AI product descriptions: problems, possible solutions, and tools
Tabla de Contenidos
Large ecommerce cannot generate quality product descriptions on a massive scale. At keytrends, after 1 year working on it, we have several experiences to share.
The problems encountered by ecommerce when generating their product descriptions are:
❌ It takes too many hands, time, and money
❌ It becomes impossible to maintain consistency in brand voice and tone
❌ SEO optimization requirements don’t apply equally
❌ The arrival of new products exceeds the writing capacity
❌ It is not an efficient process, nor scalable
The advent of AI opens the door to being able to fully automate the writing of product descriptions and categories through bulk processes, but there are still many doubts about this:
→ Will the content be original and position in search engines?
→ Will the contents have errors and hallucinations if I use AI?
→ Will it bring value to more potential buyers and convert them into customers?
→ Is the price commensurate with what I get?
→ Will I be able to scale faster?
The answer depends on the option you choose to create bulk content. To do this, you have to know what AI solutions exist today and what they offer you, but also what are the most common problems when generating this type of content for e-commerce: SEO optimization, strategic approach, the quality of artificial intelligence texts…
The AI solution you choose must be able to solve them with technology that avoids the most common pitfalls of generative AI. Not only is the quality of your content at stake, but it directly affects your business goals.
All this is what you will be able to read in this guide. We can tell you first-hand because we have our hands in the dough all the time, and only through trial and error have we managed to generate original, optimized, and quality content for online stores like yours.
Of course, we’ll tell you what we’ve tried, what other options we’re considering, and how we arrived at a solution that works and why 😉
Common e-commerce problems with product and category descriptions
Two fundamental issues have been giving us some headaches since long before the emergence of AI, and that should be taken into account before generating any bulk content: the SEO positioning processes of the content and the strategy to follow to prioritize the content and achieve better results.
SEO issues that affect your visibility
1. Differentiating your content from your competition
One of the big problems with content in e-commerce is duplicate content derived from a bad practice: copy-pasting the manufacturer’s or distributor’s specifications. This creates 2 problems:
- Lower web positioning due to the absence of original content and useful information for the user
- Lack of credibility and authority in the eyes of Google and in the face of potential customers’ purchase decision
Two aspects that have come to the fore with the emergence of AI and the emphasis that Google is placing on E-E-A-T (Experience, Expertise, Authoritativeness, and Trustworthiness) and the Helpful content.
What is the problem? That AI-powered chats and text generators are prone to producing duplicate texts if they don’t have enough technological resources and a focus Data-driven AI to generate SEO content for people
2. Quality in product and category texts
Creating original content, and not similar or the same as other pages is one of the aspects that Google rewards, including online stores. To find out if you create quality content, just go to the search engine’s documentation and ask yourself these few questions:
- Does the content provide a substantial, complete, or comprehensive description of the topic it covers?
- Does the content provide valuable information when compared to other pages that appear in search results?
- Does the content have spelling or stylistic errors?
- Do you notice that time and effort has gone into creating the content or, on the contrary, does it seem sloppy or hastily created?
- Has the content been checked for errors that are easy to check?
Using the manufacturer’s specifications is not enough. Many times, the information is very brief, does not add value, and does not answer the user’s questions about the uses, characteristics, benefits, and other properties of a product.
Can an AI do this for you? With technology, access to user search data, and the right training for your brand and product type, yes. We’ll see.
3. Detection of possible SEO cannibalizations
SEO cannibalization is very common in e-commerce, so there is no need to be alarmed except in cases in which we identify that:
- One or more relevant product or category pages are losing authority and dropping rankings.
- One or more relevant product or category pages are not being indexed by the search engine for possible duplicate content and consequent crawl budget spending.
Here it is important to re-adjust the optimization of the page or pages that are stealing the spotlight from others that are key, and incorporate detection processes into your workflow (this is why we decided to incorporate a function to check for cannibalizations in our platform 😉).
❕Before you start creating the texts of your product and category pages with an AI for e-commerce, it is important that you audit your online store for serious cases of cannibalization so as not to damage the visibility of your online store.
