How AI generates descriptions for 1000+ products overnight - without a copywriter

Last week the owner of an online auto parts store wrote to me. He has a little more than four thousand items in his catalog, and for about three thousand of them the description looks like this: name, article number, three technical characteristics. All. Not a single human suggestion, not a single word about why this filter is better than another, how this block differs from an analogue at half the price, why this particular timing belt should be installed on the second generation Kia Ceed. He asked: “How much time and money does it take to properly describe the entire catalog?” I counted. With the average cost of a copywriter for technical topics - a thousand to one and a half rubles per card, if you want not just a rewrite from the supplier’s website, but normal unique text with an understanding of the product. Three thousand cards. It turns out from three to four and a half million rubles. In terms of time, even if you hire a team of five copywriters and they produce ten descriptions a day each, that’s twelve working weeks. Three months. And this is provided that all five understand auto parts, that texts do not need to be checked and that no one goes on sick leave.

He laughed and said that he has been living with these empty cards for three years now, because every time he calculates the budget for content, he understands that for this money it would be better to buy another batch of goods. And you know what - he's not the only one. I talk to dozens of store owners every month, and the story is the same. The catalog is growing, the cards are empty, hands are not getting to the content, and in the meantime search engines are increasingly screening out pages without normal text. Google says it straight: thin content is a problem. Yandex is even stricter - a page without a description for it actually does not exist in the search results. And so the business finds itself in a trap: without content there is no traffic, without traffic there is no budget for content.

Actually, it was from this trap that the idea was born, which we later implemented in our plugin. Not because we wanted to make “another AI text generator” - there are already dozens of them on the market. But because none of the existing ones solved the problem the way a real WooCommerce store owner faces it. Let me tell you how this works in practice, why the result is different from what ChatGPT produces in the browser, and - most importantly - what pitfalls we discovered over two years of working with AI content for e-commerce.

Why "just copy from ChatGPT" doesn't work for the store

I hear this argument all the time: “Why do I need a plugin when I can open ChatGPT, paste in the product name and get a description?” Formally, yes, you can. I myself started this way two years ago, when AI text generation was just gaining momentum. I sat down, opened the interface, copied the name of the product, added “write a description for the online store,” received the text, read it, edited it, copied it into WordPress, and moved on to the next product. On the twentieth item I realized that two hours had passed. On the hundredth - that a week has passed. And this despite the fact that I type quite quickly and formulate prompts well.

The problem is not the quality of the model. Claude and GPT-4 write excellent texts if asked correctly. The problem is scale. When you have three thousand products, manual copy-paste from chat is not a solution, but an imitation of vigorous activity. But even scale is not the main problem. The main problem is in the context.

Look here. When you ask ChatGPT to write a description for “Shell Helix Ultra 5W-30 Motor Oil, 4 L,” it gives you a pretty decent text. But this text doesn't know anything about your store. He doesn't know that you have fifteen other Shell oils in your catalog that would be worth linking to. He doesn’t know that the product has attributes “viscosity”, “OEM tolerances”, “engine type”, which are stored in WooCommerce, and they need to be organically woven into the text. He doesn't know that there are three photos for this product, and each one needs alt text with keywords. He doesn't know that you already generated a description for a similar product an hour ago, and the new text should not be similar to it. He knows nothing about the structure of your site, about your categories, about which pages you need to put internal links on.

In other words, ChatGPT in the browser is a powerful brain without eyes or hands. He knows how to think and write, but he does not see your store and cannot act in it independently. And we needed to make sure that the AI ​​did not just generate text in a vacuum, but worked inside WooCommerce - it saw product data, understood the catalog structure, knew the history of previous generations and could independently sort the result into the required fields. This is the difference between an “AI chatbot” and an “AI e-commerce tool” - it determines everything.

And here I want to talk about the architectural solution, which, in my opinion, turned out to be key. We did not become tied to one provider. The plugin implements the Strategy pattern - a single interface through which you can connect Claude from Anthropic or GPT from OpenAI. For what? For several reasons. Models behave differently depending on the type of content. Claude, in my subjective opinion, copes better with long, detailed descriptions - it litters less with template phrases and is less likely to repeat itself when you generate hundreds of texts in a row. GPT-4 is good for short, marketing-charged descriptions - it is a little more aggressive in sales language. But it's not just about taste preferences. If tomorrow Anthropic doubles prices or OpenAI rolls out a new generation model that turns out to be much better, you just switch the provider in the settings, without migration, without data loss, without rewriting prompts. Prompts are stored in the system, rather than hard-coded into code. A provider is simply a transport that delivers your request to the model and returns a response.

