Practical Applications of Neural Networks in SEO: Experience of a 16-Year-Old Advertising Agency
Meta description: How an advertising agency with 16 years of experience integrated neural networks into SEO work: tools, traffic growth case studies, article pricing from 300 RUB and product descriptions from 1 RUB.
When interface-based neural networks like ChatGPT appeared, many SEO specialists began to wonder: how can this be applied in practice? In this article, I will share the real experience of our advertising agency, which is almost 17 years old. We have gone from early experiments to creating an entire neurocopywriting department that generates 500-600 texts per month for various clients.
You will learn which tools we use, how much it costs to create content with neural networks, which traffic growth case studies we achieved, and what awaits the SEO industry in the future. Spoiler: neural networks do not replace humans, but they speed up work many times over.
In this article, you will learn:
- Which neural networks the agency uses for SEO tasks
- How to collect semantics with the help of neural networks
- How to generate product descriptions for 1-6 rubles per item
- Real traffic growth case studies
- What will happen to search and SEO in the next 5-10 years
Where the experience comes from: 16 years in SEO and the emergence of neural networks
Our advertising agency is almost 17 years old. I started as an SEO specialist, built SEO departments in various marketing companies, and have been deeply involved in search engine promotion throughout this time. When interface-based neural networks appeared — ChatGPT and others — I started thinking about how to apply them in practice.
The first goal was obvious: to speed up routine tasks for agency employees, save time and resources, and offer a wider range of services. We started with texts — last year at a conference I gave a talk about how we wrote a huge number of SEO texts for media outlets and increased traffic. Now that story has grown into an entire neurocopywriting department.
It is important to state right away: neural networks do not replace humans. In SEO, there are not that many ways to speed up work, because specialized tools for collecting semantics, writing meta tags, and more already exist. Nevertheless, I found a number of useful tricks, implemented them, recorded an internal course for employees, and it all works.
Which neural networks the agency uses for SEO
We tested many tools and selected those that really work. Here is our top 5.
ChatGPT — the main tool
We have two paid accounts at $20 per month, and they pay for themselves immediately. Several employees use them simultaneously. ChatGPT is great because it has everything built in: you can generate images, work with files and spreadsheets, go online, feed it a website link or a chart.
New analytical models help SEO specialists at the mid and senior levels when strategy needs to be developed. We feed the neural network data about the company, the current state, traffic charts, rankings — and ask it to find bottlenecks or growth opportunities.
Gemini from Google — voice control
Gemini has excellent voice recognition. It is very convenient to communicate by voice and ask questions when it is difficult to formulate thoughts — and that is the main problem when working with neural networks. Gemini helps filter out everything unnecessary from the stream of consciousness and turn it into a summary, article, or instruction.
Claude from Anthropic — the best texts
If you need a high-quality text that is pleasant to read and indistinguishable from something written by a human — Claude is the best. It works excellently with styles. Plus, it has convenient workspaces and additional features.
Perplexity — fact-checking and source search
Perplexity is designed for finding facts and source links. It provides more of them than other tools. This is very helpful when preparing articles that require figures, data, and links to reliable information.
One illustrative example: my student was writing a scientific article on microbiology and, with the help of Perplexity, found a connecting article she had not seen before, even though she was an expert on the topic. The neural network dug quite deeply.
GigaChat from Sber — an alternative without VPN
If you have problems with VPN and cannot use foreign tools, GigaChat from Sber is a viable option. It is also worth mentioning Yandex with Alice — they are deeply integrating neural networks into search, although I do not like the texts yet, but the images turn out great.
Specialized SEO tools based on neural networks
In addition to general-purpose neural networks, there are tools created specifically for SEO specialists.
SEO Chat — based on Google Help
The guys uploaded all of Google’s help documentation for webmasters into ChatGPT, including the guide for raters. It contains a lot of useful information, screenshots, and charts explaining the essence of search. By reading this guide, you can understand all the algorithms.
