---
title: "AI Implementation Strategies in E-commerce with Damian Sałkowski"
description: "Senuto CEO Damian Sałkowski explains why AI in e-commerce starts with a process audit and an API-based stack, not with ChatGPT alone."
podcast: "Globalne Horyzonty"
host: "Rafał Moszkowców"
guest:
  name: "Damian Sałkowski"
  role: "CEO of Senuto, partner at the agency Vestigio"
  company: "Senuto"
topics:
  - "AI in SEO"
  - "e-commerce"
  - "process automation"
  - "AI agents and the future of shopping"
  - "the future of search"
duration_min: 71
published_at: 2025-05-30
video_url: "https://www.youtube.com/watch?v=l36f5-AUs68"
url: "https://moszkowcow.com/podcast/damian-salkowski-ai-in-ecommerce-seo/"
---

# AI Implementation Strategies in E-commerce with Damian Sałkowski

## About the guest

- **Who he is:** Damian Sałkowski, a pioneer of Polish SEO, since 2014 CEO of Senuto, a platform for visibility analysis and content optimization used by users around the world. He is also a partner at the agency Vestigio and co-runs AI education (Sensai Academy).
- **What he does:** He combines traditional SEO with the possibilities of artificial intelligence, implements AI at Senuto (customer service, process automation, mini SaaS products, enterprise products) and in his clients' e-commerce projects.
- **What we talk about:** Where to start implementing AI in SEO and e-commerce, why ChatGPT is not enough, what technology stack a team needs, what can be automated in e-commerce (descriptions, categories, linking, search, customer service), shopping agents, failed implementations, Google's approach to AI content and the future of SEO.
- **Where to find him:** LinkedIn (Damian Sałkowski), Sensai Academy and its newsletter, and a weekly YouTube show he runs with Robert Niechciał.

## Best quotes

> “Our margin should be built on the efficiency of our processes.”

> “Whoever treats a language model as a source of knowledge and not a knowledge processor is doomed to those hallucinations.”

> “It is not a question of whether, but when, and the advantage lies not in whether we use it, but in how we use it.”

> “Let us make the solution, the problem will be found.”

## Transcript

Host: Rafał Moszkowców. Guest: Damian Sałkowski.

### Introduction

**Rafał:** Hi everyone. Today on Globalne Horyzonty I have the pleasure of hosting Damian Sałkowski, a true pioneer of Polish SEO. Since 2014 Damian has been CEO of Senuto, a platform supporting content optimization, used by thousands of users around the world. Everyone who does SEO in Poland surely knows Senuto. At the same time he acts as a partner at the agency Vestigio and tirelessly combines traditional SEO with the possibilities that artificial intelligence brings and gives. Damian, welcome to the podcast.

**Damian:** Hi Rafał, thank you for the invitation and thank you for such a nice introduction.

**Rafał:** You are welcome. I could introduce you like that for probably a good 10 more minutes, because you have a lot of achievements and keep doing new things, so I had to shorten and shorten. Damian, I have an idea that today we talk a lot about AI, but from the perspective of an e-commerce shop that wants to implement it. Imagine I run an e-commerce shop and see that everyone is talking about AI in SEO. I would like to treat this episode as a practical guide: what I need to know, what I should do, how to effectively use AI in my SEO strategy. Do you think that is a good idea?

**Damian:** Yes, let us try to tackle it. Sounds good.

### Why is a strategy needed before implementing AI in SEO?

**Rafał:** I wanted to start with what I called strategy before implementing AI and how to do it wisely, because that you can do it stupidly everyone probably knows. Let us start with why, because the fact that everyone uses AI is not really a strategy.

**Damian:** First of all it is worth understanding that AI, like any tool, can be used well and badly, and most often if we start doing it without a plan and without a strategy, we will most likely use the bad methods. In most cases I notice simply an attempt to use ChatGPT or Claude for everything, whether describing categories or products. That is not the way. This strategy is needed so as not to fall into the trap of the ease of AI, because the promise on the market is that AI will do everything for you, and that is not quite what it looks like in reality. It requires some collaboration with artificial intelligence. The strategy is there so as not to fall into the trap of promises made by various wizards who appeared on the market. You know how a new market shrouded in mystery gets created, as SEO once was: usually between the recipient and the sender there is a black box containing who knows what. You have to peek into that black box a bit, understand how it works, and only then implement anything, because long term we do not yet know what effects low-quality work by artificial intelligence will bring. Google keeps changing its mind in this respect. A few years ago AI-generated content was something Google did not want. Later that narrative changed and for Google that AI content was okay as long as it was well and correctly written. Recently, this year in April, signals reached us that Google is again updating its guidelines for quality raters, meaning people who assess the quality of search results, and in those recommendations there is special attention to the quality of content delivered by AI. Google indicates that if that content does not bring new expertise, maybe it is not worth indexing. It is worth looking at this long term and anticipating what decisions Google may make here, and I think a few more changes may lie ahead.

**Rafał:** Good that you referred to today's AI market being like SEO once was: there were loads of shamans and around it lots of dark tactics. That is why I think a strategy is needed and an understanding of why we do it, what for and into what to lay it out as a plan.

