How We Automated Article Publishing on the Website
By spring 2026, automation involving artificial intelligence had ceased to be an experiment in the corner and had become a working tool. The open Model Context Protocol, which connects language models with external data and services, was introduced in November 2024, and by March 2026, its packages were downloaded about 97 million times a month — in sixteen months, it went from a niche idea to an actual standard. In December 2025, the protocol was transferred under the management of the Linux Foundation, and according to industry reviews at the beginning of 2026, about 80 percent of the largest companies were already running AI agents in real processes, and about a quarter of them connected such agents to their systems through this protocol. Against this backdrop, we decided to test a simple hypothesis: is it possible to put not a one-time task, but the regular publication of expert articles on a conveyor belt — from choosing a topic to publishing on the website and social networks.
In our practice at West Star Ltd, we have been involved in the automation of business processes, integration with accounting systems, and the implementation of AI in routine tasks for many years. It was logical to apply the same approach to our own content: not to hire a separate person for each article, but to assemble a process that finds the topic itself, checks the facts, writes a draft, checks it for errors, and publishes it on schedule. Below, we honestly analyze how this conveyor is arranged, what it can really do, and where human intervention is still needed. This is not an advertisement for the technology but an analysis of an engineering solution with its strengths and weaknesses.
WHERE THE CONVEYOR BEGINS
The first and most underrated stage is choosing a topic. A poor article on a popular topic is more beneficial than a brilliant article that no one is looking for. Therefore, the conveyor starts not with text but with demand data. We look at the performance report in the search console: which queries already lead to impressions but with a low click-through rate and are in the second or third dozen positions. These are the topics where the site is already visible to the search engine but does not sufficiently answer the question, meaning new material has a chance to rise higher.
If there is no suitable query, a pre-prepared pool of topics related to our expertise is activated: integration with accounting systems, chatbots, tax accounting in Kazakhstan, development and DevOps, applied AI. Before taking a topic from the pool, the conveyor collects the titles of already published articles and checks them to avoid writing a second material on the same thing. This is protection against so-called cannibalization, where two similar pages compete with each other in search results and both fall. The rule is simple: one topic — one article.
HOW FACTS ARE GATHERED
The main danger of automatic text generation is confidently written falsehoods. A language model, by its nature, sounds good, but its knowledge is limited by the training date, and it tends to invent details. Therefore, in our process, fact-checking occurs before writing, not after.
Before each article, the conveyor sends a series of search queries and checks everything that might have changed: numbers, rates, dates, product versions, current conditions. For example, the VAT rate in Kazakhstan cannot be taken from memory — it changes, and only the value confirmed by current sources gets into the text. If a fact is not confirmed, it does not go into the article at all. This order eliminates most errors at the input because the model writes based on fresh data, not its assumptions.
This is a key difference from the naive approach, where a neural network is simply asked to write an article on a given topic. Without external verification, the text looks convincing but easily contains outdated figures and non-existent details. A separate rule is not to attribute to companies and products what they did not do and not to publish statements about partnerships that cannot be confirmed.
GENERATION AND CHECKING OF TEXT
When the facts are collected, the conveyor writes a draft according to a strict set of rules: a specified volume, an analytical tone from the first person, a mandatory section on limitations, practical conclusions for different readers. A single template is needed so that articles maintain one level and recognizable style, not turning into a set of random texts of varying quality.
Then begins what distinguishes an engineering process from simple generation: programmatic verification of the result. The finished text is automatically run through a set of checks. The word count is calculated — it should fall within the specified range. Markup artifacts that break the layout when inserted into the content management system are searched for. It is checked whether Cyrillic and Latin characters have merged within one word — a typical typo of machine generation. The length of the title and the absence of prohibited elements are controlled. If the check is not passed, the text is refined before publication, not after. Formally, this is ordinary unit testing, only applied to an article, not code.
Such a double circuit — first checking input facts, then checking the output text — removes most of the complaints about automatic content. It does not make the article brilliant but reliably protects against gross errors that most damage reputation.
PUBLICATION ON THE SITE AND IN SOCIAL NETWORKS
Writing the text is half the battle. The second half is delivering it to the reader in several channels at once. In our case, the article is created in an internal panel, from where it is sent with one action to both the site and linked social networks. The cover is generated by the system itself so as not to depend on manual designer work for each material.
