, but what will be the context and what will be the overall question? People will ask ChatGPT and what ChatGPT need to answer those questions.

Okay, I love that. I mean, ultimately you do have people intents, you know, how to do something, what is something, you know, all of these I'm seeking information type of questions or even product or service. So when you generate content for that intent, is there a difference between I'm going to build a huge page cover multiple questions and multiple intents within one page, or is it just better to do a certain amount of information? And I'm coming with this question from the, from the perspective of LLMs working on a, you know, set memory limitations where potentially giving too much content means that some of that content is going to get missed from that page.

Yeah, it's, it's where I would say that I'm perhaps less an expert. Also, because it's changing every day, there is a new model with bigger context, etc. And I don't know exactly, I would say I'm not an expert, so I don't know how to answer the best question. Okay.

I have, I mean, I don't know as well, Quentin, that's why I asked, right? But we had a couple of SEO experts on the show, which, you know, they do a big SEO, we're talking hundreds of thousands keywords and, you know, big profiles and they said this one intent, one page. They basically say it works best if you have a content piece which expresses one intent and that's it. That's just one page. It seems to work better when it comes to ChatGPT respecting that information and giving that page as a source as well. Right. So moving on, let's talk a little bit about no-screen devices. And I just want to see your take on it and based, you know, from Technology of Search as well. So I'm talking about wearables, I'm talking about Amazon Echo, you know, maybe Siri, maybe mobile phone. I actually seen an ad quite recently and I think it was Google pushing Gemini out where it just shows a person talking with the phone and it goes like, ask me anything, you know, how to make a meal or just have a chat with me, right? So these interfaces, they will of course power the search and power discoveries and stuff like that. Do they have, any different approach from typing, or are we saying that everything we're saying just getting translates in a text, and from that point on it's literally the same as using ChatGPT?

It's so far, it's exactly the same. Yeah, there is no difference. It's just that we are translating the, the voice to text, uh, before managing it.

Uh, yeah, and that's it. So, so that means there's no, uh, if you want to enter that market, let's say like this, if you want to be more discovered and more offered in, in these, uh, new, um, conversational type of searches, it's really no— there's no strategy, there's no difference. Just, just do the same as you would want to be discovered by ChatGPT. That's, that's what we're saying.

So it's, um, how to get the question, uh, will stay the same. I believe there is still some very tiny difference, but when it's come very different is more about how do you gives a response. Because when you are giving some text, the human eyes can move across a lot of information very quickly. When you are responding with voice, you don't want to wait 10 minutes before getting the full answers. So you have at least— on the experiment we did, you need at least to explain to the LLM that will answer the response that it will be said, it will be spoken, because it will impact a lot what is written. It will be way more concise, way more straightforward, because you don't want to have an answer that repeats 10 times the same thing or that takes too long to say one precise thing. And when you are talking to an AI with your voice, if the AI can just say yes or no at the end, it's much better because you don't want to wait before being able to ask a new question. Yeah, I would say that the promptness of the answer in, in voice, it's much more important. But for the question, it's almost the same.

Okay. So I got this one, which I want to share. And that, that comes directly from Google Search Central blog. So what they basically saying is that AI-powered features such as SGE, which is Search Generative Experience, which you can see right now in Google, like if you put in some type of query at the very, very top, you're going to see Gemini just giving you an answer, right? Not the, you know, not the links. Links come below, but it will, it will sort of give you the same as you would get from Gemini. So they're saying that prioritized content that clearly answers the question includes structured data or structured data and reflects expertise and trust. And this is why I have this, this, this question about expertise and trust. In your experience, Quentin, like, what is expertise? From the eyes of AI agent?

Very good question. I hope I will not miss the question, but what I will say that at the opposite of Google that is searching across all the internet where you can ask anything, anything possible, with ManySearch, when you are contacting a search engine that is based on business data, website data, marketplace data, whatever, you already have kind of the context. Of why you are doing the search. So that's one of the big difference between Google and Mailsearch is that on Mailsearch, people are already trying to find information that we have. At least we should have somewhere the information, we know that it's here. Google have to find exactly where it is, so they can have some limits or some— it's more complicated for them to clearly notify what exactly is the intention and what exactly they are looking for, because they don't know at all among the massive amount of websites they could answer where would be the best answer. For me, search is easy because if you are asking the question to a knowledge base, for example, for one of your SaaS applications you are using, you will be talking about this SaaS. You will not be asking about something else. So I would say that it's much more easier for us. We don't have so much issues that Google could have.