### Where to start with AI in a company? With a process audit

**Rafał:** Moving to the plan: how, in your opinion, should a plan for implementing AI in SEO look, step by step?

**Damian:** First of all I think this is a problem of many organizations, it touches even mine. I observed that we often do not really know how we do certain things. There is no place where we have written down how certain processes are carried out, step by step. At the beginning it is very hard to automate them, because we do not know de facto what, step by step, leads from the beginning to the end of completing a task. So the first step I would take is a process audit and checking whether we know and have control over how given work functions in the current structure. Let me use an example. Suppose we have a problem to solve: automating the creation of product descriptions. It is a common problem in e-commerce, because most often those descriptions come from XML files pulled from wholesalers, suppliers or manufacturers, and the scale of products is too large to create those descriptions manually. To automate product descriptions, we first have to understand how we have done the process so far. Do we hire a copywriter, how does that copywriter create descriptions, where do they get data from. It is worth understanding each step and checking whether AI can help us. That is the first step I would start with.

**Rafał:** I think that makes sense. That sometimes we do not know what processes we have and whether we have them written down at all is a pain for most companies. I often see the desire: let us make a workflow, let us make agents, let us automate. And the question is: automate what? If we do not have it written down and tested, and everyone in the company does it a bit differently. I saw a situation in e-commerce where the marketing director and the marketing team, every person including the director, did the same things in three different ways and it was impossible to implement anything there.

**Damian:** Notice that we work in a somewhat creative industry, so there are many ways to achieve the same goal. Every task can be done in many different ways. That is why I think it will be very hard in Poland to implement artificial intelligence in organizations, because most of ours are small and medium companies that rest on human capital and human minds and have very few processes. Probably there will still be work for people, and that is probably your specialty, to understand how an organization works and describe it in some way. Artificial intelligence cannot do that yet. Maybe soon we will be able to walk through a company with a camera, see what everyone does, and that intelligence will understand it. For now we need to provide information on how certain processes work and how they should be carried out. And it is worth understanding that a big problem is a sum of small problems. If we can lead that sum of small problems to the big sum, it is a good starting point for thinking about automation. I see that AI is implemented very well among employees: it is very easy for employees to start using AI, but the organization is not so simple, precisely because it consists of many processes not quite described, and it is very hard to express those automations at such a high level. From my own team I see that the adoption of AI for independent work is very high. Take programmers: if we show them, for example, Cursor, a tool for programming, most of them absorb it easily.

**Rafał:** An interesting and right perspective. On one hand it will be hard for us to implement AI in Poland because of small companies, but maybe on the other hand that will force those companies to professionalize operations and consequently implement AI. It is a bit of a race: whoever does it faster and more effectively will have competitive advantages.

**Damian:** Although I point out that it changes so fast that I would not do super hard implementations. I would try to approach them softly, because maybe what we implement and spend several months on will be completely out of date in those several months, because the problems we had have been solved in another way. You can on one hand chase those solutions and implement them on an ongoing basis, or organize a big AI implementation project for several months, which later may turn out not quite right. It is a market trap: we would like to implement it, but it changes so fast that it is not exactly clear what to implement.

**Rafał:** Before you finish implementing, it may turn out it is already inadequate, because it can be solved another way. Very good advice.

**Damian:** It also leads to the thought that one-person companies will have a big advantage, because there you can implement very quickly and change very quickly. In a large organization you can implement a lot and have everything described, but implementation takes a lot of time. In a mid-sized organization you can quickly implement something, but you do not know what, because you do not have it described. Both of those types of organization are a kind of trap. An organization that consists of one, maybe at most five people, will be able to adapt very fast and change directions of development. Maybe that is a model of the future. I also try to defragment my organization a bit, divide it into smaller interdisciplinary teams, so as not to have to operate on such a large organism and change its direction, because the bigger the ship, the greater the inertia. We drifted off the SEO topic a bit, but I hope it was needed to give a picture of how I think about it.

### How does Senuto implement AI in its organisation and products?

**Rafał:** Absolutely, because you yourself brought us to Senuto. I prepared questions about how you approached implementing AI in Senuto's tools. In which direction is Senuto going? Senuto has you, so it is easier, right?

**Damian:** There are many vectors of AI implementation, so one by one. Implementing AI in the organization as regards employees and people involved: here we have full freedom. Everyone can use the AI tools they want to use. There is great permission and encouragement to test different solutions and solve existing problems with artificial intelligence, especially when building technology, where we can use Cursor or other tools that speed up programming. At the level of employees we introduce no mandates, but a lot of encouragement, we hold internal trainings, teach each other, and that is the key to strengthening the productivity of individuals. As for the organization, we first went toward customer service, because it is fairly repetitive and simple to automate. We decided on a native solution from Intercom: Intercom has a chatbot called Fin. We could build our own chatbot and it is not hard, especially now, but we used what we had natively. Now, if my memory serves, I do not want to give distorted numbers, over 50 percent of requests are handled by that bot and satisfaction on problems solved by the bot does not fall especially. It is an area we managed to put in order reasonably with AI and we are satisfied. At the level of process automation we mainly use the tool Make: within one infrastructure it connects 1000 tools and between them we can design linear flows, meaning data flowing from one tool triggers actions in another. It has a simple interface, so basically anyone can use it, and it helps us optimize small processes, like: someone signs up for a webinar, we add them to a mailing list, operations in a CRM. The most popular now, and you use it, is n8n. In our organization a bit less, because it is more technical and harder to adopt among less technical people. It is very easy for us to implement AI among employees, a bit harder at the organizational level. There is a range of solutions we could apply, but because Senuto is a dynamic organization that very often changes development directions, we do not really know what we do day to day, meaning there are not many repetitive operations. We have the basics: everyone has meeting transcription enabled, we use Fireflies, so all meetings in the organization are transcribed and anyone can return to them. We implemented such basics, but being able to set the organization on autopilot, I cannot say that yet.

