AI content and SEO

What has Google actually said about AI-written text?

Google's position has been written on the official Search Central blog since February 2023 and has not changed since. The heading of that page's main section answers the question by itself: rewarding high-quality content, however it is produced.

The sentence to memorise is this one: using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results is a violation of the spam policies. Split it in half. The first half does not condemn the tool. The second half condemns the purpose.

In the FAQ of that same page Google gets blunter. First: using AI does not give content any special gains, it is just content. And second, which we think settles the decision for most businesses: if you see AI as an inexpensive, easy way to game search rankings, then no.

So AI content is not a separate category in Google's vocabulary at all. The category that does exist has a different name, and the next section is about that one. You can read the full position in Google's own guidance about AI-generated content.

Is the line drawn on the tool or on the number of pages?

On the number of pages and the value of each one. In March 2024 Google announced three new spam policies, and one of them lands exactly here: scaled content abuse. The spam policy document defines it as generating many pages whose primary purpose is manipulating search rankings rather than helping users, typically producing large amounts of unoriginal content that provides little to no value, no matter how it is created.

Google explains in that same announcement why the policy was new: the previous one was about automatically generated content, and this one replaced it so that it makes no difference whether the content came from automation, from human effort, or from some combination of the two. The escape route many people were looking for was closed before they got there.

The examples the spam document gives are worth reading, because not one of them is about writing style:

• using generative AI tools to produce many pages without adding value for users
• scraping feeds, search results or other people's content into many pages, including through automated translation or synonymising
• stitching together pieces of several pages without adding value
• creating multiple sites to hide the scaled nature of the content

Read the list once more and notice that no row asks who wrote the text. Every row asks how many, and what was added. That difference is the difference between one page drafted with a model and a two hundred page shelf produced from one template with the keyword swapped; the second is covered by this policy even if a person typed every page.

Put the same thing on one line and you get four modes. Help writing one page sits outside the spam policy, and so does a machine draft, edited and sourced. Bulk publishing, light edit, is the borderline, and the answer there depends on what each page actually added. And bulk generation from one template is the scaled content abuse the policy was written about.

The reverse holds too. If you have a page that tells a real experiment, a measured number or a failed attempt, and the model only wrote the first draft, no clause of this policy touches you. The full definition is in Google's spam policies.

Why does the pattern show even with no detector involved?

This section is the most important part of the lesson and the part least often written in Persian.

Everybody worries about sentences: does this paragraph sound machine written. But what a crawler sees without any special tooling is not a sentence, it is the shape of the site. Picture a site that had twelve pages for eighteen months and then publishes a hundred and sixty new ones in a week. Every one of them has exactly four second-level headings. Every one opens with a definition sentence and closes with a summary paragraph. Every one carries the same internal link block at the end. Not one of them contains a measured number, a client name, a real date or a linked source.

None of that needs a text detector. Publish dates, heading structure, template similarity, the internal link graph and the amount of textual overlap are things any search engine already computes, for other reasons. The pattern is visible in data that has been collected for years.

We cannot see inside Google's systems and nobody outside Google can either; take that sentence seriously, because anyone claiming otherwise is selling a guess. What we can say is this: the written policy names quantity and value rather than style, and quantity and value are exactly the two things that are visible from outside in the pattern above.

The practical consequence is a simple rule. If you want to know whether your publishing carries risk, do not look at the text; look at your list of URLs. If the dates, the lengths and the structures all resemble each other, and no page carries something only you know, the risk is right there, even if you typed every sentence yourself.

Can you trust an AI text detector?

Not enough to make a decision on, and it is Google DeepMind that wrote that, not us.

On the same page that introduces the SynthID watermark there is a passage about the detection tools in common use: those tools are usually classifiers, classifiers often only perform well on particular tasks, and when the same classifier is applied across different kinds of platform and content its performance is not always reliable or consistent. The consequence, in that text's own words, is that a piece of writing can be mislabelled, for example incorrectly identified as AI-generated.

So the percentage those tools hand you is meaningless in both directions. Rejecting a writer over a score of eighty-two percent is not defensible, and neither is publishing a weak text because the score was four percent. No clause of Google's policies is about the output of these tools either; the policy is about something else, which you read in the last two sections.

One note for the Persian market, stated honestly as a limit of our own knowledge: we have not seen a published evaluation on Persian text from any of these tools. If you find one, read the sample size and the method before trusting it. Deciding on a number that was never evaluated for your language is guessing about a guess.

Our position is plain: if your editorial process depends on getting past a detector, you have built the process around the wrong thing. The thing you have to get past is a reader who came looking for an answer.

What is SynthID and what does it mean for your site?

SynthID is Google DeepMind's watermark: a digital mark embedded directly into AI-generated images, audio, text or video. According to DeepMind's own page the watermark is embedded across Google's generative AI consumer products, is imperceptible to humans, and can be detected by SynthID's own technology.

How it works for text is interesting and worth understanding. A language model builds text one token at a time and gives each candidate a probability score. SynthID nudges those scores at the moment of generation, and that pattern of nudges is the watermark itself. For images and video the mark is added as the content is created, and is designed to survive cropping, filters and lossy compression.

For checking, the official page gives two routes: upload the file to Gemini and ask whether it was created or altered by Google AI, or use the separate SynthID Detector portal, which is currently being tested with journalists and media professionals.

Now the most important part, which the same page states honestly and which is rarely quoted. The text watermark holds up under cropping, changing a few words and mild paraphrasing, but its confidence scores can be greatly reduced when a text is thoroughly rewritten or translated into another language. It is also weaker on short factual answers, because there is less room to adjust probabilities without touching the accuracy of the answer. And DeepMind writes that SynthID is not a silver bullet.

The consequence for you is two sentences. First, a negative check does not prove a human wrote it; it means no watermark was found. Second, none of this is about ranking; it is about whether somebody can later show that an image or a text came out of a Google model. For a site publishing ten generated images a day, that is a reputational fact rather than a ranking factor.

So how do you actually use AI in your publishing?

The rule we run on this site is one sentence: the model gets the drafting and the checking, a person keeps ownership of the facts, the position and the limit.

There are three things a model cannot bring, and those three are what separate one page from two hundred competing ones. A number you measured yourself. A case where this very advice did not work. And the boundary past which the advice stops being true. If a page has none of the three, publishing it only makes the shelf longer.

The order of work we suggest is simple too. First decide what the page carries that exists nowhere else; if you have no answer, stop here. Then take the draft from a model to get there faster. Then attach every factual claim to a source or a measurement, and delete whatever cannot be attached. Last, run the mechanical check, which is in the fast path further down this page.

If you are doing this for a whole site and do not have the capacity, we do this work as content production; but that is your decision and not the conclusion of this lesson.

One last thing experience taught us: for most sites the serious danger is not a Google penalty. The danger is that six months later you own a hundred pages that have no reason to exist, and every change to that site is expensive from then on. The spam policy you can read; that other thing you have to live through to understand.