Here begins the most fragile part of any multiplatform publication — external services. The site is under our control, and the material goes there predictably. But social networks live by their own rules: access tokens expire, limits change, responses come with delays of tens of seconds. Therefore, publication in each channel is considered a separate operation with its own status. If the site and part of the networks accept the material, and one platform rejects the request due to an expired token, this is recorded, and the remaining channels still receive the article. The system does not roll everything back due to one error.
SCHEDULE AND PROTECTION AGAINST DUPLICATES
Regularity is more important than spurts. Search engines and subscribers appreciate a predictable rhythm, so the conveyor works on a schedule — on specified days and hours. But the schedule has a tricky problem: what to do if the computer was turned off at the right moment or the task did not work. Naive repeat logic easily leads to duplicates when the same slot is published twice.
We solve this through a state file. Each time slot has a record: which article is created, where it is located, on which platforms it has already been sent. Before publishing anything, the process looks into this file. If the slot is already fully closed, it does nothing. If the article is created but some social networks have not accepted it, the next launch does not write a new text but closes exactly those platforms that failed, on the same article. This approach is called idempotency: no matter how many times the task for one slot is launched, the result is the same — exactly one article, without duplicates. This is a simple idea, but it is what distinguishes a reliable conveyor from a script that once spams the feed with a dozen copies.
LIMITATIONS AND WEAKNESSES
We deliberately do not present this process as a magic button. It has honest limitations that should be known in advance.
-
Depth of expertise. The conveyor handles overview and practical articles well but will not replace material where unique internal experience, fresh data from a specific project, or an unconventional author's thought is needed. Such texts are still written by humans.
-
Dependence on external services. Publication on social networks depends on foreign tokens, limits, and platform whims. An expired access key stops the channel, and without human intervention, it cannot be reconnected.
-
Risk of monotony. A single template provides stable quality, but with a large flow, articles begin to sound similar. Diversity of topics and presentation still needs to be monitored manually.
-
Fact-checking is not absolute. Search verification eliminates most errors but not all. If sources contradict each other or the topic is too fresh, the final decision must be made by a human.
-
Search engine behavior is unpredictable. The right choice of topic increases the chances but does not guarantee a rise in positions. Algorithms change, competitors also publish, and some materials will not succeed.
-
Responsibility remains with the company. Automation does not remove legal and reputational responsibility for what is published. A mistake in numbers will be attributed to the brand, not the algorithm.
WHAT THIS GIVES IN PRACTICE
For a content specialist, such a conveyor removes routine: there is no need to manually search for a topic, check against the list of published materials, and distribute content across channels. Time is freed up for what the machine cannot do — deep authorial texts and work with real cases. It is important here not to fight with the tool but to redistribute efforts.
For a department head, this should be seen as a way to maintain a stable publication rhythm without team burnout. Key metrics are not the number of articles but the share of materials that passed the check without edits and the dynamics of positions on target queries. The conveyor makes sense where a constant flow of overview content is needed and does not make sense where one strong publication per quarter is more important.
For the owner, the most important thing is a sober assessment of risks and benefits. Automation of publications saves money and time but shifts the burden to quality control and platform integration support. This is not a reason to fire the editorial team but a reason to redirect it to more valuable work. A reasonable model is for the machine to handle flow and routine, while the human is responsible for strategy, complex topics, and final responsibility.
FREQUENTLY ASKED QUESTIONS
Will such a conveyor replace a live author?
No. It covers regular overview and practical content but does not replace materials where unique experience and authorial position are needed. The best result is given by a combination: the machine handles the flow, the human writes key texts and controls quality.
Will the search engine penalize for automatic content?
Search engines assess the benefit to the reader, not the method of text creation. Problems arise with meaningless mass content without fact-checking. Our approach with source verification and quality control reduces this risk, but no one gives a full guarantee of position growth.
What happens if a social network publication fails?
Each platform is processed separately. The site and successful channels receive the article immediately, and the failed platform is recorded in the state file. The next launch closes exactly it on the same article, without creating a duplicate.
Where to start if you want to replicate this yourself?
Start not with generation but with two things — data on demand for topics and fact-checking before writing. These two circuits separate a useful conveyor from a plausible text generator. Publication and scheduling can be added later when text quality is stable.