So it's simply because that your context is bounded, if I want to express it like this, because you have a dataset, it's all in Mailys search. So anything which comes in that is going to be that bubble. But what we're saying, if you talk about like Gemini or ChatGPT, that's literally the whole planet. That's literally the whole internet, right? So for them to give an answer, they need to somehow validate the expertise behind the answer they give?

Yeah, basically it will be the role of Google and ChatGPT to find which search or which, if we keep the search topic, which search to call. But when they reach Mail Search on the website, for example, then Mail Search doesn't have too much to do because we know that we are on a specific context. Then the big issue, mostly when we are building a RAG, is locked is how to make sure that you will only answer about the topic you have and you are into your boundaries and you will not go outside of your boundaries because we are using a lot of LLMs for doing RAG with Mail Search. So you know that all the data that will be used, real-time data is coming from internal data. But then the LLM has been trained with external data. So how do you make sure that if someone asks the chat something political, you will not answer because it's not your role to answer this question. You are— you should be strictly answering about this specific use case and this specific business, and you should not answer any kind of questions. So I would say that the issue for me, such about this, is how to know when to not answer, because you should not always answer.

How does he know when not to answer then?

It's why you have to prompt it or to put some limitation, or sometimes you have to build your own custom model.

So if you're building your custom model, it just never even knows more context. But if you, if you somehow using an open source modules, a pre-trained ones, and you have yourself on top of that, there's a huge possibility that it will, it will take more context from something which is not your content directly, right?

Yes, exactly. So you can also limit the model knowledge. You have some— on some models you have the capability of modifying the weight of the content. So everything that is political or everything that is not on your business, you can just lower it down to make sure that it will never go on one direction. But you can also just prompt it to say— and you can have a pipeline is what we do in some use cases with Message. You can have a pipeline and at some point in the pipeline, validate as an answer with another LLM that will have as a checklist to check if the answer is answering on this bounds and not going out of the bounds.

Okay, I get what you're saying. I get what you're saying. So you mentioned speed as well, especially when we talked about the voice, as in if you have a conversation, you don't want to wait too long. You want a short answer. Which raises this question around speed versus credibility of an answer. So if you think about, I want to get something very, very quickly back, and I'm talking from experience of using some of the faster modules online like ChatGPT, often the speed comes at the cost of actually having a good answer, as in in-depth answer and really thought-out answer. It just becomes like very, you know, surface level. And then you get these answers, you get these, you know, contexts back and you go like, Well, I'm sort of start losing trust in this chat because, you know, you're giving me these sort of vague answers, right? So I sort of feel like the speed comes at the cost of actually having good answers. And if you want to have good answers, it's going to be slow. What's, what's your experience like? What's the trade-off over here?

My experience is that it's very different for everyone. Expectation could be very different. You have some use case where expression will be performance over relevancy, and there is some use cases where it will be relevancy over performance. Most of the case when you are chatting, you will be more willing to having relevancy over performance. But for example, perhaps for voice or for other use cases where you have search inside a SaaS application, or when it's critical to have very quick and iterative interactions, you need to have performance. It could be more important than relevance. So it's really use case dependent. With Mailsearch, we are serving from marketplaces to media to support to SaaS application, and all of them have very different opinions on what is the most important between search and relevance. So our goal for Mailsearch, but I think it's the same for every search engine, is to be able to have some trade-off, to give the ability for customers to choose whether they want to have something precise or to have something that is also fast. At the end, our goal at MainSearch is to be able to align all those needs, so to be able to provide something relevant and fast at the same time. Yeah, it's not magic. We are trying to do our best, and today we have the faster hybrid search engine in the world. But still, if you want to have something smarter, indeed you will have to slow it down a bit.