The tool answers questions and provides a link to the document where the information is mentioned. It is great when clients ask: “Why don’t you fill in meta keywords?” — you can ask the neural network and get a reference to the document explaining why this is not necessary.
SEO Mentor — SEO training
A tool capable of determining your level of knowledge. You ask it to test you, it asks questions, points out gaps, and suggests a training plan: which courses to watch, which lessons to study. It is trained on Google and Yandex help documentation and a large number of articles in English and Russian.
Content usefulness evaluator
You upload an article — the neural network evaluates it according to its checklist: what is good, what is bad, which paragraph can be improved. Useful for self-checking and quality control.
Collecting semantics with the help of neural networks
With the help of neural networks, you can collect semantics, upload pages — a screenshot or text — and ask it to select keywords, what to focus on, and what to write about. The neural network analyzes both text and images.
The main usage experience: when a specialist is not very well versed in a specific topic. Say carpenters come in, and the SEO specialist has spent all their time in woodworking and may not know slang terms or technical specifications. A neural network, knowing every word in the world, will point you in the direction where additional semantics should be sought.
It is useful when you have already hit a ceiling: you have brought the site to the top, occupied the entire SERP for all keywords, and it seems there is still room to grow. With the help of a neural network, you can find additional keywords.
Important: ChatGPT is prone to generating zero-volume queries. Therefore, we use it to identify directions, and check search volume through WordStat and other tools.
A practical example
I fed ChatGPT the link to an SEO crawler website and asked it to compile a list of keywords that people might use to search for such a tool. It produced keywords and grouped them. One third of them had search volume — mostly low-frequency and mid-frequency queries of 3-4 words.
This could have been assigned to a junior specialist to create a general grouping, and then search volume could be checked manually in WordStat or through specialized tools.
Analyzing a website with the help of neural networks
I experimented with analyzing websites through ChatGPT. An important difficulty: some SEO specialists prohibit crawling by neural networks in robots.txt. I have seen bans on crawling a site by all neural networks. If such a ban is in place, ChatGPT cannot access and crawl the content.
You have to manually copy the page HTML or take a screenshot. The result is the same — whether it's a link, a screenshot, or copied HTML.
Important: ChatGPT does not crawl the entire site — it analyzes only the page you feed it. You can work around this through the API, but for a 100-page site you need 20-30 seconds per page, which becomes expensive.
From practical use: junior specialists use neural networks to find underdeveloped commercial factors. They upload a page, ask it to check commercial factors, describe what has been accounted for, what has not, and what should be added. Then they check whether they included everything in their report.
Content generation: meta tags and articles
Meta tags through neural networks are just insanely good. Our workflow looks like this:
- SEO specialist uses ChatGPT to select article topics and agrees on them with the client.
- After approval, tools are used to collect LSI keywords and additional keywords.
- Neurocopywriter generates the article according to all SEO rules.
- In the same thread, the neural network is asked to write meta tags for the article and inserts them immediately.
This saves time: the SEO specialist does not need to invent the basics or adjust the length — everything is already there. In addition, using neural networks, SEO specialists take ready-made text, Title and Description, upload them to ChatGPT, and ask it to check whether all keywords from the technical assignment are mentioned. The neural network highlights the selected keywords, including inflected forms or exact matches.
This expands the quality-check funnel for texts beyond neurocopywriters. Previously, the editor could not handle the volume; now there is self-checking through ChatGPT, which highlights the necessary points, notes the number of words mentioned, keyword stuffing, and more.
Cost and volume
We generate 500-600 texts per month for different clients. One article costs us less than 300 rubles in cost — a good 7,000-8,000-character article that loads quickly, includes images from neural networks, and drives traffic.
Microdata via neural networks
Microdata has always been my pain point. The last time I worked with it manually was 7-8 years ago. You tell programmers: “Here is a link to the instructions, implement microdata” — they reply: “We need to figure out what to use.” I was also too lazy to dive into it.
Now, with ChatGPT, this is solved very easily:
- We upload the page HTML code.
- The neural network creates a final technical assignment: what to wrap, which markup types to use.
