道場Dojo To the corridor
A samurai leans over a printout under a desk lamp and marks lines with a red pen.
All posts

道Dojo · · 3 min read

Google now says it plainly: fact-check AI copy before you publish

Since 1 October, Google's guidance says AI content must be fact-checked by hand before publishing, alt text and meta descriptions included. Where the errors really hide on company websites.

An AI wrote the samurai a profile: "confident with touchscreens". Deleting that line took him four attempts.

Sounds plausible, is wrong. That is the problem with AI copy on company websites, and Google now says so in writing. On 1 October 2026 Google updated its guidance on AI-generated content. The key new sentence: it is "critical" to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing.

What actually changed

According to Search Engine Journal (1 October 2026), it was the only change listed that day in Google's documentation changelog. The new part is the reasoning: language models don't look up facts, they predict a likely sequence of words. So their output can contain errors, which Google explicitly calls hallucinations.

Worth knowing:

  • No new ranking signal. This is a documentation change. Google points to its Search Quality Rater guidelines and states itself that rater scores don't directly influence ranking.
  • AI isn't banned. But using it to churn out lots of pages without adding value can breach the spam policy on "scaled content abuse". The rater guidelines cover that in section 4.6.5, low-effort, low-originality content in 4.6.6.
  • Metadata counts too. Google says the review applies to titles, meta descriptions, structured data and alt text, because they can show up in search results.

Where AI errors really hide on company sites

Our opinion: the risk rarely sits in the blog post. It sits in the sentences nobody reads twice. Founding year, certifications, delivery area, opening hours, guarantees. An AI fills gaps with whatever is typical for similar companies. It looks credible right up to the moment a customer invokes the guarantee you never offered.

The second weak spot is metadata, because it lives in CMS fields nobody ever expands. AI alt text: "A smiling team in a modern office." The photo shows a forklift.

Wrong alt text hurts twice. It misleads people using screen readers, and according to Google it can appear in search.

Reviewing doesn't mean skimming to see if it reads smoothly. AI copy always reads smoothly. Reviewing means every statement that could be proven gets proven or cut. That should be done by someone who knows the business, not just the language.

At MED1A we write website copy so that, in the end, someone from your company has signed off every factual claim. Whatever tool wrote the first draft.

This week's drill

Take 30 minutes and the five most important pages of your website: home, about, your two key services, contact.

  1. Mark every sentence that states a fact: years, figures, certifications, partners, guarantees, times, places.
  2. Tick off each one you can back up. Correct or delete the rest.
  3. Export the alt texts and meta descriptions of those pages as a list, from a crawling tool or your CMS, and compare them with what is actually in the images and on the pages.

If nothing turns up: good. If something does, it was worth the half hour. And if any page says "confident with touchscreens", check with the samurai first.

Sources

  1. Google Search Central: Guidance on using generative AI content on your website (updated 2026-10-01)
  2. Search Engine Journal: Google Tells Sites To Fact-Check AI Content Before Publishing (2026-10-01)
  3. Google Search Central: Latest documentation updates (October 2026)

Topics: ki-textecontentseoredaktion

読Read on