Yeah, that's a trade-off you have to do, right? So it's a balancing act, I would assume. Okay, so let's say we have, you know, Mailysearch installed, or we can even imagine ChatGPT and Gemini, and all the queries which people are sending towards, you know, that's a goldmine, right? And because it shows so much intent, it shows what people want, what they're looking for. That's a lot of very, very important information which we effectively giving either to the tool such as Mailer Surge or, or, you know, to ChatGPT and Gemini to effectively to them. And, uh, I want to quote this Harvard, uh, Business Review which effectively said customer search data is one of the most overlooked tools for innovation. It reveals real-time intent, pain points, and unmet customer needs. At the same time, I worked with, you know, product teams which don't really utilize that in a sense that they go into the search queries, what people looking at, and go like, look, there's so much gold over here, we're gonna build the roadmap around that. So when it comes to your experience, uh, Quentin, it's like how you could do that. Let's say, you know, I have a mail search, you know, I'm getting all these queries, like what's that, what's the thought process, or maybe an actual workflow where you can extract that and influence what you're actually going to build next or how you're going to serve your customers?

So basically what we are doing today is that we are analyzing every query that are made, every search intent that are made. And we are capable of saying like this person have searched this thing and found this amount of results at this speed. We have a lot of information. And then the goal and what we are doing is that we are just transforming the data. To have reports that could be easily understood by our customers, like knowing what are the searches that have no result. Because today it's like, it's a big issue for people that are doing mostly full-text searches. Like they could have, or people that are using old-school technologies such as Elasticsearch or Postgres or MongoDB, they have tons of searches that have no results. Like they could have 25-30% of their search that have no result. And I'm like, God, it means that you are probably losing 25 or 30% of your money at the end of the month. Like, you should react. And that's our goal, is also to prove that first we could have some leverage to make those 30% go to 5% or even 0% of no response rate. But also we can learn what has been answered that was relevant or not relevant. And we can do it because people are saying with MySearch, they can tell us like this person clicked on this result, or for Marketplace they can say like this product has been bought after having been searched, or whatever the use case. They can basically send custom events telling us what happened after a search so we can correlate all of this and expose this information to the business to give them some insight about how to improve as such. Then The next step for us, it's what will come soon, that we want to automatically convert this information to improve the overall search experience. First by adapting all this content to the overall strategy of the business, saying like, for example, you want to answer always the most sold product by countries. So someone that will search from France in your marketplace, you will answer this result because they are the top selling result, but we want also to use this data to personalize the search experience for each individual because we will be able to categorize them, to know what they have bought before or what is their behavior on the search so far.

Okay, so you mentioned adopted content. So we're basically getting all the queries captured which are getting searched, we're getting some type of insights which comes in the reports. Actually, can we talk a little bit about these reports? Are these like baked into Mailys Search or are these something I would need to do it myself? Are we talking about somehow clustering all the keywords to express a certain intent area? It's like, what are you— how do these reports look like?

So today it's pretty basic. What you have is all information regarding the number of users, the number of searches you got, etc., per day, per hours. Whatever. And on top of that, you have specific report to know what are the search with no result, what are the most searched queries. You can have some, um, uh, it's splitting by countries, those kind of things.

Yeah.

Um, and for the moment, it's pretty like, it's almost profit. Uh, but then we can think about being able to use AI also to labelize the query, to, to make a to analyze the content, the queries, and to make group of queries, like to say, like, overall people are more willing to look for shoes in your marketplace instead of t-shirts. It's something that we might do in the future that is not possible yet. Okay. The biggest things today is that people don't want necessarily to go on MageSearch to have access to all this data. So one of our challenges is also to be able to export this data, to manage it, to compress it, and to send it to another tool. It could be like Mixpanel, Amplitude, Google Analytics, whatever, where they will be able to aggregate this result, this search analytics with all those analytics they have from the product, from the business, etc., to take decisions.

Okay, so it's analytical in the sense that it's statistical, it's numbers, it's what happened. And when it comes to that strategic part of like, how do you look at that report and build your future plans around your product or content or support or service, that part is where you need to do it yourself effectively, right?

Yeah, you can do it yourself. Most of the team are doing it themselves. Most of the team also are discussing with us to know how they can adapt the strategy. So it's something we are doing a lot with our customers today. If they have any question about how to improve the search relevancy or to improve the result for specific queries, we definitely help them improving the search relevancy. It's one of the services we have.

All right, so, so you mentioned this, uh, adaptive content, which I, I really wanted to understand more. So, so we're looking at the report and, uh, and we're seeing something which then leads to us going like, okay, well, that's a good query and we need to change something in the content to match against that query so the results are being shown. Is that, is that what we're saying about adaptive content?

