Google’s line: a person checks the AI draft before it goes up
Google’s page on generative AI content, last updated 1 October 2026, added a plain paragraph. The doc says generative models do not retrieve facts. They predict a likely next word from training data, so the output can be wrong. It then says it is critical to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing. The same paragraph includes titles, meta descriptions, structured data, and image alt text, because those strings show up in Search.
Critical is a word their docs spend rarely. For a small firm it means the draft may start in a model, and the version that goes on the site is read by the person who owns that page. “Reads smoothly” is not the check. The check is a fact that would send a buyer to the wrong person or a promise you do not make. On a site like ours, drafts invent a phone, a LINE ID, a rank promise, or a price band we do not use. We do not publish a phone or LINE. Contact is the form. If the draft says “call us,” that sentence stays off the site, however clean the prose is.
Before a service page ships, the person who owns that page checks the items below. You do not need a twenty-page list. The price band on the page is the band you quote to buyers now, not a band the model wrote. The address is the base in Taipei, Zhongshan, Fuxing North Road. Do not let the model change it to Xinyi, and do not write only “we serve the whole country” with no base. Check whether work you do and work you do not do got swapped: something you do not offer written as if you do, or something you do offer written as if you do not. Both get corrected. FAQ answers have to match the sentences a buyer can see on the page. Image alt text describes only what is actually in the picture, not a certification badge the picture does not contain. When you add structured data, repeat only sentences visible on the page. Do not add a phone or a price that exists only in the markup.
Titles and descriptions are easy to generate ten at a time. Search shows one of them. If that one says you are first in Taipei or that you guarantee a homepage slot, a buyer sees that sentence before the body, and treats it as your promise sooner. The person who publishes reads the title and the meta description first, then the body. Google put those two lines in the must-check list, not only the middle of the article.
This sits next to the spam updates, not instead of them. A large set of pages nobody has read can hit the policy. A page a person has read, about work you actually do, is not that set. The earlier note on not washing the blog with the same skeleton is about how often you publish, and whether each post is the same frame. This doc adds: even one post a week still needs a person to have read the numbers and the contact line in that post. When we cite Search Console, we use figures from an export we can match. If the export has no click-through rate, we do not let the model type one in.
Who checks belongs on a calendar, not only in a prompt. On a small team that is the person who owns the page. Before it goes up they read it themselves, correct a wrong phone, address or price, and then upload. When a draft comes back from outside, you still check the price and the address yourself. Replying “the copy is fine” is not that check. Writing “do not invent a phone” in the prompt does not make the model obey forever, so the check stays at the step where a person reads the draft, not at the prompt. If the price band on the service page changes, it changes to the band you checked, not to the first number the model replied with in the chat.
Two drafts that show up before upload
One service intro is tidy and ends with “hotline 02-xxxx.” We do not have that number. The person who publishes does not let that version go up. Delete the phone, point to the form, and search the whole file for “hotline” and “LINE ID.” If an FAQ in the markup copied that same phone sentence, delete that sentence too. Do not delete only the line a buyer can see on the page.
The other is a blog post. The date in the opening is right. The middle says conversion rose forty percent last quarter. The company has no conversion report that matches the words “forty percent.” Cut that sentence. If you write about search, write only what you can match, for example: the brand name spelled in phonetic form has impressions in Search Console, few people click through, and the service queries sit in a weak position. A percentage with no report behind it stays out.
Questions
Does using AI on the site get a penalty?
Google spam rules target a large set of pages that do not help the person reading them, not a draft that once passed through a model. If the person responsible has checked the page, and the page describes a service you actually sell, it can go up. A whole site generated with nobody reading it can hit scaled-content abuse.
Do FAQ markup and titles count?
Yes. The help page lists structured data, titles, meta descriptions, and image alt text because they show in Search results. If the markup contains a phone or a price the page does not show, it disagrees with the text a buyer can see, and the person publishing deletes that extra sentence from the markup.
Does a weekly note need the same pass?
Yes. The person who publishes checks the news date, the product name, and your own figures before the note goes up. A sentence you cannot match to a source gets cut. Do not keep it just to hold the length.
A page someone has read Markup that matches the page
Source: Google Search Central, generative AI content, page marked last updated 2026-10-01 UTC.