4. Control of the entire indexing process
While this is something you can do with Google Search Console, Bing’s Webmaster tools, or a indexing monitoring tool like ours, you need to ask yourself this question to make sure your e-commerce content is indexable:
→ Are you creating the differentiating, useful, and quality content for your potential customers that Google expects?
Now more than ever, Google is pushing content created for people that doesn’t just summarize what already exists in other sources. This will not be considered quality and may or may not be indexed, or be de-indexed at some point when detected by algorithm updates.
This was made clear by John Mueller in a Google SEO office-hours 3 years ago:
Why do we say this is so hot today? Easy: the emergence of AI has caused the internet to be filled with an “unprecedented avalanche of content”, as USEO points out. The result: “Google can’t and won’t index the entire internet.”
More than text: strategy issues that affect your business goals
Entrusting the creation of the most commercial content of your online store to a bulk AI tool or service is not the only way you have to save resources and time. It’s the second step you have to take.
First, ask yourself if you have a strategy for prioritizing and directing content creation. Is it is essential to prioritize the products and categories that you are interested in launching and positioning earlier because they will contribute to the numbers of your business.
What that strategy dictates are:
- Product trends and other news in your industry
- User demands for product, resources, and information
- The profit margin your products give you according to the growth and market penetration they have
Again, you have two ways to do this:
🅰️ With a tool that detects those trends, growing words, and related user queries (you can do this, for example, with Google Trends).
🅱️ With the help of the tool or service you hire to create your product sheets and categories in bulk with AI. We do this in the accompaniment of all our clients who start using the platform or hire the service of content generation for e-commerce.
Potential drawbacks of generating product and category texts with AI
Chatbots, text generators, AI content marketing platforms… All of them allow any e-commerce business to write their content automatically. However, not all of the one that is generated is valid:
➡️ Does it meet the quality demands of search engines like Google?
➡️ Is it optimized for the user’s search intent and does it offer real help and value?
➡️ Is it imbued with the style, tone and all the verbal identity of e-commerce?
The answer is no, because the language models we have today (including GPT) have not been trained for it and are prompt-dependent.
This means that the quality of their answers depends on the information and context they receive through the questions, commands, or prompts used to obtain them, but also on the corpus of data from which they have learned or have at their disposal (for example by connecting to different databases, documents or APIs)
👉 That’s right: the more information the AI has and the more relevant it is to the industry and the purpose for which it creates content, the better the texts will be.
Achieving this is in the hands of each tool or platform, and this is where a large part of the quality problems of AI-generated e-commerce content arise. Among the reasons are:
1. Inadequate prompts and poor or no refinement with prompt engineering
The problem with most AI tools for content writing (Ryrt, Qopywriter, Jasper.ai) is that they are based on ChatGPT and its base language model, trained with a closed corpus of documents and information, not adapted to the sector of each business or online store.
Therefore, unless there is a subsequent refinement of the prompts to give the AI more contextual information, the content and information provided by these tools will be generic and will not be optimized for the user’s search intent or personalized to the characteristics of each e-commerce.
Choosing the right prompt-engineering technique(s) can greatly improve AI responses:
- Zero-shot prompt: the LLM will be given a task or question without having been specifically trained on that topic, so their answer will be based on their reasoning ability from very general knowledge. This is the basic technique from which all AI-based content tools based on ChatGPT
- Few-shot prompt: a task or question is provided to an LLM along with some examples of expected responses (various product descriptions). Following that pattern and adapting to the context, the LLM tries to generate a response.
- Chain of thought: a technique that provides a series of step-by-step instructions that help LLMs create content that is much more coherent and tailored to the context provided. For example: detail each of the SEO optimization tasks that you have to carry out with examples of competitors in SERPs.
- RAG (retrieval augmented generation): a method in which, in addition to a prompt, the LLM receives additional information from a knowledge base with examples of related texts. In this way, more personalized content is achieved adapted to the type of business (e-commerce), the project (a specific online store) and its sector (type of product).
Others, such as Backpropagation or decision trees, themselves discriminate between prompts that give a good or bad result for the given command.