Yes, absolutely. Or for example, if you So it's less and less the case with AI because now with AI it's almost automatic with message, with the AI capabilities of message. But basically it's what people were doing. They were looking like, when people are looking for these shoes, they have no result. Why?

Okay.

We should find another way to promote some result when people are searching this, or we should improve the description of those products to have those keywords in it, those kind of things.

Are we saying this, the, um, so two-part question, Quentin. One is how exactly do you know what to change in the content to make, to make it adapt to the query you want to target? And the second one, considering the semantic capabilities of AI, is it even a problem in like another two years or so?

I don't think so. Today, to be honest, uh, for people that have issues, what we do is just we activate people that have issues are mostly people that are using only the full-text search. So what we do is we just activate the hybrid search and solve almost all the issues. But also what we can do, most of those are images. For example, if they are products on marketplaces, they have images. And perhaps one of the issues with marketplaces is that they are sourcing a lot of data coming from different providers. And so they have sometimes data missing, description missing, they have category missing, those kind of things. So what we can do today is taking all the pictures and using AI also to get information about the product and put back those information into the main description or into the semantic description of the product. But today most of the issue in that range are solved by AI either by just having a hybrid search. Or what we call multimodal search, meaning being able to search not only on the text, on the semantic meaning of the text or of the description, but also the images and stuff.

Okay. Okay. So I still have this sort of pick on Trust a little bit, but maybe it's not directly with like Mailys Search because I think you mentioned that there's no real challenge in that sense because you have a a small contextual window, which is just the business content, just the business product or all that information. But for the bigger players, you know, like ChatGPT and Google, I still have this like, um, a quote which I wanted to share, which is basically, uh, it's from Pew Research and they interviewed people when it comes to AI answers across the globe, you know, not only like ChatGPT or Gemini or other tools, but as well AI agents, which are chatbots within, you know, bigger companies. And don't, they don't really trust. It's they saw only 19% of users fully trust AI answers. And then when they go and ask why, 60% of them said that, well, I don't know where the answer is coming from. And they were basically like, okay, well, if you give me citation, if you give me some type of context where that information came from, maybe I'm going to trust it more. Like, do you see anything around Mailly like that, that's being a challenge? And, you know, your opinion in general when it when it comes to trust to AI-generated content?

So it's not a challenge because it's a solution. By using Maesearch, you are sure to rely on some facts that are inside the database because we are a search engine, finally. We are just not a vector database or whatever. We are a real search engine. So what we answer is precise. Then the issue could come to hallucination. But then you have some some solutions to limit the hallucination with Maesearch. But overall, this is a huge topic for B2C LLMs such as Gemini or ChatGPT. It's not a real issue for us because we always have the data that people can trust. If they are looking for information on a specific website and searching with Maesearch on this specific website, they will see content from the specific website. We will not get content from other websites or whatever. One other thing is that, I don't know if it's mentioned on your thing, but people could also be more and more afraid about just who trained the model. Like, can we trust how the model will behave in front of my question? What if I ask any question that could be political or anything else? That's why with Mail Search, we try to stay agnostic with the model. People can choose whatever models they prefer. We can use OpenAI, we can use Anthropic Claude, or we can use open source one. Today we are really moving more on open source because it's also in our heart. We made the main search open source, so we are also promoting a lot of open source models, LLMs, embeddings, re-ranking models, etc. That are open source, and we definitely think and believe that it's the future. But definitely there is this question about can we trust where the data is coming from? And when you have this capability of just being agnostic and let people choose what models they want to use, it's easier to be trustable.

Okay. Can you talk a little bit about The modules you guys, you mentioned, you know, you open source as well and people can choose. So you're trying to stay agnostic, right? So effectively you're saying like, look, we're not going to make a choice for you, you're going to choose that yourself. But if I have no idea, it's like, do I need to know to be able to choose? It's like, you know, you're delegating that responsibility to me, great, but will I make the right choice?

