---
title: "Editorial Policy: How AI Articles Are Written and Checked – Probeli AI"
description: "How every Probeli AI article is researched, written, fact-checked and scored before it reaches you, and how published facts are checked again."
url: "https://probeli.ai/editorial-policy"
---

Editorial policy

# How an article is made, *and who approves it.*

Every article is researched, written, checked and scored before it reaches you. *Nothing goes live without your OK.*

[Start free trial](https://app.probeli.ai/signup) [AI articles](https://probeli.ai/features/ai-articles)

The stages of a sample article for Webflow as the example brand: research read 18 of 20 pages and kept 42 facts, the brief planned 6 sections, a person approved it, then the sections, the fact review, the FAQ and meta and the images were done.

## The same steps, *for every article.*

## 1. Research of the top results

Probeli AI reads up to 8 of the top Google results for the article's keyword, in its market's country and language. Pages about the keyword's subject come first: in "best agency templates", "agency" counts and "best" doesn't. Code repositories and license files on the topic, and up to 2 of your own pages about it, count as primary sources. The words the top results share are taken only from pages on the subject.

It then plans what the article has to cover: its type (a list of picks, a how-to, a comparison, a review…), who it's for and what they came to do. The type follows the keyword's own words first ("best" or "top" asks for picks, "vs" for a comparison, "how to" for steps, a symptom such as "not getting hot" for a fix, a year's changes for news), and only then what the top results are; forum threads don't count. Someone who searched for "best" gets help choosing, whatever the format.

Before writing, it checks your own pages. When one already has the keyword in its title, the brief waits for your approval, so you can update that page instead. You can review the brief before any writing starts, and add evidence only you have, such as a customer's result or a quote. A case study, original research or an opinion piece is written only from evidence you add. Without it, the article becomes a roundup of published case studies or findings, or a guide to the decision, and the brief says what to add to write the original. A statistics roundup with few dated numbers lists only the numbers its sources give, and the brief says so.

## 2. A fact sheet from your own pages

Your products, plans, prices and claims come from a fact sheet built from your own website, read for each article. It's what the writer may say about you, and nothing more. Products that fit the keyword, such as an agency theme for "agency templates", are found on their own pages, and the writer reads those pages when it writes.

## 3. The writer

A long-context model writes the article from the brief, the research and the fact sheet, with structured data, images, sources and links to your own pages, and a FAQ where the article's type takes one and searchers ask questions about the topic.

- Your products that fit the topic are named in the part they fit, not only in the conclusion.
- No part is about the article itself or praises it, and the call to action says "top-rated" or "#1" only when your own pages do.
- Stock constructions such as "not just" or "rather than" are on the writer's list to avoid, and the checks count them.
- The summary says it in other words than the first paragraph, and the meta description holds the keyword. It is the writer's own, or built from the article's steps, checks, cases or figures with their sources, and the writer's is kept whenever the built one fails a check. It never repeats a sentence of the article word for word or cuts one short, and a summary that can't meet that is left out rather than padded.
- Your own row in a comparison table, and the row of every pick with a section of its own, is kept even when a source leaves some of its cells blank (they read "Not listed"); a blank price is filled only with the price a page that row already cites gives for that product. Your call to action reads plainly (no dashes or exclamation marks), in the article's own voice, and a part written to point readers to your offer keeps its link there. When the article has a FAQ, your call to action comes after it, as the article's last block.
- A table never loses a column's name, a column holding your own price or free trial stays, and a troubleshooting table keeps every symptom whose fix it gives. A figure the article credits to a named source is cited to that source and listed under Sources.
- When videos are on, a YouTube video is added only when its title names what the article is about and no competing platform, after the part it fits. It is never a video about a product you call yourself an alternative to, never one about another kind of the article's subject that the article gives no part of its own ("short cruises" for an article about cruises), and never a repair the article leaves to a professional unless it shows a safe step the article itself tells you to do. In a how-to or troubleshooting guide that may be any video showing one of the guide's own steps, such as a breaker reset or an element test with the power off; elsewhere it must be one Google lists, and its step names no breaker or panel work. Gas, pilot, burner, wiring, live electrical work, voltage checks and replacing an element or thermostat are never shown, in the title or the description, and an article that leaves gas work to a professional shows no gas or pilot video at all. A video Google lists for the keyword comes first, and a newer video is preferred to one named after a past year. A how-to guide prefers a video of ten minutes or less and places it right after its steps. No video is better than a wrong one.
- Pictures show a scene, never text: no screen, chart, title, platform name or brand, yours or another's, goes into an image prompt, since image models garble words. The cover shows the thing the topic is about (a person unclogging a bathroom sink, running shoes, a cruise), whatever your kind of business, else a scene that suits it, and its description says what was drawn. A part that names one of your products can show that product page's own preview image, credited to it; a pick in a list shows an image from its own site or code repository, never one from a review of it.
- A stock photo is never one tagged with another brand than the part's own; it must be tagged with two of the words it was searched for and, when its tags are in the article's language, with a word of its part's heading or of the keyword, and never shows a pandemic scene the part isn't about. Its first tag, when it names an object, is one the part names, and in an English article about a thing (a water heater, a bathroom sink, running shoes) the photo is tagged with that thing, never another kind of it: a part with no such photo gets none. Its description is the planned one when the photo shows everything searched for and every object that description names, else a sentence of what the photo shows. No paid picture is bought of a generic office or people scene. A picture never sits between a sentence ending in a colon and the list it introduces. An infographic is drawn only when it sums up a list that says more than its labels and that no table beside it already lists; a statistics or research article's infographic shows a summary table's figures as tiles, each with its source, and its date when the table gives one, never a comparison table's and never beside the table it draws, else a share list as a bar chart, and a part of such an article whose picture would be a chart shows its own figures as a bar chart instead of a stock photo. A bar chart compares figures of one measure on one base, never a change, and names its sources under the bars; figures no source credits get no chart. An infographic's description names its labels, never just how many there are.