And one more point about cost, which many people miss. I have already said that a copywriter for technical topics costs one thousand and a half per card. How much does AI generation cost? We measured it on real projects. When using Claude - from half a cent to three cents per item. This takes into account entry and exit tokens, taking into account the fact that we are sending not just the name, but the full context: attributes, categories, related products. Let's say we take the upper limit - three cents. For three thousand goods, that's ninety dollars. About eight thousand rubles at the current exchange rate. Eight thousand rubles versus three or four million. The difference is four hundred times. Even if you spend another forty thousand on manual finishing of the most important cards, it will still be two orders of magnitude cheaper.

Of course, someone will now say: “Well, AI text is worse!” Yes and no. If you compare the best AI text with the text from a copywriter who deeply understands the product and spent an hour on one card - yes, the copywriter will write better. But let's be honest: how many of these copywriters are there in nature? And how many of them are ready to write ten descriptions of oils a day for three months? The real alternative to an AI description is not a perfect text from an expert. The real alternative is a blank card. And between the AI ​​description and an empty card, the choice is obvious.

Batch processing: how to set a task for a thousand products and go to sleep

Now we come to what makes our approach fundamentally different compared to manual work in a chat interface. Batch generation. It sounds boring, like a title from a corporate presentation, but in practice this is exactly what turns AI from a toy into a working tool.

What it looks like. You go to the plugin panel, select products - you can by category, you can by description status (empty, short, not updated for a long time), you can simply check the boxes. Choose a prompt - a template on which the text will be built. Click “Launch”. That's it, you can close the tab. You can turn off your computer. You can go to bed.

Behind the scenes there is an engine that is built on Action Scheduler - this is the same queue engine that WooCommerce uses for its internal tasks. Not cron, not an external service, not a separate server. Everything runs inside your WordPress using a proven infrastructure. BatchProcessor takes a task, breaks it into chunks - say, ten products each - and puts each chunk in a queue. One chunk is processed - the next one is launched. If the server is overloaded and the process is interrupted, it’s okay, Action Scheduler will automatically restart the task. If an error occurs when generating one specific product, it will be marked as problematic, and the rest will continue to be processed. In the morning you open the panel and see a report: how many products were processed, how many errors, how much it cost.

I remember the first time we ran batch generation on a real directory. It was a building materials store, with about one thousand two hundred items - from dry mixes to power tools. The owner was skeptical, saying that the AI ​​would probably write nonsense because it doesn’t understand construction. We prepared a prompt that took into account the specifics of the niche - technical characteristics, areas of application, comparison with analogues. Launched on Friday evening. On Saturday morning he opened the store and had one thousand one hundred and eighty new descriptions. Twenty products fell with errors - mostly those that had no data at all in the database, not even the name of a normal one. He leafed through the rest for an hour and a half, found about fifty that he wanted to edit, and said a phrase that I remembered: “They wrote better than my last copywriter, whom I paid eight hundred rubles apiece.”

Honestly, this didn't surprise me. Not because AI is magical, but because a good prompt, supported by real data about the product, gives the model context that a copywriter usually does not receive. The copywriter receives a technical assignment on a piece of paper: “Write a description for Knauf Rotband plaster, 30 kg. Key words: buy plaster, gypsum plaster.” And AI receives the full name, all attributes (composition, packaging, consumption, layer thickness, drying time), category, related products, price and even the text of reviews, if any. And all this is automatic - no need to collect and transmit by hand.

But I would be lying if I said that everything works perfectly out of the box. The key element of the system is prompts. It is the prompt that determines what the text will be: long or short, technical or emotional, with or without SEO keywords. And here we hit a lot of bumps.

The first version of prompts was universal - one prompt for all products. The result was average: the texts were read normally, but they were somehow... nothing. It's as if they were written by an intern who is trying hard but doesn't understand the audience. Then we made a system of customizable prompts with presets for different niches. For auto parts - one tone, with an emphasis on compatibility, OEM numbers, differences from fakes. For cosmetics - another, with an emphasis on ingredients, effect, skin type. For building materials - the third, the most technical, with consumption calculations and recommendations for use. Prompts can be created and edited directly in the admin panel, they are stored in the database, and if you wish, you can make at least ten different prompts for ten product categories.

And this is where the fun begins. When you have a batch generation system and custom prompts, a natural question arises: how can you figure out which prompt works better? Which text gives more conversions - short and aggressive or long and detailed? Technical with numbers or emotional with stories?