Yeah. So yeah, Basically, you can have your own choice. The fact is that we have a lot of people asking us, what is the best model in my context? So we just have to understand better the context and we can advise on, is it this model or this model? We have internal benchmark about all the models for each industries, etc. So we can say, yes, for your use case, because you really need to have very high relevancy, we will require to have this proprietary model because it will be more relevant on marketplaces that need to answer in German, French, English, and Japanese. But in that case, you are only in English and you have a very tight piece of information. So perhaps an open source model will be enough. So we have this capability internally to promote or to advise our users or customers about which model to use. But the goal also in the future of our product will be to have a marketplace where we we can clearly define which model to use in which use case. And perhaps also because we have the data already inside MeSearch, being able to advise in advance, like perhaps if you want to move to semantic search or use an LLM, because we know your data, we know what kind of data you need to promote to your customer, we would recommend this one.

Okay, I love that. And you mentioned a marketplace. So you guys trying to create some type of module marketplace which plugs and plays with Mailysearch or just not necessarily Mailysearch, but a general sort of like, you know, Hugging Face type of situation?

No, definitely we are trying to create a marketplace. It will not come in the coming month, but probably before the end of the year where you will be able to choose all your models, but also custom ones that are coming from Hugging Face. So being able to be connected to Hugging Face because what we can see is that today, when it comes to— so there are 3 main kinds of models. There is the embedding models that are transforming a text to a semantic meaning. There is the re-ranking models that are capable of re-ranking results based on the context or stuff like this. And you have the LLMs models that are capable of transcribing results into a message and vice versa. For the embedding model and the LLMs model, generic ones are are good enough. But for the re-ranking model, most of the big companies today have a team dedicated to build their own re-ranking models that will be tailored to their business logic. Airbnb, Netflix, all those companies have a dedicated team of data engineers and ML engineers that are building this re-ranking model. And our goal will be never to build one re-ranking model that works for everyone. It's totally impossible. First, it's going too fast. And second, we will never know as best as the team what is their use case, what is their business challenges, etc. So our goal is to stay agnostic and to tell people, like, we have those free models, you can put whatever you want on those, like, plugs. You can plug whatever model you want. If you want to bring your own models that you built yourself, you can. It's definitely yours. And today, those people are using Hugging Face for storing those models. So we just want to connect to Hugging Face to be able to download a Hugging Face model into our infrastructures and to be able to use them for the search.

That's very nice. I love that. Everyone's going to be waiting for that. Hopefully it goes out quite— based on the time you said, by the end of the year. So I think people will be tuning in. So we got this rapid question situation in the show. So it's like 10 questions, you know, see if you can keep the answer short, but if you want to expand, you know, of course, fine. And these are often the questions which I, you know, wasn't able to sort of put in the main theme of the topic, but it's still very relevant, right? So let's go. What's something most people misunderstand about how AI-powered discovery actually works behind the scenes?

The importance of having well-structured and good data.

So what was the first? The second was good data. The first one?

Structured.

What is that?

So having data that could be manageable because sometimes it's a mess and you have to spend some time enriching the data or making it clean to make it easier to be retrieved.

Okay. So we're basically saying for the people thinking about their own small databases and putting a search on top, serving the customers One thing which you're probably not thinking about is the cleanness of your data.

It's— I would say like the advice, think about what your data, what your user expect and manage that your data to look like what your user will expect.

Okay. It goes back to that data structure, which we discussed in MCPs, isn't it? It's so important. I mean, it's literally the fundamental piece, right? It's your data. Then you have the proxy, as you mentioned, your search engine. And then you have your customer intents, right? Okay. How do you think AI search will shape the way we design interfaces for non-screen devices like wearables or smart home?

Mixed feeling because I think it will not replace entirely screen, but definitely we will have to find new ways to expose data. The question, as I mentioned, is not about how do you express your intention, but how do you receive it? And for me, that's the challenge we have to solve in terms of user experience.

Okay. AI answers are fast, but not always clear about where they come from. We sort of touched a little bit on that. If you had to prioritize, would you focus on speed or transparency to build user trust?

Transparency.

Transparency. How would you do that?

Being able to always get the source. And with exactly the point of the source where the information came.

So effectively, I'm talking with your search engine, I'm talking with Mailysearch, which is AI-powered, I'm getting an answer, and at any point in time I can go like, what did you get that from? And it goes, you know, I got that from these sources and just listing sources.

We can already do it.

Okay.

And we can tell you like it was exactly in this paragraph, in the this pages.