❗New OpenAI models and latest news for developers
At the past OpenDev Day 2023 OpenAI announced new features that can mean a leap in content generation: they facilitate the training of models and open the door to greater customization. The most important are:
- GPT-4 Turbo: a new version of the generative language model with up-to-date knowledge as of April 2023 and a 128k context window, allowing it to process more than 300 pages of text in a single entry. With more information available, the quality and consistency of the texts increases significantly.
- Agentes GPTs: the great novelty and revolution. The GPTs agents are customized versions of ChatGPT that each developer or user can create for a specific purpose (for example, analyzing a website’s content and suggesting improvements to optimize SEO). They are easily created by initiating a conversation and contributing web pages, documents, databases, or any additional instruction and knowledge to the pre-trained model. A GPT agent could improve the quality of product descriptions by adapting to the brand’s tone, style, and vocabulary, avoiding repetition and plagiarism, and incorporating feedback from customers and market trends.
- Assistants API: a new API that allows you to call models and tools and makes it easier for developers to build their own applications. It also offers new capabilities: Code interpreter, Retrieval y Function calling.
💻 Code interpreter allows you to write and execute Python code, generate graphs and tables, and process files with various data and formats. This could be useful for automating tasks such as extracting product data from databases or generating descriptions based on code templates.
🔃 Retrieval expands ChatGPT’s knowledge with external information, such as domain-specific data, product information, or documents provided by developers. This means that there is no longer a need to compute and store embeddings, nor do you have to implement sharding and search algorithms. The API itself will decide which recovery technique is most optimal to use. In the context of generating product descriptions, it could be used to search for information about similar products or market trends to inform the generation of descriptions.
📣 Function calling it is used to call predefined functions to perform specific tasks. For example, sentiment analysis of product reviews or generating SEO-optimized descriptions.
But you have to think that a new problem arises. Many ecommerce companies block access to ChatGPT from their robots.txt and it will not be possible to read, train or refine any model because we will not be able to scrape that content. Amazon already does this and many other marketplaces and online shops of all sizes will follow.
Google has made a masterstroke in this regard as its AI scraping bot is Googlebot. Therefore, if an ecommerce is capping Googlebot, it will be preventing Google from crawling and indexing the site. With this, it is more than likely that the majority of ecommerce will choose not to cap it (and that it can be scraped freely).
Anyway, AI product description generation can change completely if we add all these new features to the equation. We can achieve high quality, value, and customization for each project or online store with less complex development processes.
In addition, this increasingly sophisticated refinement or extra training has an added advantage: mitigating the hallucinations inherent in generative AI.
2. Hallucinations in product and category texts: techniques to mitigate them
One of the great handicaps of AI and the greatest danger of generating content for any business are hallucinations, responses that contain information that seems true but is not. Other times, that content doesn’t make any sense or is unfaithful to the source provided.
The good news is that hallucinations can be mitigated by applying a number of enhancements and techniques. Xavier Amatriain brings together in his Guide to Hallucinations more recent:
- Reinforcement Learning from Human Feedback (RLHF): This technique involves training the AI model using a reward signal derived from human evaluations of the generated responses, as he points out Xabier Amatrian. This other University of Cambridge paper, studies and analyzes the effectiveness of RLHF in improving AI responses
- REACT prompting: model to generate prior reasoning in the LLM. This means that the model generates a verbal explanation of its reasoning and makes decisions or actions based on that reasoning, alternating between these two processes. This alternation allows the model to perform dynamic reasoning, that is, to adapt its reasoning and actions as it receives new information or feedback. In other words, it allows the AI model to “think” and act more similarly to how a human would.
- COVE o Chain of Verification: a method developed to reduce hallucinations that allows language models to deliberate on the answers they give to correct their errors. They do this through self-formulated verification questions to check that the answers contain correct information and are free of hallucinations. Meta AI researchers collect their experiments and verifications with this technique in this study.
- DERA (Dialog-Enabled Resolving Agents): technique that improves the accuracy of LLms by using agents with different roles tasked with resolving (or improving) the output through dialogue. One agent role, the “Investigator,” works to identify information relevant to the problem and suggest areas of focus to the other agent. Another agent role, the “Decider,” has the autonomy to react to that information and make final decisions about the output. This interaction allows language models to improve their responses in an iterative way.