- For standard pages, it writes that for all pages of this type, wrap the author, do this and that.
You can analyze the current microdata: upload the page and ask whether everything is marked up and whether all markup types are used. You can generate new microdata, optimize, and expand the existing one.
For mass implementation, there are plugins for WordPress and Bitrix that automatically create microdata via API. You only need to approve or edit it.
Example from practice
I have an SEO crawler that goes through pages and checks for microdata. I filter out pages without microdata, export a table, feed it to ChatGPT, and process all site pages in bulk. If there are many, I split them into clusters.
Mass generation of product descriptions
Huge savings in time and resources. A client came to us with a million similar products — microchips, differing by just one technical characteristic. SEO specialists do not understand them, and writing technical assignments for a million products is unrealistic.
Solution: in popular CMS platforms there are plugins that let you generate texts in bulk. You select a category, write a prompt: “This is a product for such and such, these delivery conditions, interaction methods, volumes, prices.” You upload all the store information, brand, tone of voice. You define variables — what changes (size, color, bit depth).
The neural network substitutes the variables into the same prompt and generates descriptions.
Results
We needed to describe 10,000 products. There were duplicates, so we massively downloaded the texts, sent them for uniqueness checking, and the ones that did not pass were sent for regeneration.
The cost of describing one product card: from 1 to 6 rubles.
Important points for mass use:
- Monitor the API balance — so it doesn't stop halfway
- Prompts will have to be rewritten and improved for a long time
- Set limits on the number of requests so it doesn't loop
- Fine-tuning is better done together with a programmer
Traffic growth cases
Several real-world cases from practice.
Case 1: The impact of texts on traffic growth
A media project where initially 10-15 texts per month were written by copywriters — growth was gradual. Then, with the help of neural networks, they started generating topics and articles — search traffic began to rise. Search engines — both Yandex and Google — treat content created with neural networks equally. They have no special preferences.
Case 2: New domain
A newly created site. Not a single text written by a human — everything with the help of neural networks. Everything is growing perfectly. Articles, commercial pages — everything works.
Case 3: An old site with technical issues
There was an old domain with a CPA network that loads every time. It was necessary to create preloaders and solve a bunch of technical issues. Google was fine, but Yandex was not. We wrote to technical support and studied the errors.
With the help of ChatGPT, we found even more problems than we could initially identify ourselves. We fixed everything — and it produced results. Traffic increased.
Working with external platforms: Zen, VC, Habr
When publishing articles on external platforms, our final goal is not rankings, but traffic and the number of clicks to the client's website or store. For this, we use intermediary pages through which we can see the conversion of clicks into the final product.
The same topic can be rewritten in bulk for all types of platforms that can bring traffic. For SEO traffic, we choose Zen, VC.ru, and Tinkoff Journal. We make the texts unique for each platform.
Advice: Lately, I suggest to clients with budgets that instead of a blog they create media within the product — build an audience and work with it properly: pop-ups, engagement, UGC. If the money is there, this works much better than just an article.
What is still done manually
Not everything can be automated. Here is what we still do manually:
- Classic semantics: collection, clustering — everything has been semi-automated for a long time, but is still done the old-fashioned way
- Publishing on the site: content managers handle this
- Audits: the audit creation algorithm for a commercial proposal is laid out step by step. A strategic SEO specialist's opinion on the client's growth point cannot be replaced by a neural network
- Link building: neural networks cannot massively crawl data and find sites for placement. We tried — it produces a lot of made-up results
Google and neural networks: official position
Google's rules clearly state: artificial intelligence provides new opportunities. It doesn't matter to us who made the text — a human or a neural network — it doesn't matter. The main thing is that the text is high-quality, useful, and reliable.
When clients ask, “What happens if a neural network writes the text?” — I send them a link to Google’s rule, and the questions disappear.
AI-content checkers that marketplaces require work fifty-fifty — like encountering a dinosaur. But they are still set as a requirement.
Creating your own GPT assistant
In the paid version of ChatGPT, you can create your own assistant. Go to “Explore GPTs” and click “Create a GPT.” In dialog mode or manual setup, explain how it should communicate, what it should do, and what it should check.