Yeah, I love that. I really hate when I'd work with ChatGPT and I'm like, give me a source. I don't know. It's like, well, what do I do with this now, right? If it's like an email, you don't care, you know, you can just put it in. But if you've— if you're writing something monetizable, let's say like this, like if you're writing maybe a PDF you want to sell about certain things, right, which is very popular right now where people are just reusing the content in different ways and then sort of selling low-ticket type of PDF content. Well, if you can't back this up by the source, it's like, how do you even know what you're writing here is actually valid, right?

It's a challenge.

Yeah. Okay. Do you think every product team should be running prompts through tools like ChatGPT to test how they show up?

What do you mean by how they show up?

Let's say I'm putting what is mainly searched in ChatGPT and I'm getting an answer. Should product teams and services care about what ChatGPT is saying about them?

Not entirely. It's a good insight, not the main one, because as you might know, if you are a product team, you are sometimes in advance about the marketing also. And so if you are in advance against the marketing, it means that you will be also in advance to LLMs, that would be 6 months late, basically. So I say, like, take it as advice, but not as a trust. Okay. Any information you get from ChatGPT, like, just take it, use it, but be careful about, like, don't just take it as 100% pure fact.

Okay. So if we are a product team, we put it in and we see something which we don't like, that might be just old data. Uh, that might be— it probably is old data. How do we go about changing that? Like, what would be your go-to? Let's say you put in Mailysearch, maybe your name, and you get something you don't like, and you go like, that's not true. Like, do you just ignore it because you said, like, you know, like, yes, it's a good reference, but don't you know, don't go over concentrate on that, or would you actually go like, okay, well, if I'm seeing this, that means other people are seeing this. I don't like what is getting served, like, how do I go and fix that?

Uh, it's important to think about what that people are also seeing it, but I would be more willing to think like, okay, if I got this, it's because the probability, uh, means that most of the people think this about me already because it means that the model has been trained with those kind of information, meaning that probably those kind of information are online already. So I would be more willing to know why ChatGPT tells this, why people told this about message in front before, and how do I communicate better in the future to make sure that now people see the value or change their mind or whatever. And say positive things about me.

Yep, love that. How do you explain structured content to non-technical teams who just want to be found with AI?

How do I explain structured content? What do you mean by that?

So we talked a little bit about MCPs and what the future might be when it comes to content being given to AI engines. So in current day, they just go in and crawl the websites. Like, what's the structure of the content should be to make sure that they're getting picked up and represented? Well, effectively, the content is just getting picked up by AI agents.

Yeah, I would say just use Mail Search to index your website. It will be well structured.

So about how you, like, if you index website, right? So Mail Search index like how do you look at the website structure? Do you just follow a normal sort of like HTML SEO guidance or is there something specific where the content just becomes better when it comes to structure to feed into LLM?

So there's many things to do to get relevant data. It's not just like today you can see more and more just AI looking at web pages and try to extract some information. For me, it's one part of the result, but there is something old that's called schema.org that is used by every content website and every marketplace, e-commerce, etc. That is a way to structure content for Google. So just following the, I would say, the old advice is good, but also with Message Search you have the capability, and our crawler is doing this, to split the page into some relevant segments. Because one of the challenges when you search for content in pages is to know exactly what is the part of the content that matches the user needs, and then how to put this into a context that will be probably the full page. But you need to retrieve exactly where the match is, so we have a huge work at splitting content into multiple pieces. Okay, and so that's one of the challenges.

So the old advice of schema.org, that's still like SEO tactics. We basically saying that's still very valid. There's no new ways of serving contents, at least yet. And you as a search engine and likely others are doing the heavy lifting to making sense of whatever is the content being given by, you know, by the public, by the internet, and then putting that into LLM, right?

Yeah, I think instead of if we invent a new standard, there is an LM.txt that could be also useful. We index them, for example, for Maysearch with the crawler. If any— do you know what it is?

No.

No. So basically what you can do is like for every pages you have on your website, you can have the shadow, a shadow page called LM.txt. That is a text a text file that LLM could use as a context to answer a question.

Okay, that's great.

And very useful.

Yeah, that's very useful. So how does that TXT look like? It's just literally a text like this? Like just pure text. Okay. Is that like a standard now or is it just something like a want-to-become?

Kind of.