A/B tests of descriptions and quality control: when AI checks itself

We built A/B testing of descriptions into the plugin, and I think this is one of the most underrated features. It works simply: you generate two description options for the same product - say, one through Claude with a long detailed prompt, the other through GPT-4 with a short selling prompt. The system shows option A to half of the visitors, option B to the other half, and collects data: views, additions to cart, purchases. After a couple of weeks, you see that option A gives a 3.2% conversion to cart, and option B - 4.1%. Switch all products to the winning option. Testing a new hypothesis again. And so iteratively come to descriptions that actually sell, and do not just fill empty space on the page.

It sounds like something out of a growth hacking textbook, but in practice it changes the entire approach to content. You stop arguing about taste - “I think we need a different tone here” - and start making decisions based on data. Real data, from real customers. I've seen cases where a store owner was absolutely sure that his audience wanted long, expert descriptions with tables of characteristics, and a test showed that short, emotional texts of three paragraphs gave a forty percent higher conversion rate. He didn’t believe it, restarted the test - the result was confirmed. It hurts the ego, but it's good for business.

But testing is already the second stage. Before showing the text to customers, you need to make sure that it is generally adequate. This is where quality control comes into play. We have built into the plugin several mechanisms that catch typical AI generation problems.

Preview before publication. Any generated text first goes into draft status. You can view it, edit it and only then publish it. Moreover, the interface has a diff comparison - the old text is on the left, the new one is on the right, the changes are highlighted. This is especially useful when you are regenerating descriptions for products that already had some text. You can immediately see what the AI ​​has added, what it has removed, and whether any important information is missing.

Next - SEO audit. Each product card is checked against more than fifteen criteria. The length of the meta description is not too short, not too long. The presence of a keyword in the title and in the first paragraph. The uniqueness of the meta title tag - whether there are any duplicates with other products. Alt-text content for images. Availability of structured data. Internal links to related products. The system highlights problems and suggests what to fix. Moreover, this works not only for AI texts - you can run an audit throughout the entire catalog and find problems in the descriptions that people wrote.

I thought for a long time about how much AI can be trusted in terms of content quality. After all, if AI both generates text and checks it, isn’t it a vicious circle? In part, yes, it works. Therefore, we did not create a fully automatic “generate-publish-forget” pipeline. The person remains in the loop. The system generates, checks against formal criteria, highlights potential problems, but the final decision on publication is made by the store owner. In our experience, with a well-tuned prompt, about eighty percent of descriptions can be published as is, another fifteen percent require minimal editing, and only five percent need to be significantly rewritten. For comparison, when we worked with freelance copywriters, the ratio was about fifty to thirty to twenty. That is, AI, even in terms of “how many texts need to be redone,” turns out to be more stable than the average copywriter.

But let's look at it from the other side. There are products for which the AI ​​description will be frankly weak. Highly specialized industrial products, where you need to know not just the characteristics, but the context of application, industry standards, and the history of testing at specific enterprises. Medical products where imprecise wording can be dangerous. Handmade goods, where the value is in the history of the master, and not in a list of parameters. For such cases, AI is a starting point, a rough draft, not a final product. And that's okay. A tool should not replace the brain - it should save time on routine so that a person can focus on where his expertise is really needed.

And while we're on the topic of SEO, let's dig deeper. Because generating descriptions is just the tip of the iceberg. Full product card optimization is a whole complex of tasks, and we tried to automate as many of them as possible.

Content is not just text: FAQ, alt-texts, linking, images

Do you know what irritates me most about most e-commerce AI tools? They solve one problem. They generate a description and that’s it. And then you yourself deal with meta tags, alt texts, FAQ sections, internal links. Each task must be done separately, in a different tool, often manually.

We took a different route. We decided that if the AI ​​model “sees” a product with all its data, it can do much more for this product than just a description. And gradually they built up an entire ecosystem around the main generator.

FAQ - One of my favorite stories. For several years now, Google has been promoting FAQ schema markup - structured question-answer data that can be shown directly in the search results, under your link. This is a rich snippet that visually takes up more space and attracts more clicks. But to add a FAQ, you need to actually write this FAQ. For each product. Three to five questions with answers. For three thousand products, this is nine to fifteen thousand question-answer pairs. Manually? Good luck. AI does this automatically. When generating a description, the plugin can simultaneously generate a FAQ block and immediately place it in the JSON-LD schema format. Questions are generated based on real product attributes - “What is the viscosity of this oil?”, “Is it suitable for diesel engines?”, “What is the consumption per square meter?” Not abstract, but concrete, tied to data.