**Damian:** In the product itself the perspective is interesting, because we carry a 10-year technological debt that we built up. We did not design this application with the idea that AI would later develop it. It is very hard for us to make fundamental changes in our main application. So we went sideways and opened other vectors of development. The first is so-called mini SaaS products, as we called them: an internal accelerator of mini startups, where anyone in the organization can start their own small team and release a product that solves a single problem. We released a product for fetching SERPs, a simple API. Many people now build processes in n8n for writing articles, so they can buy that API and connect it to their processes. We are also releasing a product called Human TXT, which serves to humanize text written by artificial intelligence, and we have four more products in the queue. We want to release new products and test the market that way, and those that pass a six-week incubation phase we will develop further. The products are made in small, 3-5-person interdisciplinary teams with heavy use of AI, basically artificial intelligence creates everything there and we only control it, and it is a bit of an escape from the technological debt we incurred. The second area we decided to open is Enterprise products, because there is a great deal of software on the market and there will be more, now basically anyone can build software over a weekend if they have basic technology knowledge. We decided that very many clients, instead of software, will look for results. We set up an Enterprise unit where we use artificial intelligence and our data, our unique combination, to deliver insights, reports and operations for clients. The application that has been the core product for 10 years has its rhythm and we cannot turn its direction overnight. Mini SaaS and Enterprise are the areas we want to explore now.

**Rafał:** On mini SaaS, that is the incredible change AI brought: we can quickly set up and verify solutions, solve micro problems with micro tools. I have a feeling it is somewhat the future.

**Damian:** Yes, but I have a mixed opinion here. On one hand the tool market had professionalized: you gained trust that when you register in a SaaS, someone invested money, did it properly. Now the market has turned into a wild west: you register in a tool and do not know what level of security there is, you do not know whether someone made it for 5 dollars, you do not know where your data will end up. Maybe it is a change for the worse, because people will gain distance. Code created by AI carries many dangers related to security, and I have heard many times that data leaked from such applications. We will see what time brings. That is why we launched the Enterprise unit, because maybe in the future people will want results, not interfaces like now. And on the other hand mini SaaS products capitalize a bit on the data we have and its enormity, because that is something AI will not do: it will not fetch data at scale and will not design the architecture of a database to store it. There are no-code tools like Make and n8n, so instead of tools people need only a specific endpoint to automate their processes.

### Why does AI in SEO need an API-based stack, not just ChatGPT?

**Rafał:** I wanted to move to the section Go AI, but stay human, in a way. How, in your opinion, should an SEO, content, marketing team be prepared to work effectively with AI?

**Damian:** The first step is choosing the technology stack. The team that will implement AI should not do it only and exclusively in ChatGPT. The first step: give the team access to the technology, and what will be needed 100 percent are API keys for services like OpenAI, AI Studio from Google or Claude. I recommend the tool OpenRouter, which within one API structure gives access to all language models at the prices at which we would get them from providers, and we use it most often. The team implementing AI will use mainly the API and not chat interfaces. Chat interfaces serve everyday work, ideation, creating plans, but for repeatable processes we will not use them. The second element is a decision: do we go with code or no-code solutions. That is a big discussion. On one side we have no-code like Make or n8n, which reign in this area, and on the other code solutions, meaning written code that does the same, in some programming environment, most often Python. Both no-code is okay and code is okay, because even a non-technical person will be able to write Python code and run it. The pluses of no-code solutions are speed, because we will definitely build an automation in n8n faster than generating Python scripts. On the side of Python stands flexibility: we can basically do everything and no limits touch us, while for example in Make we have timeouts and so on. The next element is a database that will store data and operations. We can store it in Google Sheets if we have a small set of operations, but remember that Sheets works within our computer's memory, so after exceeding a certain amount of data it will stop working, not effective at large scale. We can use no-code databases like Airtable or NocoDB, open source, or classic databases like MySQL, or now the very popular Supabase, a layer over Postgres. In most cases it is worth the team having access to a computing cloud where such services can be run. Lately Google's cloud works very well for me: we can run a database there, so our automations can pass data, and also scripts and applications. In Google's cloud there are services like Cloud Run, which lets you run web applications, Google Colab, which lets you run Python scripts, Cloud Scheduler, which lets you run processes on a schedule. If we have all this and have written down the processes we want to automate, we have everything to start. The basic step is building this technology stack, and along the way several important decisions must be made, because they will affect whether the work runs uninterrupted or whether we spend a lot of time servicing what we built.

**Rafał:** That is a comprehensive guide. You give the lie to the common claims that having a ChatGPT license is enough and solves the problem. To use AI at any scale and of good quality, you need many tools and to prepare sensibly.

**Damian:** I have a comparison. ChatGPT has "chat" in its name for a reason. It is a solution created for chatting, meaning natural conversation between a human and a machine. If you use the chat, you have surely noticed that it tries to start a dialogue with you. It can do research, summarize information, but at the end of the day it is a dialogue. The processes we perform in organizations usually do not consist of that. They consist of processing data, moving it from one place to another, changing a status based on an action in one place. These solutions are not designed for that, so we need the API to be able to steer the behavior of the language model, which is why we need the whole stack. It is very hard to do such process work in ChatGPT, because we have to instruct it step by step what to do, and each time it may do it differently, because language models are non-deterministic. If we break those small problems down and, within automation, design it step by step, we have control that the process will be performed the same way each time.

**Rafał:** That is the important matter of repeatability, for example in translations, transcreation. Through a chat created for interaction between a machine and a human you cannot easily get a repeatable effect at scale.