- Ajuste de la temperatura del modelo: to balance the level of randomness when predicting subsequent tokens. Higher temperatures make the model’s output more diverse (low-probability tokens), but it can compromise consistency and accuracy. Conversely, a lower temperature promotes greater adherence to high-probability tokens, causing the model to lean towards the most secure and predictable options.
- The rest of the prompt refinement techniques to improve AI outputs that we’ve seen before, especially RAG. According to this study, the RAG method that works best to mitigate hallucinations is the “Post-hoc retrieval,” which involves retrieving relevant documents after generating the text and then editing the text to correct factual errors. It is more accurate than “On-time retrieval“, which retrieves documents before generating the text, and “Iterative retrieval“, which retrieves documents during text generation.
The ideal scenario is to combine various prompt-engineering or hallucination mitigation techniques to tailor responses as much as possible to the specific type of content, project, and specific product type:
💡 You can find out more about why they occur in our guide on content generation with data-driven AI
Best practices for generating ecommerce descriptions with quality AI
One thing is clear: ChatGPT’s pre-trained model is not enough to generate product and category descriptions with AI. The application of prompt-engineering and other techniques to mitigate hallucinations improves the results, but does not personalize them to the specific case (type of content, product, and specific project).
For this, there are advanced solutions that enrich the startup model with new data:
1️⃣ Fine-tuning or re-training with additional database
This technique allows you to customize the pre-trained ChatGPT model by adding an additional layer of training with project data. In the case of online stores, the product database. These include all the attributes needed to generate the right content and vocabulary and technical jargon.
A fact: according to OpenAI, fine-tuning reduces errors in AI responses by 50%. This means that up to 95% of your answers are correct.
In addition, the great advantage of fine-tuning is that it avoids training an LLM of 0 (where we would dispense with all the information that the GPT model already knows as a baseline), and allows it to offer the model relevant examples of the type of content it has to create. This not only ensures quality product sheets or category pages but is also adapted to the brand and its personality.
2️⃣ Advanced pipelines with data scraping and response cleansing
An automated process collects and processes product information from a variety of sources, such as the web (through scraping) or directly from the manufacturer’s databases. This can include extracting product features, specifications, images, pricing, and others.
Scraping not only allows us to obtain data from any website, but we can also analyze the tone and voice used in social media accounts or other documents to replicate it in the content to be generated. In this sense, a huge door is opened for us to personalize content!
And what about text generation? For this, we don’t give total freedom to the AI, but create a sequence of prompts that guides the model step by step. For example, first describe the general characteristics of the product, then its technical specifications, and finally its benefits.
In addition, we would not accept the answer as valid the first time: we would clean and correct incorrect, incomplete, irrelevant or poorly formatted data. All in all, it is a huge leap in quality with response to the responses of the pre-trained model (ChatGPT).
3️⃣ Sectoral embeddings
With embeddings we could go further and not only provide the model with specific information about your products through a database (using, for example, fine-tuning), but also the entire sector of your project.
This requires training the model with a very large corpus of text, but they have a great advantage: they (almost) completely eliminate probability in the texts. The reason? That AI more accurately identifies the context and semantic meaning of words, and that thanks to this it better recognizes industry-specific language.
Currently, all these techniques and strategies for improving AI-generated content are being tested by many tools and platforms, including ours (we already told you about the measures we take against hallucinations). However, to mitigate all these problems, there is already a solution available: data-driven AI.
Data-driven AI as a solution to generate quality and useful AI content for people
When we rode the wave of artificial intelligence, there were hardly any alternatives to generate texts with AI beyond generic writers based on ChatGPT’s pre-trained model (Simplified, Jounce or Copymatic).
Without additional training and without a good context, the ecommerce content generated by these tools did not have the quality or sufficient optimization to position in search engines; Nor was it imbued with the verbal identity of the project.
To go beyond all that, data was needed, and it needed to be present in each of the processes that are carried out to create content. This is how Keytrends approaches a data-driven AI approach, based on:
- User search data: to identify potential topics and information needs of users.