I created a “digital twin” for myself: I uploaded all my posts from my Telegram channel, articles, and video transcripts. I asked it to answer questions in my style and based on my knowledge. I stopped remembering what I had ever written — I ask my assistant.
The only advice: any information you upload should be converted to text. Images and PDFs can be uploaded, but that will quickly eat up memory. Convert everything to text format and upload it as text.
The future of SEO and search engines
A big area for reflection: where search is headed and whether SEO will die.
I think SEO will die someday — but not now. In the medium term, search will be replaced by personal assistants that stop offering options and instead deliver answers tailored specifically to you. You won’t type “buy pizza” — you’ll say, “I want to order pizza” — and the neural network, knowing your preferences, will give the best option.
What Google and Yandex will do with contextual advertising, how they will make money — that’s a different question. But they are already implementing recommendation systems. Yandex already provides recommendations with sources: you ask a question — the system writes a summary and gives source 1, 2, 3 for each point. If you need more depth, you click through.
What to do now: No specific actions are needed yet to get into recommendation systems. They simply analyze content, and if it is high-quality and directly answers the question, they cite it. Build good websites that meet the requirements of search and neural networks.
Practical tips for working with neural networks
Do not use prompt templates
Top-10 prompts for writing technical specifications or top-10 prompts for high-quality text are templates. I do not recommend using them. The most important thing is to explain the end result to the neural network. Not just “I want traffic to grow,” but “I want traffic to grow for this company; they hold these positions and have this traffic.”
Train the neural network on your own texts
Provide examples of high-quality texts. The more complex and elaborate the tone of voice, the more content needs to be uploaded. For authorial texts with humor, 10-15 texts are enough for the neural network to capture the style.
Check uniqueness
We check all texts on Text.ru. My mandatory minimum is 5-6 points on Advego. ChatGPT often writes straight away at 10-12 points, Claude — from 4 to 6 points. The risk is minimal, and the result is fast, high-quality, and requires no rewrites.
Frequently Asked Questions (FAQ)
Can SEO be fully entrusted to neural networks?
No. Neural networks speed up work, but strategic thinking, audits, link building, and competitor analysis are still done manually. Neural networks are an assistant, not a replacement for a specialist.
How do search engines treat content created by neural networks?
The same — both Yandex and Google. They have no preferences. The main thing is the quality, usefulness, and reliability of the content, not who created it.
How much does an article written with neural networks cost?
In our agency, the cost of one article is less than 300 rubles. A good article of 7,000-8,000 characters with images that brings traffic.
How do you bypass the ban on neural-network crawling in robots.txt?
If the ban is in place, ChatGPT won’t be able to access the site. You have to copy the page HTML manually or take screenshots. The best solution is to ask the client to remove the ban.
What will happen to contextual advertising when search is replaced by recommendations?
That is an open question. Google and Yandex are testing formats, but it is still unclear how they will integrate ads into recommendation systems. This is something to watch.
Is special microdata needed to get into recommendation systems?
No. For now, no specific actions are needed. Neural networks analyze content, and if it is high-quality, they cite it. Build good websites with quality content.
Conclusion
Neural networks in SEO are not the future — they are the present. In our agency, we use them to generate 500-600 texts per month, collect semantics, analyze websites, create microdata, and mass-produce product descriptions. The cost of one article is less than 300 rubles; product descriptions cost from 1 to 6 rubles.
At the same time, neural networks do not replace humans. Strategic thinking, audits, link building, and client work are still done manually. Neural networks are a powerful tool for speeding up routine tasks.
Search engines treat content from neural networks the same way. The main thing is quality and usefulness for the reader. And what will happen to search in 5-10 years will tell in time. But it is already clear: whoever does not use neural networks is falling behind.
All links to the mentioned tools, the presentation, and the interview are available in my Telegram channel.
Sources: Google Search Central Documentation, Yandex Help for Webmasters, advertising agency experience (16 years in the industry).