**Damian:** It also results from the fact that language models have parameters we can steer. Let me give one example: the temperature parameter. GPT models, General Pre-trained Transformer, are the transformer technology Google published in 2017. They work by having taken in lots of data from the internet and on that basis, very simplified, being able to predict with probability calculus what the next token should be. If we have the sentence "my favorite animal is", on the internet it will most often be a dog, then a cat, a tiger much less, the probability of a tiger is low. In ChatGPT we cannot steer that temperature. Depending on what we type, for example "be creative" or "be consistent", ChatGPT will set the temperature somehow. If it sets it high, the randomness of tokens will be high, meaning it will have a greater tendency to choose tiger rather than cat. If we set it low, for example say "be consistent", it will always choose dog. Between you and the model there is a system that may draw that set of parameters slightly differently. With the API we have full control: sending a query directly to the model, we can set the temperature at 0 or for example 0.7 if we want the model to be more creative. Of course it is not about creativity but a tendency to choose less obvious tokens, but that is how you can extrapolate it.

### Does AI give a competitive edge in SEO, or only help you survive?

**Rafał:** Do you think it is possible to build a competitive advantage based on AI in SEO?

**Damian:** Definitely. I recently spoke at the I Love Marketing conference and concluded there that it is not even about building a competitive advantage with artificial intelligence, but fighting for survival with artificial intelligence, because maybe the players that do not do it will not survive the next change. I used a reference to the economics of content. Thanks to AI we have the ability to create an unlimited amount of content, so supply is very high. There is also the demand side. First, people have other solutions for finding information, ChatGPT, Perplexity, and in the future will not use only Google, and they also have a tendency to watch video. Third, there is AI Overview, which strongly changes demand for content. Demand falls, supply grows strongly. Those two forces, which should go in the same direction, go in completely different ones, so if you want to survive and be able to generate benefits from generating content, you have to do it at large scale and very cheaply. If you can drive supply at a very low cost, then demand does not quite interest you, because the cost curve is at the same level. If you stay at the same cost of creating content, demand for it is so low that the profitability of that investment will fall strongly over time.

**Rafał:** So let us not build a competitive advantage but try to survive.

**Damian:** Using artificial intelligence is necessary: either we will want to do it or the situation will force us. We can build competitive advantage based on how we do it, how we use AI.

**Rafał:** That is the clue. In the context of return on investment in content: on my own example I will say that within three years ROI fell by 100 percent, basically, 99.9 percent. So what you said about supply and demand, absolutely. I see first-hand how much more content there is: working with sites in various niches in English, the internet is absolutely flooded. The Helpful Content Update, which came out in I think 3 or 4 iterations, cleaned it up hard. The name itself indicates it cut content that brought no additional expertise. Many of those copywriters who wrote that content today unfortunately have no work, because they did it at an average level and that content is no longer needed by anyone. I do not want to go into what they should do in life to be useful, because everyone should deduce their own path. AI does not end the adventure with content, it fundamentally changes the playground.

**Damian:** An interesting perspective. Maybe the gravedigger of SEO is not AI, but for example video, because people like to consume video content, so such copywriters should widen their horizons as to the type of content they create. Returning to the question: do we need to build a competitive advantage based on AI? Yes, we will be forced to. It is not a question of whether, but when, and the advantage lies not in whether we use it, but in how we use it.

### How can AI turn one piece of content into many formats?

**Rafał:** I will add a small digression in the context of video. Video absolutely eats traffic. You mentioned AI Overview, behavioral factors, people consume content differently. We, in the SEO industry, got used to SEO being Google, because Google has a gigantic market advantage, but that is changing toward other formats. I would encourage agencies and content people to think that SEO is not only written content, and if you write great content, it is very easy to make a podcast, a video out of it, with little work. From one input of work into content you can make several formats of engaging content adapted to how people want to receive it. It is not about flooding the internet with low-quality video, but that what AI makes today can be very high quality and adapted to how people consume.

**Damian:** So far in SEO we tried to squeeze one type of content into various intents and ways of consuming it, and AI is great at reformatting content. When you write an article, you can generate an infographic or ask the model to generate, based on the content, charts in JavaScript that you place on the page, so instead of a wall of text the page can have many charts and tell a story with that data. You can make a transcript into a video, record shorts, a podcast, and it will not take much time if you automate it properly. What you said can be called a process of content recycling: what we did once, one piece of work, is worth reformatting into various message formats and in this way reaching users in many places. SEO should consist of trying to be where our users are, and not looking for them where we chose to look.

**Rafał:** Exactly. SEO is not keyword positions in a search engine, which is often forgotten.

**Damian:** If our listeners are familiar with the SEO industry, then many new names with prefixes will probably arise: LLM SEO, AI Overview SEO and so on. Do not let marketing headlines mislead you. There is nothing wrong with creating such headlines, it is standard practice to find a market position. Once many agencies grew by starting to offer white hat SEO services when Penguin came in 2013, and that was okay too. But the only path is a multiplatform, broad approach to fulfilling intent with different forms of content, not only the text layer.

**Rafał:** That is how I try to approach it: not only multilingual but in different formats, so the return on investment in a piece of content is reasonable, because from just a piece of written content it is minimal. You said about visualization that AI can be used to create vectors, graphics, interactive ones, there is a brilliant field to show off there. Even the Wall Street Journal, certainly not 100 percent with AI, but many publishers in the world could look at how that organization makes content: great written content, and for those who do not read, audio right away, the content visualized everywhere in condensed form, adapted to the user without losing the substance.