- Google Trends data: To detect in-demand products and new, uncompeting, and growing content opportunities.
- Data from the SERPs: to analyze and define the right search intent through the content you already rank on Google. They are also essential to optimize content for SEO.
- Internal project data: to know which keywords are attacked, what trends are covered, what content has already been created and how it is working. The goal: not to create for the sake of creating, but to build a strategy (and avoid, incidentally, duplicating or cannibalizing content).
In short, everything you need to create user-centered content, and with strategy. To this we have added, little by little, improvements in the response of the AI through the refinement of prompts and fine-tuning. But we’ll get into that later!
How we create product and category pages with data-driven AI in Keytrends
Let’s get to what you’re interested in: How can you generate your product and category pages with Keytrends? Today we support the generation of content for ecommerce in 2 pillars:
» Our AI Content Platform with Trends data
We developed Keytrends so that any project or business could stand out through optimized and strategically created content based on content opportunities, keywords and (above all) new trends in their sector.
In addition, this strategy can be worked on from start to finish from the platform itself and without the need for extra tools, extensions or scripts.
- Spotting trends and growing topics ✔️
In e-commerce, trends are an invaluable source of data to detect products in demand, but also to define their seasonality and prioritize the generation of the most profitable content for each time of the year.
- Optimizing the current strategy and planning new content ✔️
Not infrequently a topic or a trend is discovered and it turns out that there is already a content created. Detecting this is necessary to avoid duplicate content and, in many cases, improving a piece of content will speed up the positioning for a keyword more than generating a new one.
That’s why we’ve given so much importance to including features on the platform that help both detect cannibalization, and be able to identify opportunities in your own content.
- Writing with AI and doing automatic SEO research ✔️
We’ve integrated a large portion of our data into our AI Content Assistant: user searches, SERP data, and competitor content.
Including them within the product page or category page is the only way to cover the search intent, and you have them at your fingertips within the Copilot to generate and optimize any briefing or content.
» Our tech stack for creating content (also in bulk)
As we told you above, we are also testing and implementing new techniques to improve AI responses when creating content. Not only for our platform, from where you can now create product descriptions with the help of Copilot, but also for our product and category page service in bulk for e-commerce.
What we’re testing:
- Prompt engineering: few shot learning and chain of though
- Fine-tuning: with OpenAI (GPT-4) models.
- In context learning: another training method that overcomes the shortcomings that other techniques may have. We find out from this study by several Google Research researchers.
- New alternative models: LLaMA, a large language model launched by Meta AI in February 2023.
- Scraping: for all kinds of data extraction.
» Upcoming features
Our roadmap is growing according to the demand in ecommerce and the advances of AI. Today we are contemplating expanding the generation of automatic content from the platform to the categories (for now we have a prompt for product descriptions) and incorporating the generation of images to accompany SEO articles, product sheets and other types of content.
But above all, we are obsessed with being able to automate the prioritization of products based on trends (something in which today we accompany projects in a personalized way) and internal linking, and to be able to connect to the product feed to have all the data to generate the content and that any change is visible in one click.
As always, we’ll keep you up to date on our releases notes.
5 other options on the market to create commercial content for your e-commerce
Like us, the rest of the AI content tools and platforms have also been evolving. Some of them have the option to create product cards or category texts one by one based on your instructions.
But others are already able to create them in bulk from a specification file for each product, even connecting to your CMS to upload them automatically. Of course, you will have to work and be very clear about the strategic part on your own. Let’s see what options you have:
In this AI writer you can choose what suits you best: either use the specific template for e-commerce for each file you want to generate and insert the instructions for the AI (including the characteristics of the product); or connect to the tool’s API to generate multiple product cards at once.
You’ll need to create a document with all the products and their data, but thanks to integrations with Amazon, Shopify, and others, you’ll have them uploaded in seconds. One of its drawbacks is that it doesn’t detect keywords and doesn’t have SEO optimization options integrated into the generation of bulk product sheets.