Alt texts for images is another task that everyone puts off until later and never does. You have twenty thousand photos in your catalog, and nineteen thousand of them have either empty alt text or “image-001.jpg.” Search engines can’t look at pictures the way people do—they need a text alt to understand what’s in a photo. And this is not only SEO - this is accessibility for people with visual impairments, this is a requirement of standards. AI generates alt texts in batches, based on the context of the product and analysis of the image itself. No, it does not “look” at the photo in the sense that it identifies the object in the photo - although this is also possible with multimodal models. But he takes the name of the product, its characteristics, the photo number on the card and formulates a meaningful alt: “Shell Helix Ultra 5W-30 motor oil, 4 liter canister, front view.” This is a hundred times better than an empty alt or filename.

Internal linking is perhaps the most complex and at the same time the most valuable function. I've been doing SEO for many years and know that internal links are one of the most powerful and yet most underutilized tools. Search engines use internal links to understand the structure of a site, determine the importance of pages, and distribute PageRank. The problem is that manually setting internal links is painstaking work that requires knowledge of the entire directory. You need to remember that you have a page with a selection of oils for trucks, that there is a category “Hydraulic oils”, that there is a blog article about the differences between synthetics and semi-synthetics. AI analyzes the catalog and places contextual links directly in the description text. Not stupidly - “see also product X” - but organically, within a sentence: “If you need hydraulic oil, check out our line of hydraulic oils.” This works both for SEO and for the user, who gets additional navigation paths through the catalog.

I have to say that it was the linking feature that caused the most controversy within the team. Some said that AI will put irrelevant links and this will harm the site. Others say that the links will be too formulaic and look unnatural. We tested for several months on real directories, and this is what we found out: with the right prompt, AI places links no worse than an SEO specialist who knows the directory well would do. The key word is “with the right prompt.” If you simply ask “add links to similar products,” the result will be mechanical. But if you give the model context - a complete list of categories, top products, active landing pages - it begins to build a meaningful grid of links.

I’ll tell you separately about image generation, because this is an area that few people take seriously, but it solves a real problem. Not all products in the catalog have photographs. This is especially true for stores that sell goods from suppliers - often you only receive a technical description and, if you're lucky, one photo from the catalog in low resolution. We have integrated two image generation engines: DALL-E 3 from OpenAI and Stability AI. The plugin takes the context of the product - name, category, key characteristics - and generates a prompt to generate an image. This is not a replacement for a photograph of the actual product, of course. But for products that have no visual at all, an AI image is better than a gray square with the words “no photo available.” This works especially well for banners, article covers, and illustrations for collections. It’s too early for actual product photos. The buyer wants to see the real product. But the industry is developing quickly, and I think that in a couple of years AI product photos will become indistinguishable from studio photos.

And look at the picture that emerges. AI generates a product description. At the same time - meta tags for SEO. At the same time - FAQ in JSON-LD format. At the same time - alt-texts for photos. At the same time - internal links to relevant catalog pages. If necessary, images. And all this for thousands of products in one night, while you sleep. One copywriter would spend six months on this amount of work. The whole team is a month. AI - night.

But I'd like to talk about another aspect that is often overlooked - content beyond product cards. Because descriptions are good, but full SEO requires much more.

Blog, content calendar and special content: AI as a marketing department at minimum wages

Any SEO specialist will tell you that an online store website that only has product cards is inferior to a website that also runs a blog, publishes guides, and makes seasonal selections. The reasons are obvious: articles attract information traffic, people who are looking for “what oil to put in a Kia Ceed” or “what is the difference between gypsum plaster and cement plaster” end up on your website, read the article and, with a certain probability, go to the catalogue. The classic content marketing funnel.

The problem, as usual, is resources. Running a blog for an online store is a separate task that requires planning topics, writing articles, layout, and publishing. For small businesses this is often unrealistic - there are simply no people for the blog. This is where AI can help again.

The plugin has a blog post generator. It works a little differently than the description generator. You select one or more products and say: “Write an article.” AI analyzes products, their characteristics, category and creates a full-fledged article of two to four thousand words. Not a set of facts from a card, but an article - with an introduction, development of the topic, practical advice, and a conclusion. You can choose a tone: expert, practical, marketing, technical. The article is created as a draft in WordPress - with all meta tags, SEO optimization, and markup. If you want, you can automatically add AI-generated illustrations. And all this can also be launched in batches: selected one hundred products → received one hundred articles → scrolled through and published the best ones.