**Damian:** If we have a text layer, creating a podcast from it is no problem. ElevenLabs is enough, which will generate a podcast from the content, we can even use our own voice. There is NotebookLM, where you put in text and it generates a podcast, and they have reached the point that you can talk with the AI hosts, enter a discussion. Any text can be automatically processed into such a form. You can come up with various forms of conveying the information: quizzes, widgets with described data, polls. Considering the direction Google is going, meaning those new guidelines for quality raters from April, Google will expect you to give something more than AI-generated text. Even though it will be superbly generated, in the near future it will not be enough, because Google gets too much of that content, indexing costs, a number of problems arise. You have to approach it creatively and think how that content could be presented differently so the user satisfies their need and does not necessarily throw themselves at a wall of text.

**Rafał:** Paradoxically there will be an advantage of quality over quantity, in my opinion, and people will stop making blogs for the sake of blogs. My favorite SEO advice, which still persists: sir, you must have a blog. What for? I do not know, but you must have a blog, because it is good for SEO. I hope it dies a natural death. And what you nicely said: we must know our recipient. I often meet companies that do not know the recipients of their site, do not know who they are writing for. They know roughly that they are between 20 and 45, but not how they consume content, what problems they have that we can solve with articles or videos. You have to make the investment at the start and later fit the implementation to it, not do the implementation because we must and then wonder how it resonates with the market. That is a recipe for burning a haystack of money and for the classic "AI, SEO does not work".

**Damian:** It used to be simpler, because you built a set of keywords they searched for, so you did not have to be interested in who they are, because they had a defined need. You classified the keywords and answered with your offer. If we approach multiplatform, on every platform a slightly different language is needed, you communicate with each group differently, so it becomes all the more important.

### How do you measure the effectiveness of AI in e-commerce?

**Rafał:** How would you suggest measuring the effectiveness of AI implementation in an SEO strategy, for example in e-commerce?

**Damian:** The function of cost and time, above all. I suspect artificial intelligence will not give us better results than a human, but will give definitely broader possibilities for automating operations and definitely lower cost. If we can get the same effect at lower cost, that is the success we should look for. Our margin should be built on the efficiency of our processes, not on how much value we can deliver from that search engine.

**Rafał:** Besides, we can test and experiment more often, almost for free.

### What can be automated in e-commerce with AI?

**Damian:** Let us go to what in e-commerce can be automated with AI, so we can refer to examples. I will use examples of our implementations for e-commerce clients and products we try to make for e-commerce.

**Rafał:** Good that you went straight to it, because I was ready to dive into further philosophical digressions. I have a point: the most effective AI implementations in SEO in e-commerce that you observe or had a chance to do.

**Damian:** I will divide it into several types. The first is the text layer: we can automate fully product descriptions, category descriptions and guide texts. For product descriptions you have to build a process that learns as much as possible about the product. A few tips: based on the SKU code or another universal product code, it is worth crawling the internet and seeing what features we can gather. If we have for example an iPhone 16, which has an SKU, in other shops it will also have that SKU, so we launch a crawl that pulls that data, and with prompts we extract the features the product has. With multimodal models you can put photos of the product into the model: if we are a furniture manufacturer, we can throw photos at AI and it will describe the features, colour and so on. When we have the set of features, we automate product descriptions one by one. The second thing is category descriptions. Some shops use them, some do not, but if we decide to automate them, we most often study which products the shop has in a given category, because for example car accessories at OBI and at Leroy Merlin are two different pages: there they sell hubcaps, here air fresheners. A category description should refer to the products we have in that category, so the most important thing is to gather all the products from the category and put them in the description. These two text layers carry very little risk, because we do not make Google's index swell, we complete pages we already had in the index, which is the safest place to test AI-generated descriptions. Guides are another element: if we have a category axes, we can build a process that tells which guides about axes users search for, and complete them. We do it for many clients and it works.

In the information architecture layer we can do very much, starting with internal linking. We recently did a project for Answear, a fashion shop. In some shops you have similar categories: if there are green Adidas shoes, at the bottom a link to red Adidas shoes appears. Such shops use boxes with internal links, because internal linking affects the internal page rank of the site and shows Google which pages are more important. Using the semantics that AI provides, it is easy to determine pages similar to the one we are on. At Answear we completely automated the choice of those boxes: to each category we add 20 similar categories with an optimal anchor based on visibility from Search Console and our internal data. The second thing in information architecture is studying which categories we are missing, because we have a set of products and they can be categorized in different ways, and in most shops some category is missing. For example shoes: some serve for walking in the sea so as not to get cut, so people search for a category "water shoes". Answear's information architecture does not provide for such a category, and there are a few thousand such categories in this shop where it would be worth creating a static page with products answering that specific user need. With AI and its semantics we can do it: we take all keywords connected with shoes, filter those where our competition does not appear, then with embedding models and similar techniques we do clustering, then we can turn all our products into embeddings and compare whether we have products that correspond to given categories. We are also implementing that and it works great, so entirely with AI we can manage information architecture.

The third thing, and very important: in most shops the search engines do not work great. The most common problem is the internal search engine. Here we can use embedding search engines, a total semantic engine: turn products and their descriptions into vectors with embedding models, store them in a vector database, and when a user asks a question, vectorize it too and match products. These internal search results often strongly raise conversion. Finally customer service: we can automate that fully with AI. I will just point to concepts like RAG, Retrieval Augmented Generation, embeddings, tools like Dify that can generate chatbots. Recently Google showed, if there are notes for this podcast, I can add it, a chatbot that has access to a camera, a microphone and an internal knowledge base and can steer the shop. In the demo a guy had a flower, showed it to the camera, asked the chatbot whether the shop has complementary products or what fertilizer to treat that flower with. Google put it on GitHub, so you can download it and use it yourself. Those are the applications we implement daily in large, small and medium e-commerce shops and which work great for us.