❕ Keep in mind that Copy.ai always recommends that review, edit, and supplement AI-generated content, So you must have an editing process and a person from your team who can do it: “We always recommend using AI writing tools as supplements to your writing process (rather than a complete replacement for it)”.
(In fact, no tool of this type, including us, can guarantee flawless AI content despite all the tweaks and improvements we make)
The alternative to Copy.ai for the bulk generation of product descriptions, with an identical system, integrations to e-commerce platforms and marketplaces (including eBay and Walmart), and CMS such as WordPress, and the possibility of custom development by the tool’s team.
The functionality is intuitive, easy to use and allows you to upload product images (you can take a look at the tutorial that they have published):
As Copy.ai, it does not include keyword research nor does it allow SEO optimization when creating bulk content, but it is the user himself who must enter the keywords conscientiously.
Two advantages: in Hypotenuse you can manage more than one catalog and the tool ensures that its generator has been trained with specific ecommerce datasets, so it would give better results and -in theory- never duplicates.
Junia is a template-based content generator that is supported by a very powerful AI text editor and created with the SEO research and optimization process in mind.
To do this, it combines keyword research functions, generation of briefings based on competitors’ articles positioned in SERPs and advanced options such as structured data markup, E-E-A-T references or the definition of the brand voice.
The product description template, is used first, then it is edited and enhanced. Although it can integrate with Shopify, WooCommerce, or Magento, there is no option to generate multiple content in bulk.
Along with Junia in research and SEO optimization of content, Narrato does give you the possibility to create several product sheets in bulk, although with a much-reduced functionality than that of Hypotenuse.
What you can do is upload a CSV file with your products and their features so that the platform’s AI generates a basic description. Then, you’ll need to edit it with the built-in content wizard to add headings (if you need them), bullet points, keywords, and images.
To transfer them to your online store, you can export the generated content or connect via API or Zapier with almost any platform, marketplace, or CMS.
➕ The added bonus? Narrato is a little closer to the strategic approach to content and allows you to manage, plan and work with your team within the tool itself. But, as you can see, they all require the user’s involvement to create product texts.
The latter does not change with the latter tool either, but the process of generating product pages (whether single or bulk) and subsequent editing is done from the same platform and the additional functionalities to be used are integrated.
You don’t have to jump from the template to the editor, from the editor to the keyword research function to optimize, and from here to bulk creation. Describely has everything at hand and both the process of importing and exporting products is done in one click thanks to integrations. You can even publish the changes directly to your store:
❓Does its AI create a definitive text that you can post with your eyes closed? No: you will have to adapt it to the platform or marketplace (for example, by adding titles and bullet points) and you will have to review and add keywords. As we said, the advantage is that you will have the editor and the chatbot of the tool right there to expand the information or paraphrase.
Which solution to choose: comparison table between the top 6 ecommerce AI content tools
We have already seen the tools that today allow you to generate your product descriptions with AI and the most interesting functionalities they have to optimize content, mitigate hallucinations, upload it to your CMS, and more. To make it easier for you to choose, we’ve compiled them all in one table (including our platform):
|Importing and exporting files
What the future holds: advancements in AI-powered product cards and category pages
What’s to come promises to further revolutionize content generation for e-commerce. Hint: everything is going to be even more automatic and instantaneous. Here are a few brushstrokes:
🖌️ All-in-one tools to create complete automatic tabs on marketplaces
The main marketplaces are being the first to work and implement this possibility, starting with Amazon. From the Seller Center, sellers will be able to generate the texts of their products from a simple description. The AI would be responsible for filling in the title, main text, and bullet points, which could be edited later.
In addition, Amazon would be working together with a 3D + AI visualization platform so that sellers can create 3D images of their products and that users can try them on virtually. In addition, there could soon be the ability to create product images in context:
🖌️ Creating instant product descriptions from an image: GPT4 Vision
LLM models with text generation from images are already a reality, and one of the use cases is the creation of commercial texts for online stores. In the SEJ example screenshot, ChatGPT automatically throws up product specifications:
Another option is to complement the image with a prompt, as in the test done by Álvaro Peña of iSocialWeb:
As of today, this is only available to paid users of the tool, although platforms such as eBay yare already testing it to offer it to their sellers.