Content calendar is next level. Instead of thinking every time “what should I write about,” you tell AI: “Plan a blog for a month for a motor oil store.” It analyzes your catalog, the current season, trends in the niche and produces a publication plan with topics, keywords and recommended dates. January - article about winter oils and engine warming up. March is about changing the oil before long trips. June - about oils for garden equipment. This is not magic, it is data analysis and common sense, but framed in the form of a concrete plan that you can immediately begin to implement.

And finally - special content. Landing pages, seasonal selections, selection guides. This is content that is usually made by a marketer or content manager, and each such page takes from several hours to several days. AI generates a draft in minutes. “Make a landing page with a selection of oils for trucks.” “Make a seasonal guide to preparing your car for winter with our products.” “Make a comparative review of three brands of hydraulic oils from our catalog.” All this is created based on real data from your WooCommerce - with correct prices, articles, links to products.

I'm not saying AI is replacing the marketing department. But for businesses that don’t have a marketing department—and that’s ninety percent of small online stores—this is an opportunity to have content that simply didn’t exist before. The difference between “no blog” and “there is a blog with twenty articles written by AI and modified by the owner” is colossal. For SEO, for customer trust, for positioning the store as an expert in its niche.

What if we look at it from the other side? There is a well-founded criticism of AI content: search engines learn to recognize AI texts and can pessimize them in search results. Google has explicitly stated that AI content itself does not violate the rules, but “content designed to manipulate rankings” does. A fine line. I believe the key factor here is value to the user. If an AI description really helps the buyer make a decision, if it contains relevant information that was not previously on the page, it is useful content, regardless of who wrote it. If this is a meaningless set of keywords generated for the sake of SEO, search engines are right to be pessimistic about it. Our system is based on the first approach: AI works with real product data and creates content that helps the buyer. Not “buy motor oil inexpensively, reviews, price,” but a meaningful text about the product.

I've spent a lot of time thinking about where the line is between useful automation and spam. And I came to this conclusion: the border is drawn by intention. If your intention is to help the buyer understand a product and make a decision, AI is the tool that allows you to do this at scale. If the intention is to fill empty pages with at least some text for the sake of positions in the search results, then no AI will save you, because Yandex and Google have long learned to distinguish one from the other. And AI has nothing to do with it - a copywriter who writes for the sake of keywords, and not for the sake of the reader, creates exactly the same garbage, just manually.

Real numbers and honest talk about limitations

I promised real numbers, and here they are. We have collected statistics on dozens of stores that use the AI ​​generation module.

The average cost of generating one product description through Claude Sonnet is about one or two cents. Through Claude Opus - three to five cents. Through GPT-4 - two to four cents. This takes into account everything: both the prompt context for the input and the generated text for the output. For a large-scale catalog of five thousand products through Claude Sonnet, it turns out to be fifty to one hundred dollars. Through Opus - one hundred fifty to two hundred fifty. For comparison: the average stock copywriter charges five hundred to eight hundred rubles for a description. A good one is a thousand and a half. Specialized technical - two to three thousand. Five thousand products for the average copywriter is two and a half to four million rubles. The difference is two orders of magnitude.

In terms of speed: batch generation of a thousand products takes approximately six to eight hours if the server is not loaded. This takes into account delays between requests to the API, so as not to hit the rate limit. Two thousand - twelve to fifteen hours. There is plenty of night to spare. By comparison, a team of five copywriters writes a thousand descriptions in four to six weeks. AI does it overnight.

In terms of quality - and here I will be as honest as possible - the result strongly depends on three factors. The quality of the initial product data. If the card only contains the name and price, the AI ​​won’t pull any punches, the text will be weak. Prompt quality. The default prompt gives an average result, a custom one for a niche gives a good result. Niche. For standard consumer products (cosmetics, household chemicals, clothing) AI writes excellently. For highly specialized ones (industrial equipment, medical devices, chemical reagents) - serious modification is needed.

And here's another thing I want to talk about frankly: AI generates text based on its training data. He doesn't know that your specific product was certified in 2024. He doesn't know that you have twenty left in stock. He may not know that a particular model has been discontinued. All factual statements related to the specific circumstances of your business should be verified. AI does not intentionally lie—it has no intentions at all—but it can “invent” facts that are not in the source data. We minimize this problem by feeding as much real data from WooCommerce into the prompt, but we cannot completely eliminate it. That is why the system has a review stage before publication.