**Rafał:** Brilliant, especially in the context of internal linking, because it is always a gigantic problem at high volume: handling linking so that it is linked not only for SEO but for conversion. SEOs always implemented it somehow, but in my opinion often not quite sensibly.

**Damian:** We automated it completely, so it takes 10 minutes now, the whole site. It took some work to tune the system, but I think anyone using the technologies I listed can do it.

**Rafał:** With building competitive advantage this plays right into it: advantage not from implementation itself, only how and what we implement. Send the links, we will add them. RAG, Dify I know, but what Google released I was not aware of. I saw something similar in HeyGen, a tool mainly for video and audio translation and for making UGC, but it is not that. I am surprised Google put it out as open source.

**Damian:** Google is working hard in open source now, they simply slept through the moment on the market, now they want to make it up by giving everything away for free.

### How might shopping agents change e-commerce and SEO?

**Damian:** I imagine that in e-commerce an agent will arise that, like Ceneo, has access to all shops: a personal agent with access to your camera and all your orders. You can record for it, say, a video from home and it will order the right products in the right shops.

**Rafał:** A tempting and frightening vision at the same time.

**Damian:** InPost is doing something like that too. In the States a few years ago a product like OneBasket arose, I do not know whether that is exactly its name: in many shops you have the same basket and make the transaction in one place. You can be on Zalando's page and on Answear's page, but have the same basket. There is an external company that supplies that basket, and InPost is implementing it now too. Maybe if InPost manages to connect all shops in one basket, they might also be able to implement an agent in the future. I also raised the topic that maybe in the context of SEO the future is not optimizing for language models, because I do not believe that in the long run we will receive traffic from ChatGPT. It is a trend now, but soon in the era of agents it will not matter. Maybe it is worth optimizing for the agent's decision. I imagine the future like this: I need water filters, I am not interested in where I order them. I will have an application with access to a certain number of shops, I will say what I need or take a photo, and the agent will decide itself which shop to enter and buy from. It is worth anticipating what we can do in our shop so the agent is able to choose us. A few issues: if we have an API in the shop, it is worth having it written in one TXT file for agents: the agent enters the shop, gets the API as a TXT file and knows how to perform various operations with that API. Two: when we write descriptions or describe our product, we must strongly make sure the language model understands what distinguishes this product from others on the market. Probably that is how it will decide. Three: in the context of e-commerce, rankings of various kinds, a ranking of washing machines, may return to favor. Language models will not know how those rankings are made, but they may matter, because from the text alone it is hard for a model to infer what is better. Probably it will often use such rankings, so probably the affiliate model may return to favor a bit too, and that direction seems interesting to explore: optimizing for the agents' decision. It is hard to say how that world of agents will be built: whether a company like InPost will arise that links all baskets into one, and you will be a shipping warehouse, or whether there will be many applications with agents and you will have to build your shop so the agent can make a transaction in it.

**Rafał:** An interesting perspective. It somewhat resembles voice search in the States and attempts to optimize shops for it. To this day there are countries like Australia where, I think, 26 percent of users use voice search through Alexa and Google Home. I imagine it would be ideal if through my Google Home I could buy certain things through an agent.

**Damian:** It will be the first hands-off technology. If you have a device like glasses with a camera and the agent knows your diet, when you open the fridge it will automatically decide which products you are running out of and order them. Of course it is a futuristic vision, but remember that the world changes faster than so far and that vision may appear sooner than we think. Of course the level of trust in such products is very important, but who would have thought three years ago that we would be in such a place today.

**Rafał:** I certainly would not, and I think I will have to publish this episode very quickly, because in 4 weeks it may be too late. Half joking, half serious.

### Whom should you follow to learn AI implementation in SEO?

**Rafał:** We talked about how you implement with your team. Outside your yard, have you seen the most effective AI implementation in SEO, someone doing it spectacularly on many fronts? It need not be in Poland.

**Damian:** I like to observe Dejan Petrovic, a specialist who I think is in Australia. He produces a lot of software to automate processes related to SEO and AI. It is worth following him. He is very technical and uses very technical language, but if we read him carefully, we can take a lot from him. As for AI implementations in SEO, 99.9 percent of people focused on automatic generation of content at large scale. As for a broad set of applications, I do not have such knowledge, because the vectors are two. First, the industry threw itself at no-code solutions like n8n and various solutions arise: agencies and SEO specialists did with automations what they used to do in Google Sheets, they do nothing differently, only automatically. And the second area is AI content and there is a great deal of it. What I said on the Answear example is not so visible yet, but probably it will go that way: dealing with the information architecture of the site, putting in order what exists, because if demand for content falls, they will move away somewhat from content, deal with what they already have and try to optimize it.

**Rafał:** I have seen Dejan somewhere, but he operates in ultra technical language, so you have to be strongly in the topic and keep up, because if you stop looking at him for a few weeks, it is hard to return if you are not an engineer or programmer.

**Damian:** And there is a lot of noise around AI. There are many people who deal with circulating news without quite understanding it, causing a big FOMO effect. There is Julian Goldie, whom I know personally, who adopted a FOMO tactic: when he publishes something, everything is insane. There is nothing that was not insane in the last week. Because of that it is very hard to decide what is really important, so it is hard for me to recommend someone who approaches the topic honestly. I try to explore this knowledge directly using the technology providers: I read Google's documentation and try to understand how I can implement it.