🖌️ Adhoc pipeline or workflow services for online stores
Our experience and feedback from large e-commerce sites tell us that the way forward will be personalized pipeline or workflow services with automated processes to create and manage all e-commerce content.
A large amount of data would be handled in those processes, such as product information, images, descriptions, prices, inventory, customer reviews, sales, and profit margins…
We, for example, aggregate trend data to set week-by-week content goals. In this way, the most in-demand and, therefore, most profitable product categories are prioritized.
But automation could go beyond content generation, and optimize the way it is uploaded and updated in the online store. This has only just begun! 🚀
As always, to end this guide, we leave you with a mind map that summarizes everything we have seen 👇
❗First, we’ll leave you direct access to our service to generate your e-commerce content with AI. There, we’ll tell you how we do it, what you can expect, and the steps to take.
📃 That’s all we’ve seen in this guide…
- Search Engine Journal. (n.d.). What is Keyword Cannibalization? Retrieved November 14, 2023, from https://www.searchenginejournal.com/on-page-seo/keyword-cannibalization
- Google Developers. (n.d.). Creating helpful content. Retrieved November 14, 2023, from https://developers.google.com/search/docs/fundamentals/creating-helpful-content?hl=es
- uSeo. (n.d.). Problemas de indexación en Google: causas y soluciones. Retrieved November 14, 2023, from https://useo.es/problemas-indexacion-google/
- TELUS International. (2023, July 6). Generative AI Hallucinations: Explanation and Prevention. Retrieved November 14, 2023, from [TELUS International]
- Amatriain, X. (n.d.). Mitigating LLM Hallucinations: a multifaceted approach. Retrieved November 14, 2023, from [Amatriain.net].
- Dhuliawala, S., Komeili, M., Xu, J., Raileanu, R., Li, X., Celikyilmaz, A., & Weston, J. (2023, September 25). Chain-of-Verification Reduces Hallucination in Large Language Models. arXiv preprint arXiv:2309.11495. Retrieved November 14, 2023, from [arXiv.org]
- Nair, V., Schumacher, E., Tso, G., & Kannan, A. (2023, March 30). DERA: Enhancing Large Language Model Completions with Dialog-Enabled Resolving Agents. arXiv preprint arXiv:2303.17071. Retrieved November 14, 2023, from [arXiv.org].
- Prompting Guide. (n.d.). Introduction: LLM Settings. Retrieved November 14, 2023, from [Prompting Guide].
- Prompting Guide. (n.d.). Techniques. Retrieved November 14, 2023, from [Prompting Guide]
- Prompt Engineering. (2023, May 20). An Insightful Journey into LLM Embedding and Fine-Tuning Techniques. Retrieved November 14, 2023, from [Prompt Engineering]
- Carrasco, D. (2023, September 14). Amazon implementa una IA generativa para crear fichas de producto a partir de una descripción. Retrieved November 14, 2023, from [Marketing 4 Ecommerce]
- Amazon Staff. (2023, October 25). Amazon rolls out AI-powered image generation to help advertisers deliver a better ad experience for customers. Retrieved November 14, 2023, from [About Amazon]
- Hines, K. (2023, September 27). GPT-4 With Vision: Examples, Limitations, And Potential Risks. Retrieved November 14, 2023, from [Search Engine Journal]
- Raad, E. (n.d.). E-Commerce Product Descriptions Using GPT-3: From Generation To Fine-Tuning. Retrieved November 14, 2023, from [WordLift Blog]
- React-lm. (n.d.). React-lm. Retrieved November 14, 2023, from https://react-lm.github.io/
- OpenAI. (2021, julio 15). Customizing GPT-3 with your own neural modules. Recuperado el 14 de noviembre de 2023, de https://openai.com/blog/customizing-gpt-3
- OpenAI. (2023, noviembre 6). New models and developer products announced at DevDay. Recuperado el 19 de noviembre de 2023, de New models and developer products announced at DevDay (openai.com)
- OpenAI. (2023, noviembre 6). Introducing GPTs. Recuperado el 19 de noviembre de 2023, de Introducing GPTs (openai.com)