Kind of. Okay.

Yeah, there is more and more CMS and stuff that are taking this stuff Excellent.

All right. So what one thing companies should stop doing when it comes to getting discovered today? Again, with AI, by AI agents.

Your questions are hard.

Too much content? Maybe as an idea. I had, I'm behind that question. That's what I had. I was like, I was reading this blog post and it was, I don't know, it was like 16 minutes read or something like that. It's just so, so huge. And I was like, I don't think that would, that would, I feel like that would be too much for any sort of like, you know, small memory agent to consume that information to actually be given then, you know, as an answer with ChatGPT. And I was like, I want to ask Quentin if he has any ideas about what you should stop doing.

Definitely what you should do is still content that will interest people. And what you should avoid to do is crap content that will serve only SEO purpose.

There you go.

Because yeah, definitely it could work for a time. It will not scale. And at the end, it can drive some people. But if the content is not interesting, they will just leave and they will have a bad feeling about your content.

Yeah, value-add content. I love that. I'm seeing as well a lot of ChatGPT real written content. You know, like you have these like very like surface-level definitions, you know, no real depth within the content. And you see a lot of that getting just spit out, spit out into internet, and it's getting annoying to some point. So, you know, so I get what you're saying when it comes to quality read and quality content. So in your view, is there any opportunities getting missed when it comes to LLM agents and companies not being on YouTube, not doing podcasts, you know, TikToks and other social media. And where I'm coming with this from is that you are likely producing a lot of content which is not written content, right? You know, like videos and talking like us having conversation right now. Our companies are losing opportunities to actually take that content, somehow structure it for LLM, you know, maybe write some tools and then shove it in. And make it yours?

If the question is about, should we take more carefully about, should we invest more into other than text format for LLMs, the answer is definitely yes, more and more, because all the text format has been indexed, almost all the video format has been indexed by the bigger company yet, but I think the podcast, I don't know if all the podcasts have been indexed by ChatGPT and other stuff, but because they are willing to have more and more content to train the models, you should be on every kind of information. And for example, all the information on Reddit, X, or those kind of websites will be not indexed in the future depending on the models still. But I think that podcasts and YouTube could be a real source of truth for the long term for those models because you have a lot of learnings to get on those. It takes some investment and time to do, at the opposite of just tweeting something. And you have to put your— you are also exposing yourself as a human when you are talking to in a podcast or on YouTube. So definitely I think that they will put more importance in the content that is produced on those kind of format instead of text.

I love you saying that. I share that opinion, by the way. I really do. Right, so personalization shaping future of AI search, helpful or harmful? And what I mean by that is, uh, I do a search and it, you know, it knows who I am, it knows my habits, it probably connects up maybe with even other information sources, what I buy, what I look, what I watch, and then it gives answers through that filter.

I would say both, depending on who is doing it. If you are looking at personalization from OpenAI, not trustable.

Yeah, that's scary, right?

Yeah, that's scary. But if you are looking for personalization in a specific website, for example, I would say it's a good thing because the website is probably just using your information, not sharing this information to other provider or whatever. They are not using it for, for any ads or whatever. We are doing it a lot for MySearch and the data stay on MySearch. We are never selling it to anyone else. Like, we are getting analytics and, and getting personalization for each of the users. The data is never shared. Like, our business is to answer the queries, is not to share the data as other providers.

What if you get acquired by OpenAI?

Hmm. I would probably erase all the content before.

Before you leave. Delete everything.

Yeah. Right. People have to think about this. And that's super important as someone coming from open source. I think that technology is always good when it's on the good hand. Meaning on the hand of everyone, and things are made transparently and all that stuff. So I would say like, look about where is your data, what people are doing with your data, and if you don't know, ask.

Yeah, I love that. How far do you think we are from fully voice-first commerce, like buying something without ever looking at the screen? Far. Far? Like, what, like 2 years, 5 years, 10 years?

It will almost never happen.

It will almost never happen? Okay, give me your thinking.

When you are buying something, you want to see it.

Okay.

So perhaps you will have a 3D representation in front of you, but at the end you need to see it. It's difficult to buy something just on the audio description or something like this. You need to see the product, the good you will buy. If you want to buy a seal on my Talking about my sneakers, if you want to buy your sneakers, you want to see it. If you don't, how do you know what you will receive? So the same for everything, I think.