## 4. Fact checks

Before anyone scores it, the draft's claims are checked against the sources it read and the fact sheet, each number against the page its marker cites. A figure that page gives is kept and cited to it, never cut, and what a page says is coming is read as of the page's own date. Steps and instructions, which state no fact, need no source.

## 5. A judge on a second model

A different model from the writer's scores the draft. A draft that fails a hard check (a figure no source gives, a claim your pages contradict, a missing call to action or price of yours), or that leaves out a page, price or product of yours the app has on file, gets a targeted rewrite (1 at most): the writer fixes only the problems named, on at most two parts, and keeps every other sentence word for word. A part it changed stays only if the judge scores that part higher, no score of the article falls and it loses nothing the draft had (a step, a table row, a cited figure, a safety step, a link to your pages); otherwise the part stays as it was. A draft with only some of its changes kept arrives marked Needs review, since no judge read it whole; one that already passed keeps its changes only when every changed part stays. Nothing is carried over from a version judged worse, and a weak score alone never buys a rewrite. If it still doesn't pass, it arrives marked Needs review, with the weak parts named. A refresh written without an article type keeps the earlier check and may get a second rewrite. Your call to action's words and page are set by code, so a rewrite that changes or drops it gets it back. A rewrite that drops a subheading, a step that turns the power, gas or water off, a table column, a named case or an answer no other part gives, or adds a stock construction, hearsay or a statistic with no citation, is refused and the part stays as it was. Code's own cuts never remove a safety step either. Before the draft arrives, code itself fixes a checklist count that disagrees with the list, a scorecard that promises more rows than it has, a comparison row left blank for one side, an intro over its length and a sentence said twice.

## 6. Code checks

Rules that no model decides: a price of yours that isn't on your pages, a plan given the wrong price, an article that recommends a competitor (a table of other sellers' products that leaves out yours counts as one), a quotation no source holds, or a title or heading claiming a test, proof or a #1 rank fails the check, and so does an article for your business that never links your call to action. A fact check's finding still in the text holds the article for review, until the sentence cites a page that gives its figure. A cited number its page doesn't give is listed for you to check.

## 7. You approve

Every article arrives as a draft with its score and the checks behind it. You read it, edit it and publish it. Autopilot can publish for you only if you turn that on, and an article that needs review stays a draft.

## Published facts, *checked again every month.*

About once a month, Probeli AI reads the sources behind a published article's dated claims again and updates the numbers that changed.

- Up to 8 source pages read again per article
- A sentence you edited is flagged, never changed
- On by default, and you can turn it off per workspace or per article
- A fact check never uses an article

[Keeping articles up to date](https://probeli.ai/faq#articles-and-publishing--refresh)

The article editor with a sample article for Webflow as the example brand, “How to choose a website builder for a small business”: its opening paragraph and headings, the focus keyword with 6,600 searches a month and the outline.

## Questions about *AI content.*

**Does AI-written content hurt SEO?**

Not by being AI-written. Google's guidance judges content by how helpful it is, not by how it was made, and thin or copied pages do badly whoever writes them.

So each article starts with research: Probeli AI reads up to 8 of the top Google results for its keyword and takes product facts from your own pages. The draft gets a FAQ, structured data, sources and images, and is checked before it reaches you. No tool can promise rankings or AI citations, and we don't.

**Do I approve articles before they go live?**

Yes. Every article arrives as a draft. You read it, edit it in the editor (or ask for an AI edit of a passage, a section or one of its subsections) and publish it when it's ready.

Autopilot plans and writes articles on a schedule you set, and shows what it will cost before it starts. By default its articles arrive as drafts too.

**Can Probeli AI keep published articles up to date?**

Yes. A refresh reads a published article against fresh results, rewrites its weak sections and adds what's missing, and leaves the parts you edited alone. You see every change before you publish it. A refresh uses one of your monthly refreshes, never an article, and gives it back if nothing needed changing. Grow includes 5 refreshes a month, Pro 10 and Scale 100, and the free trial 3. Track writes no articles, so it has no refreshes. Suggestions for one of your own pages draw on the same count.

With fact checks on, about once a month Probeli AI also reads the sources behind a published article's dated claims again, up to 8 pages, and updates the numbers that changed. A sentence you edited is flagged, never changed. A fact check doesn't use an article.

**Which languages can articles be written in?**

English, French, German and Spanish. You pick the language and the country with each keyword, so a workspace can write for several markets, and the research reads the results of that country.

[See all questions](https://probeli.ai/faq#articles-and-publishing)

## Read a scored draft *before you decide.*

Start your free trial and get your first articles as drafts, with the checks behind each score.

[Start free trial](https://app.probeli.ai/signup) [Explore demo](https://app.probeli.ai/demo)

7-day free trial, card required · 3 articles included