**Rafał:** I can recommend you as a source of information, and Robert Niechciał, whom I warmly greet.

**Damian:** If we can recommend each other, we try not to do clickbait, within reason, to approach it quite practically, not to go mad. The world runs on sensation, and AI is a rather large sensation, so that impression is hard to shake, but we try to run this education in a mode of not stirring fear among people.

**Rafał:** I think you do it well, really, without sugarcoating. It is not for everyone, but for interested people 100 percent.

### What are the pitfalls and failed AI implementations in SEO?

**Rafał:** At the end about pitfalls and threats. Let us give an example of failed AI implementations in SEO. I saw incredibly failed translations with AI, saying things after which you would sooner sue the company than buy something from it. You need not name the company, because lawyers' hours cost money.

**Damian:** There were many examples where someone hit on the idea: I will take your site's sitemap and rewrite all your content through AI. Such sites generated several million visits two years ago, but used it unskillfully, creating content through ChatGPT, and fell to zero. Dozens of such examples could be multiplied on Twitter. The main problem was that someone did it at mass scale and very fast: a site arose and in two weeks had 100 thousand articles, and later zero. There is even a Polish case of an agency that operated that way and for a long time had hundreds of thousands of visits, and now it is almost zero. The advice: when we generate AI content, not to do it at such a large scale that we publish more every day and later nothing. You have to do it with reason: if so far we published nothing, we cannot publish a thousand articles a day. The second basic problem is hallucinations. Language models are designed so that they are not quite able to say they cannot do something or do not know, but answer with some nonsense. Whoever treats a language model as a source of knowledge and not a knowledge processor is doomed to those hallucinations. According to various studies Google has about 1.5 percent hallucinations, but it is enough that in every hundredth article you will have an incorrect piece of information. Recently there was an example of a guy flying from Australia to Chile, and ChatGPT told him he did not need a visa, and then he sat a week at the airport. We do not want such things. Most often we solve it by gathering, before we write an article, a set of facts from the internet and generating the content based on those facts, not on the knowledge the model has. Failed AI implementations mainly consist of unskillful scaling of content creation by AI and that is mainly what I observed.

**Rafał:** I would like that to resonate: use AI as a knowledge processor and not a source. That is the main reason for hallucinations and those nonsenses that were in AI Overview. Today it is much better, but you remember the beginnings of AI Overview in the States. For me a harmful implementation was precisely AI Overview.

**Damian:** A famous implementation, maybe not strictly from the SEO area, but one that gives pause, is Klarna. A startup that lets you make installment purchases or with deferred payment, multi-billion. Two years ago they laid off I think 1000 people from customer service and switched fully to AI, and now they are backing out, rehiring people for customer service. I think it did not result from errors, but simply people for some reason still do not like being served by machines, as when you try to get through to a bank hotline and the only thing you say is: connect me to a consultant.

**Rafał:** I remember 15-20 years ago travel agencies or banks in Britain that used such automatons. The association in society is that it is the worst there can be, and everyone says: connect me to a consultant. Klarna is a Swedish startup with funding in many rounds from the States, in the UK regarded as a premium processor, not some PayPal giving installments on tragic terms. With a very premium brand they went for a very non-premium solution that cuts costs and misleads, because there were cases where users were misled by AI agents. They started to lose business strongly and had big financial problems in getting the next round.

**Damian:** The main advice: do not do things your clients are not ready for and interfere as little as possible in the relationship between the client and the service. There are services where clients do not require a living person's service, but there are those where it is necessary. In the case of SEO it is worth paying attention to the scale effect: if we do something at large scale, there is a high probability of making a mistake. Especially in important industries, like finance or health: if we automate content creation for a pharmacy and make one error in 100 thousand articles, it may cost someone health or life and that cannot be priced. The problem of buying a different bolt than I wanted is small, but a problem concerning health is big, so you have to approach it with sensitivity, push aside somewhat the rush linked to the FOMO produced by the glut of information, and think about long-term effects, especially since Google changes its opinion about AI content from time to time and we really do not know what it will be in a year.

### How does Google treat AI-generated content?

**Rafał:** How does Google approach AI-generated content? I have read loads on the subject. I myself have several cases where AI content was linked by Stanford University and ranked high in Google, while articles written by a human on the same site, in my opinion much better, were completely devalued and finally deindexed.

**Damian:** There are studies, I cannot cite the source, which showed that de facto for people there is no difference whether a text was written by a human or AI, in most cases. What matters is that human copywriting was also somewhat destroyed by SEO. I will cite the Sunday trading ban: how sites wrote about trading Sundays and how hard it was to find out when the trading Sunday is. In such cases AI is much better, because you get the information faster. In newsy, opinion-forming things people still prefer to read content created by a human and still you can build large reach with content. There is a service in Poland that bet on a strong community, a strong relationship with recipients and wins on that, and on the uniqueness of the content it delivers. What we do in the context of AI writing is try to understand what has already been written and write it in other words, and that has no especially deep value. Google has patents that provide information about how a given unit of content affected the knowledge Google has on a given topic. If we have unique knowledge from some area, for example we write an article about bicycles and know something unique that Google does not yet have, that is an advantage. Google even has patents where it calculates from some formula how much new information a given article delivers to its knowledge graph on that topic.