Okay, so you want to— what about this, like, I'm like, I don't know, on my phone, I'm like, hey, AI agent, you know, I've been Googling yesterday something something, and you know, is anyone selling that? 'Oh yes, you know, ShopX is selling this, has your size because I know stuff about you, you know, price is good, you know, refunds is good, you can just send them back if you don't like in 14 days.' Yeah, okay, buy it.

So first, I don't want this world. I think it will come. I don't know if it will be massive, if it will replace. I think it will come next to traditional way to search and to it will not entirely replace it. It will probably evolve. I don't know in how long it will completely be transformed. Perhaps in the future we'll have chips in our mind that will allow us to see the products that we buy without the screen. In overall, I don't see anything magic in that world. It will not entirely replace existing things.

I feel like big e-commerce brands will try to make it happen because it's just just, you know, it removes so much friction of actually looking for stuff. And if we think about Discoveries, I feel like they probably start, I don't know, pushing some messages or something like that. Like, they will, they will just try to utilize this somehow. And this is where the question came from. But we'll see, we'll see what the future is. All right, so last one, Quentin. If someone is starting from scratch today, no marketing teams, no tools whatsoever, what's the first thing they should be doing to get findable by AI agents?

How is— what is the first thing they should do to be found by the AI agents?

Yeah, let's just change the PO Gemini as an example, or even, you know, MailerSearch if you guys did next thing.

Uh, definitely write content that will be, uh, mindful and that people will want to read. Basically do, uh, everything that was needed to do before. At the difference that today, the content, if you are sharing to the right place, etc., if you master your channel of distribution, it could be indexed quicker by some LLMs. I think it's the same, it's just that the channels that are changing over time and what is interesting to people is also changing. So basically it's the same recipe but with different ingredients.

Okay. I love that. Chain of distribution. I mean, that's, that's probably quite, quite technical for people to understand, but what we're saying, how LLMs, well, AI agents get information is evolving. And there's a, there's a protocols which are coming in place, which you probably wanna understand as well and how they, how they work. So you can, you can get your content in the right structure, in the right shape to, to just get that fed into you. But at the end of the day, the user intents and your content serving these intents, which is the quality of the content, is still there. That didn't change at all. Right. Okay. So before we wrap up content, if people care to know more about Mailysearch, maybe you personally, where they can find you, and if you want to leave our listeners with one thing, what that thing would be.

So first, if they want to find me or Mailysearch, please look either on LinkedIn for myself on GitHub for my research. And if I had one thing to say, it's like, take the control back on your data and don't let LLMs just master the information for you. Like, take back control, think about it, think about how do you want to serve what you are doing, and not be passive in that world.

I love that data is the new gold in the new era, is it? Especially now, everyone was talking about like 10 years ago, you know, data is the new era, that's the new gold. And now when AI came out, it's so easy to see why data is so, so important, right? So, you know how I know Mailstories, by the way, Quentin? Just, I forgot to give that story in the beginning. I was planning to I used to code with Laravel.

Okay.

Which is like PHP framework. And when I say used to, I have, you know, like a coder's background, you know, in my career. And I did Laravel and at some point they had this package which was their search package. So if you want to get a search feature within what you're building with Laravel framework, you could pull in different tools. And I think by default they were using Elasticsearch. And at some point they switched to Mailysearch, which was like default engine. They would pull in with the package. This is how I knew Mail Research. That's years ago. So I think it's like 6 or 7 years ago. So it was very cool, you know, to see you coming on the show. It just reminded me of all good times, right? So, yep.

No, I just say awesome. Very cool.

Yeah, I'm glad to see you guys, you know, doing so well. I'm glad to see you guys do so well, right? So everyone tuning in today, uh, if you love the episode, do me a favor, hit that subscribe button so you don't miss anything we release in the future. If this episode got you thinking about your data, of how you enable search on your website or on your data with AI-powered features, you know, Mailysearch is something you probably want to look at. There's some other tips which we shared today. So if it made you think, share this episode with whoever you think should, should listen this to. As well. And if you have your challenges, growth challenges, give, give that to me at masteringtechgrowth.com. I really want to know what you're working on and especially what keeps you up at night. And as always, until the next time, stay safe and see you online.