AI Blog Writing: How I Check Every Claim

Short answer

• What AI takes over: the outline, the first draft, the title and the description. Google calls generative AI particularly useful when researching a topic and for adding structure to original content.
• Where AI writing goes wrong: the facts. Language models sometimes guess when they are uncertain and produce plausible but incorrect statements, according to a 2025 paper by Adam Tauman Kalai and colleagues.
• Google and AI content: Google does not penalise AI content as such. Its spam policies target pages generated at scale to manipulate rankings without helping users, no matter how they were created.
• What has to happen before you publish: every claim checked against its source, including the title, meta description and image alt text. In my pipeline the article text is checked by a freshly started model that did not write it. In the first five articles through my pipeline, its first pass fixed 35 of 304 claims.
• Who this suits: businesses that publish expert content regularly and have one person who answers for what goes out.

A number without a record

What the first version promised

On 2 April 2026, the first version of this article went live on my website, in German. It promised to cut the work on one blog post from 60 minutes to 13. On 10 April I unpublished it. Other pages of my website kept the figure until 1 October 2026.

Outside the website, there was no record behind that figure, so it has not been in my facts register since 1 October 2026. That register is the list of every claim about my work an article may use. The number is gone.

Why a number about your own work needs a record

It was not the only one. On 1 October 2026 I took four figures without a record off my website. Two of them, 60 to 13 included, even had a source in my facts register: the text of my own website. The site was citing itself.

If a figure like that is in the material a language model works from, the model can carry it into the next text without hesitation. It cannot tell from the figure alone that no record backs it. What survives from the first version is the pipeline this article was rebuilt with.

What AI takes over when you write, and what it cannot

Outline, draft, title and description

Google's guidance on generative AI calls it particularly useful when researching a topic and for adding structure to original content. I leave the outline, the first draft, the title and the description to a language model.

Why AI text sounds the same

"Many AI texts are like hotel music: pleasant, clean, professional and forgotten after five minutes." Someone wrote that in the comments under a Tages-Anzeiger article from August 2026 (my translation). In the comment sections of Swiss news sites I analysed for this article, this was the most common kind of objection under articles about writing, media, advertising and books.

Against that, the draft step of my pipeline carries a list of banned patterns. One rule bans reflexive lists of three, another bans stock transitions such as "With that in mind" or "Let's dive in".

Banning stock phrases does not create substance, though. What a model does not have is your numbers and your experience with your customers. Whatever sets your text apart from the other results for the same search has to come from you.

Why language models make up facts

A 2025 paper by Adam Tauman Kalai and colleagues puts it this way: language models "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty." According to the paper, the cause is how models are trained and evaluated. Both reward guessing over admitting that the model does not know.

For a blog post, that means the wrong sentence looks exactly like the correct ones around it. Reading for style will not find it.

What Google says about AI content

Google does not penalise AI content as such. Its spam policies target scaled content abuse: many pages generated mainly to manipulate rankings, with little to no value for users, no matter how they were created. In its guidance on generative AI, Google writes that generating many pages without adding value for users may violate that policy.

The line is drawn by value, not by the tool. In the same guidance, Google calls it critical to manually fact-check all AI-generated content before publishing, including title elements, meta descriptions, structured data and image alt text.

My pipeline for one blog post

Since late September 2026, my blog posts have run through one pipeline, this one included.

The topic comes from search data

I start with searches from Switzerland: volumes from the Google Ads Keyword Planner, plus the queries my site already shows up for in Search Console. For this English version, the volume check came back with these figures:

The figures are for Switzerland, with English as the search language. The German version of this article targets "ki texte schreiben", with 1,300 a month on its own.

Google Ads returns close variants of a keyword as separate rows with identical numbers, and the same keyword checked with German and with English as the search language comes back twice. My first topic plan added those rows up and came out at up to three times the real figure per topic. Since then my pipeline works out which rows mean the same keyword before it adds anything up.

Outline, draft and a style pass

A language model writes the outline, a short research brief with a source for every fact, and then the draft. The research step does not open pages by agencies or consultants on the same topic, so nothing from there travels into my text unchecked.

A style pass follows: rhythm, cuts, a clear position. It may not add a single new fact.

A check by a model that did not write the text

Then comes the fact check, my auditor: it did not keep the books and wants to see every receipt. A freshly started model that had no part in the text reopens every source and gives each claim a verdict. A copy review then checks the tone and the words that sound machine-made.

Both reviews are stamped on a fingerprint of the final text. If I change a sentence afterwards, the stamp no longer matches, and the pipeline writes the article to my website only once both reviews are stamped again on the new text.

Nothing goes live without my go

At the end, the pipeline writes the article to my website as a hidden draft. It does not overwrite a published article. A post goes live only when I release it.

My view after the articles so far: with AI, writing has become the cheap part. The work is in the check. Skip it and you cut corners exactly where AI text goes wrong.

What the check found in five articles

Source: my pipeline's fact-check reports, first pass of each article, 28 to 30 September 2026.

About one claim in nine did not hold as written in the first pass. None had to be cut, 35 were fixed. The German Agentic Commerce article was adapted from the English version, which had already been checked, and needed one fix.

A trend figure that went stale within weeks

My brief for the Agentic Commerce article said that growth in traffic from AI sources to US retail sites was slowing. It rested on Adobe's figure for July 2026: up 62% year on year. The check found Adobe's report of 28 September, with growth of 127% for August. The claim about slowing growth was corrected.

A quote that was right and still wrong

A study by Vercel and MERJ from December 2024 states that none of the major AI crawlers currently render JavaScript. In the Agentic Commerce article, the English fact check itself wrote that sentence into the text, and the German one confirmed it. Further down, the same page names two exceptions: Google's AI assistant, which uses Googlebot's infrastructure, and AppleBot.

Only the final review, which read the whole page, caught the missing exceptions. Since then, my fact check counts a dropped exception as an error even when the quote is exact.

What you can take from my pipeline

A register for every number about your company

Keep a list of every number and claim about your company that your texts may use, each with its record, such as a report or a log. About your company, the AI learns only what is on that list. What is not on it does not go into the text.

Checked by someone who did not write it

Have the check done by someone who had no part in the text, a person or a second model. That reviewer should open every source instead of judging from memory.

That goes for fixes too. In the Forward Deployed Engineer article, a fix claimed that no Zurich posting checked, at Google or at Databricks, states pay.

The fixer had opened one posting without a pay line and assumed the others had none either. One of them does give a range, for France. Only the re-check of every changed line found it.

Never publish automatically

Let AI write drafts, but never let it publish. It is the same rule as for the AI agent in my shop inbox: it saves drafts but may not send them.

What stays out of the prompt

Keep customer data and anything confidential out of your prompts. If customer data has to go in, remove the personal details first.

When this is not worth it

If you write two texts a year, skip the pipeline and the stamps and read every source carefully. And if nobody in your company takes responsibility for what gets published, the best check will not help.

Records first, tools second

Before you pick a tool, write down which numbers about your company have a record and where that record is kept. That list protects your texts more than any prompt. If you then want a pipeline for your own blog, with search data, checks and your approval before anything goes live, I will go through it with you in an initial call. What I build in that kind of project is on the AI automation page.

Frequently asked questions