What happened

If you use AI to write your website’s articles or product descriptions, two recent announcements are worth your attention. The first is the one closest to home: Google is reminding sites that AI-written content should have its facts checked by a person before it’s published. The second happened in the EU: text that ChatGPT writes there is about to carry an invisible mark. Side by side, they happen to touch the same question: when AI writes it, who checks that it’s right?

Google first. On 1 October it updated a guide for site owners on using AI-generated content on their websites. The guide says that when AI writes, it doesn’t look facts up; it predicts which words are likely to come next, so what it produces may contain errors. That is why it calls manually fact-checking all AI-generated content before publishing “critical”. The check covers more than the body text: the blue page title in Google’s search results and the few lines of description beneath it count too. So does text written into a page’s code, such as image “alt text” and the “structured data” written for search engines: you don’t see it when you browse the site, but Google may show it in search results, so it needs checking as well. The guide is written as advice and, on the fact-checking step, doesn’t say what happens if you skip it; its emphasis is on content being accurate.

The other announcement came from OpenAI on 5 October: over the coming weeks, text that users in the EU generate with ChatGPT and its coding tool Codex will carry an invisible watermark (OpenAI calls it textGrain), because the EU AI Act requires AI-generated text to be identifiable by machines (we covered that rule in August). For ChatGPT and Codex, this round applies only in the EU and won’t be a global default at launch, so text written with ChatGPT outside the EU isn’t covered.

The tool that checks for the watermark isn’t being made public yet either; only approved researchers and expert organisations can apply to use it, partly because detection can miss watermarks and can also flag text wrongly. OpenAI’s own tests found that swapping one word in ten for a synonym in English text cut the detection rate from about 92% to 66%. The announcement also states plainly that a watermark can’t tell you whether a passage is true.

Why it matters to you

The two announcements came from two different companies, independently. Google’s has written “fact-check by hand before publishing” into its guidance for sites, with titles and descriptions named too. For any business producing content in bulk with AI and publishing after a quick glance, that shortcut is already getting less safe, because whatever the AI gets wrong reaches your customers first, in your name.

Yet the things AI easily gets wrong are often exactly the things only you know. Say you run a removals company and ask AI to write twenty service pages, one per district, in an afternoon. The prose flows, but the AI has no idea whether you charge extra at weekends, whether boxes are free or cost extra, or how you price moves between cities, so it fills the gap with what removal companies usually say. It sounds perfectly reasonable, and it doesn’t match how you actually work.

Once those sentences are live, a customer may ring up one day saying “your website says there’s no weekend surcharge”, and whoever answers can either explain or swallow the difference. Nor are the errors confined to the body text: the page titles and descriptions the AI wrote along the way can appear straight in Google’s results, so customers read them before they even click. ChatGPT, Gemini and other AI assistants also draw on web pages when they answer questions, so that wrong sentence can get passed on.

A watermark doesn’t help with any of this. Even if the detector one day opens up to everyone, it will only tell you “this carries OpenAI’s watermark”, not whether “no weekend surcharge” is right. Only your price list, your contracts and your crew know that.

What to do about it

Act now: keep using AI to write, but add a check before publishing. Every sentence about prices, service areas, timing or licences gets compared with the company’s real information before it goes out, so the site says what you actually do.

The easiest place to miss is the part you never look at. You usually only see the description under your page title if you search for yourself on Google; the alt text and structured data are harder still, because your web agency has to open the admin panel and pull them up page by page before anyone can read them. If AI wrote that too, often nobody who pressed “publish” has read it. And the two halves rarely meet: your web agency can pull up the text but doesn’t know your real prices, and you know the prices but can’t see the text.

When the words on your site are right, AI has a better chance of getting you right when it passes them on. Our managed GEO service can bring those two halves together: our team fixes your site’s content, brings what other websites say about you into line, and regularly checks how each AI engine describes you. You just review the results. We can produce new content for you too, with facts such as prices and services checked against the information you give us; if you’d rather write it yourselves, send us the drafts and we’ll edit them before they go live.

Put your name down for a demo: a session with one of our consultants, who shows you everything live. They’ll start with your website’s GEO audit score — geoweb.tw’s free audit, which checks how well AI can understand your site and scores it across 12 areas. Then they’ll put questions to ChatGPT that your customers often ask, without naming your business (something like “which movers in Taipei are the best value for a weekend move?”), and see whether it mentions you; if it does, whether what it says about your prices and service area matches what you actually do.

Further reading