Separately, I want to dwell on the topic of multilingualism, because this is a pain that is rarely talked about. We have clients who sell in several CIS countries - Russia, Kazakhstan, Belarus. And there are those who work for export and need descriptions in English, German, and sometimes even Arabic. Previously, this meant hiring native copywriters for each language, or worse, running texts through Google Translate and praying that the results weren't too embarrassing. AI models, especially Claude and GPT-4, work with dozens of languages ​​at a high level. You can generate a description in Russian, and then - using the same product context - generate a version in Kazakh or English. This is not machine translation, but rather the generation of new text in another language, taking into account all the product data. The result is an order of magnitude better than that of a translator, because AI does not translate phrase by phrase, but writes it anew, adapting it to language norms. For online stores that want to enter the international market or simply serve Russian-speaking customers in different countries, this is a critical opportunity that used to cost tens of thousands of dollars for localization, but is now solved in one overnight batch generation.

Another aspect that I consider extremely important is the reproducibility of the result. When you work with a freelance copywriter, each new text is a lottery. Today he is in a good mood and wrote well. Tomorrow - tired, wrote crumpled. The day after tomorrow - he left for a competitor, and you need to look for a new one who writes differently, and your catalog begins to look like a patchwork quilt. AI generates texts with the same level of quality, in the same style, with a single structure. The thousandth text will be written as well as the first. Not better - but not worse either. For a catalog of thousands of products, this consistency is a huge advantage, because the buyer, moving from one product to another, sees uniform, professional content, and not a discordant text from ten different authors.

I know people who run completely automatic generation and publication - no review, no verification. This works for some niches: if you sell thousands of standard products (bolts, nuts, wires), where there can be no factual errors, because all data is taken from attributes, auto-publishing is acceptable. But for goods where accuracy is important - medicines, technical safety equipment, chemicals - manual checking is mandatory. We do not impose any of the approaches: the system gives you the tools for both full automation and thorough manual review. The choice is yours because you know your business better than any algorithm.

Now let's talk about what happens when a directory is described. Because generating descriptions is a one-time process (well, or once a year when updating). What's next? Then the maintenance routine begins: new products, updating information, seasonal promotions, updating prices in texts. And this is where batch generation reveals itself in all its glory. Have two hundred new products appeared in the catalog? We launched batch generation, the descriptions are ready in the morning. Have OEM tolerances changed for the line of oils? We regenerated the descriptions for these products with an updated prompt. Has winter tire season begun? We launched the generation of seasonal landing pages and blog articles. All this is hours, not weeks.

There is another scenario that I especially like. Store migration. When a business moves from one platform to another - say, from OpenCart to WooCommerce - it often turns out that half the descriptions on the old site were written ten years ago and are hopelessly outdated. Transferring them one to one is stupid. Rewriting by hand is expensive. But running it through AI generation, using the old text as part of the context, is both fast and high-quality. The model takes what was useful in the old description, updates it, rewrites it in a modern tone and adds new data from updated attributes. I've seen a store with three thousand products undergo a migration and completely update all of its content in a weekend. Without AI, this process would take months and would most likely lead to a temporary drop in search rankings.

You know, I sometimes catch myself thinking that two years ago the idea of ​​automatically generating content for a store seemed like something out of science fiction. Not because the AI ​​couldn’t write text—it could, and quite well. Because between “AI can write text” and “AI can take data from your WooCommerce, generate description, meta tags, FAQ, alt texts, check SEO, compare with the previous version and neatly sort everything into fields” is a gap of thousands of hours of development. We have overcome this gap. Not right away, not without mistakes, not without rework. But we overcame it.

And now, when I see a store with five thousand empty cards, I no longer feel hopeless. I know this is a solvable problem. Not in months and not in millions. Per night and for the cost of dinner in the restaurant. And in the morning, the store owner will open his WooCommerce and see that each product has a description, meta tags are filled, FAQ is in place, alt-texts are set. And a search engine, visiting this site in a couple of days, will see a completely different picture - not empty stub pages, but full-fledged product cards, each of which answers the buyer’s questions.

This is not the future. It's already working. Right now, in real stores, with real products and real customers.

Try COS WP Woo - 14 days free.Install the plugin, connect the Claude or OpenAI API key, select products and run the first generation. In an hour, you'll have descriptions, meta tags, and FAQs for hundreds of products. Every other night - for the entire catalogue. No obligations, no strings attached. Just see how it works on your store, with your products, in your niche. Sometimes the best argument is the result that you can feel with your hands.