An online analytics school in Almaty has a guide for new students, "How to get ready for your first class". The school wants to put an open version of it on the blog: maybe people who are still choosing a course will find it through search. The marketer hands the guide to an AI model and a minute later has a 1,200-word article, "What you need to start learning analytics". It has figures, links to sources and a confident tone. It also has mistakes: the guide says "8 GB of RAM, no Excel needed", and the article says "a laptop with at least 16 GB and Excel 2021". The school's numbers are made up.
The model needs material, not a role
A prompt like "You are a senior SEO copywriter with ten years of experience" adds no facts. A model that calls itself an expert knows exactly as much about the school as it was given. So instead of a role, it gets material:
- the guide itself, all five sections: laptop, accounts, files for the first class, the report in Data Studio, what to do if something doesn't work;
- the 48 questions the curators got in the first week of the September cohort: these are the reader's questions;
- the course syllabus;
- a list of what the school doesn't do and doesn't promise;
- the brand voice rules with examples.
With that material the model is useful: it sorts the answers into sections, turns the curators' shorthand into language for someone who isn't a student yet, and suggests headlines. What it can't do is find out what has changed since the guide was written. Anything the model adds beyond the material counts as unverified until the editor finds a source.
Break the article into claims
Reading top to bottom and fixing as you go is a poor tactic: smooth text lulls you. First you list everything in it that can be verified. The school's article had five such places:
| Claim in the article | What the source says | Decision |
|---|---|---|
| You need a laptop with at least 16 GB of RAM | Guide: Windows or macOS, 8 GB | Change to 8 GB |
| Install Excel 2021 | Guide: no Excel needed, the course runs in Google Sheets | Remove |
| The average analyst salary in Almaty is 650,000 ₸ (with a link) | The link leads to a 2023 article about salaries in Moscow | Remove the paragraph |
| On the course you'll learn to connect any CRM | Syllabus: a lesson on exporting leads from a CRM, covering two systems | Say what the syllabus covers |
| More in the Data Studio documentation (link) | The link doesn't open | Replace with a working link or remove |
The first two rows take a minute to catch against the guide. The third and fifth need the link opened. The fourth is the nastiest: it promises a capability the course doesn't have. Someone whose company runs a third CRM will sign up for something they won't get, and two weeks later ask for a refund.
The picture shows the same fix in general form: the product description lists two systems, so the text says two.
The words that stretch a promise furthest are "any", "all", "always" and "guaranteed". Search the text for them and check each against the description: if the syllabus covers two CRMs, "any" becomes "two". And if the model has added a quote from a "graduate", it goes entirely: an invented review can't be fixed, only deleted.
Open links and read them to the end
Links fail in two different ways. The first: the link doesn't open, because the model assembled an address that looks real. Replace it with a working one or remove it. The second is worse: the link opens, but the page is about something else. An article about salaries in Moscow in 2023 doesn't confirm a salary in Almaty in 2026: different market, different year, different currency. A reader who opens it will stop trusting the rest of the text.
The rule is simple: for every link, copy out the sentence on the page that supports the sentence in the article. If there is no such sentence, there is no source, only a link.
The salary was settled more simply: no new source was needed. The article answers "what do I need to prepare for the course", and an analyst's salary doesn't belong in it. The school collects no data on its graduates' salaries, and neither the model nor the marketer may invent them on its behalf.
The article has to answer the reader's question
The reader arrives asking "can I handle the course, and what do I need to get ready". The model opened with three paragraphs about data being the new oil and the history of analytics, and hid the answer in the fifth. The editor reorders: the first paragraph gives the short answer, and the five sections from the guide follow.
For the course you need a Windows or macOS laptop with 8 GB of RAM, the Chrome browser and a Google account. You won't need Excel: all assignments are done in Google Sheets.
Say who the course is too early for: without a laptop, because the assignments can't be done on a phone. That boundary filters out people who would otherwise sign up and be disappointed.
The voice is fixed by facts too. The model readily writes "a unique proprietary method" and "the market's best experts". The school's voice uses what can be verified instead: a curator checks the assignments, each curator has 20 students, and during the course every student builds their own sales report.
If the article is translated into Kazakh, the model can produce a first-pass translation, but a native speaker who knows the subject reads and fixes it. Machine translation may name a menu item differently from the interface, and the student will look for a button that isn't there.
What Google says about AI-generated text
Back in 2023 Google explained that it rewards high-quality content however it is produced and treats automated generation aimed at manipulating rankings as spam. Its guidance on AI-generated content, last updated in October 2026, is direct: all generated content needs a manual accuracy check. That includes metadata, the title element, the meta description, structured data and image alt text, because they show up in search results. The same page warns that generating many pages without adding value for users may violate the scaled content abuse policy.
For the school the conclusion is simple. The temptation to make fifty pages, "analytics course in Astana", "in Shymkent", "in Karaganda", with identical text and the city swapped in, is exactly what Google warns against. One article that answers students' real questions is worth more.
Whether the article works in search shows up after two or three months in Search Console under the page's address, just as for the guide.
Before publishing
Spelling is left off the list on purpose. A well-spelled text with an invented salary is more dangerous than one with a typo. Readers forgive a typo, not an invented salary.
Exercise: find the lost condition
The school turns the article into a short Telegram post. The guide says:
The Data Studio report opens from the link if you sign in with the email address the curator gave access to. If the report asks for access, write to the curator and attach a screenshot.
The model cut this to one sentence: "All reports open from the link, no access needed." Find what was lost and write a version under 150 characters with the condition back in place. Then do the same with your own text: take any material a model shortened for your project and compare it with the source sentence by sentence.
In the school's example two things were lost: the condition (sign in with the email the curator knows) and the fallback (write to the curator and attach a screenshot). And "no access needed" says the opposite of the source: the model dropped the condition and kept a conclusion that is wrong without it. The word "all" is the model's addition as well. A working version: "The Data Studio report opens from the link if you sign in with the email you gave your curator. Asked for access? Send the curator a screenshot." It is 144 characters long.
In your own text, look for the same things: words like "all", "any", "without" and "instantly" that the source doesn't contain, and sentences where an "if" has gone missing. Each one either goes back to the original wording or comes out. A student who fails to open the report from the shortened post will message a curator, and the saving on length turns into one more question in the chat.