**Rafał:** That is a good summary: as long as content brings value, there should be no problem. Users will likewise not have a problem, because that content is indistinguishable today: copywriters who write nonsense, and AI that writes nonsense, hallucinates or repeats the same content in other words, as I am now repeating your content in my words.

### What might SEO look like in 2–3 years?

**Rafał:** At the very end, Damian, a few last minutes: I know it is fortune-telling from tea leaves, but how, in your view, will SEO look in 2-3 years and will there even be SEO, or should one already get a bricklayer's certificate and look for work in Norway?

**Damian:** That is mental gymnastics, because in the future we will not even predict the topic of a presentation at the nearest SEMKARKA in June, I will have to tackle this topic broadly. I see a few directions. The SEO industry will survive in one form or another. I think it will undergo some consolidation, because in some areas return on investment will be so small that it will not be possible to invest in it. There will still be a stage where salespeople manage to convince people to spend that money, but a moment will come when everyone realizes it no longer works as it did, and certain groups of clients will drop off. I think small and medium businesses will move for example to local results, and classic content-based SEO will not make sense in their case. The first stage in the SEO industry is automation, thereby an increase in financial efficiency, and those entities will survive that are able to carry it out, because probably it will not be possible to reach a moment when we can charge more for these services, as has been happening so far, because there was more and more traffic in Google and we could deliver more and more effects. Here that will be limited, so our margin should be built on the efficiency of processes, not on how much value we can deliver from the search engine. The second stage is widening horizons beyond the search engine. A large part of traffic can be taken by vertical services, for example ZnanyLekarz or Booksy, meaning internal search engines, and diverse delivery of content, whether on YouTube or to other places where search engine traffic also occurs, will become very important. YouTube is the second search engine and I think it will be the first place SEO agencies turn to. There you can also do keyword research, record evergreens and get quite decent traffic and even capture more of the recipient's attention. A passing trend will be positioning in chats. Lots of tools are being created, lots of agencies specialize in it. I simply do not believe in the long run that you can build visibility and large traffic there, because already now when you use Deep Research from Google, it can review 1600 pages to answer a question. What CTR could you get there? And in the era of agents we will not use chats for everything, the agent will use the chat and make the decision for us. So maybe in the future optimization for the agents' decision.

**Rafał:** That will be a big topic. It is not worth getting hooked on positioning for ChatGPT, it is worth thinking of it as a side effect of your popularity in Google. If you are popular in Google and have authority as a company, as an e-commerce, as your content and products, you will appear there too.

**Damian:** I read that it is worth being high in Bing, because most chatbots use Bing.

**Rafał:** I heard something like that too, I have not tested it, I do not know if it is true.

**Damian:** We will not do much more to be visible in LLMs than what we have been doing. Maybe it is worth betting on distinguishing features in the offer, in copywriting: what distinguishes our product from others. Rankings may be key, because I see that models often search that way, because what we type into ChatGPT is not what ChatGPT types into Bing. It transforms those queries and often that transformation consists of, when we search for cosmetics, typing into Bing "best cosmetics" or "cosmetics ranking". Maybe it is worth positioning that way too, but short term, because long term I think chatbots will not be a big source of traffic. Certainly they will not replace Google in any way, and that now they deliver 3 percent of Google's traffic is, I think, the peak. It will only get worse.

**Rafał:** I think there is still room for growth, but not huge, and more likely Google will be replaced by TikTok and YouTube, in my opinion in the long run.

### Which three things to implement and which three mistakes to avoid?

**Rafał:** At the very end: three things that in your opinion every company should implement right now, without which no e-commerce should do.

**Damian:** Automation of the text layer, especially product and category descriptions, is a no-brainer, because we do not expand the index, we only complete what we already have. Inside the organization one thing I would implement is meeting transcription and note summaries. A small thing, but it changes a lot in the organization. The third thing in e-commerce is all those operations concerning information architecture, meaning maybe we can change the way of categorizing, improve internal linking, and certainly the internal product search.

**Rafał:** And three mistakes to avoid? To sum up.

**Damian:** Relying on one solution, for example ChatGPT. Not testing solutions at large scale, because small errors can carry a large cost. And the third mistake, which seems important to me: a wrong technology stack, meaning we choose tools badly for our processes. If for example we use Make and have 100 thousand products to write daily, it may not be a good solution in the long run.

**Rafał:** I would add that you have to understand your recipient, whom I already mentioned today, and do that first, and then build for the recipient, and not build on the principle that whatever we do may find a recipient. AI will help us understand who buys from us, and that will also help us understand how to reach such a person once we identify them.

**Damian:** You know, Rafał, as they say: let us make the solution, the problem will be found.

**Rafał:** Yes, that will be the title of this episode. Damian, thank you immensely. You put a ton of value into this conversation. If someone wanted to find you and listen to you further, where is best?

**Damian:** Best either on my LinkedIn or at sensai.academy. We have a free newsletter there and every week on Friday we send a fresh dose of knowledge, and on YouTube we also run a program with Robert, on Thursdays at 1 pm, meaning at the time we are recording this podcast now, so you can see it there every week too. We present news from the world of artificial intelligence, if that interests someone.

**Rafał:** I will definitely be at your webinar. Once again thank you very much, all the best and see you.

**Damian:** Thank you very much, see you.

## Guest's recommendation

**Dejan Petrovic (DEJAN), an SEO and AI specialist from Australia** (source of knowledge). Damian recommends following him as the author of software for automating SEO and AI processes, though he uses very technical language.

Źródło: https://moszkowcow.com/podcast/damian-salkowski-ai-in-ecommerce-seo/
