Is AI Content Bad for SEO? What Google Actually Says in 2026

The question gets asked in the wrong shape. “Is AI content bad for SEO” assumes Google has a view about who typed the words, and it does not. It has a view about whether the page is worth showing to somebody.

That distinction is not a technicality. It explains why one business publishes forty AI-assisted articles and grows, and another publishes forty and loses half its traffic. Same tool, opposite outcomes, and the difference is entirely in what was done with it.

What Google’s policies actually say

Google’s spam policies are written to be method-agnostic, and they have been for some time. The relevant policy is scaled content abuse, and it describes the practice rather than the tool: creating many pages primarily to manipulate rankings rather than to help users. The documentation says explicitly that the method does not matter, and that generative AI tools can be part of the problem.

Note the shape of that sentence. Not “AI content is spam”. “AI can be part of the problem”, where the problem is defined independently.

The August 2026 spam update, which ran from 18 to 21 August across all languages and locations, is a useful test of this. It did not introduce new AI rules. It did not announce new spam-policy categories. Google improved enforcement of policies that already existed. Sites that lost visibility lost it against rules that had been on the books beforehand.

What people ask

Did a human or a model write this? They then look for detection tools, disclosure rules and safe percentages, none of which map to anything in Google's documentation.

What Google asks

Does this page exist to help someone, or to occupy a search result? Is anyone accountable for whether it is correct? Would a reasonable person feel their time was well spent?

What actually gets punished

In practice, four patterns cause the damage. All four are possible without AI and all four became much cheaper with it, which is why they got common.

Scale without judgement. One template turned into three hundred barely differentiated pages. “Technical SEO in Reading”, “Technical SEO in Slough”, “Technical SEO in Maidenhead”, identical but for the place name. This was a bad idea in 2015 and it is a bad idea now. AI simply made it possible for one person to do it in an afternoon.

Paraphrase of what already ranks. Reading the top ten results and producing a synthesis of them adds nothing to the web. It also produces a page with no independent claim to be right, which matters more now that AI systems select sources by corroboration rather than by position.

Unreviewed claims. Publishing figures, dates, legal positions or medical guidance that no qualified person has checked. This is where AI does genuine damage, because a confident wrong number reads exactly like a confident right one. It is also where the reputational cost lands long before the ranking cost does.

Volume as a strategy. Publishing two hundred pages because you can, rather than twenty because they were needed. The older ranking systems could be flooded. A synthesis system that picks three sources for an answer cannot.

Notice that none of those four is “used a language model”.

What does not get punished

Research and structuringGathering material, mapping an outline, finding the questions a topic needs to answer. Nothing reaches the page unfiltered.

First drafts a person rewritesThe draft is scaffolding. What publishes carries the writer's judgement, their examples and their willingness to be wrong in public.

Editing and tighteningCutting, clarifying, checking readability, catching the sentence that says nothing. This one is unambiguously an improvement.

Genuine production supportTranscribing interviews, turning a data set into prose, drafting alt text and summaries. Real work, done faster, with a person accountable for it.

There is no disclosure requirement in Google’s guidance for any of the above. There is no acceptable percentage. There is no threshold at which a page becomes ineligible. Anyone selling you a compliance number has invented it.

The two things that actually determine outcomes

Strip away the anxiety and it comes down to two questions.

Does the page contain something that is not already on the web?

An original number. A first-hand account. A worked example from real client work. A position somebody is prepared to defend. A methodology explained rather than asserted.

A model cannot supply any of these, because it can only recombine what it has already read. This is the entire reason AI-assisted content splits so cleanly into winners and losers. Teams that bring something and use AI to express it faster do well. Teams that expect the model to bring the substance publish an eloquent nothing.

The uncomfortable test: what is on this page that a competent person could not have produced by reading the current top ten results? If the answer is nothing, the page has no reason to outrank them.

Is someone accountable for it being right?

Google’s quality guidance keeps circling experience, expertise, authoritativeness and trust, and trust is the one that carries the weight. Trust is not a writing style. It is whether there is a named person whose reputation is attached to the accuracy of the claims, and whether the claims survive checking.

That is why unreviewed AI output is genuinely dangerous and reviewed AI output genuinely is not. The review is the trust.

Why this matters more in AI search than in classic SEO

There is a second reason to care, and it is becoming the bigger one.

Search results have ten positions and can afford to include a mediocre page at seven. An AI-generated answer names two or three sources. The selection is far harsher, and it favours sources that state something specific enough to be worth attributing.

Recycled synthesis fails that test completely. It has nothing that could be quoted and attributed, because everything in it came from somewhere else. The page can rank tolerably and still never once be cited, which is how a site ends up with impressions that do not convert into anything at all.

Directional weighting from visibility work with clients. The top two are things a model cannot supply for you.

• Original dataHighest
• First-hand experienceVery high
• Named accountabilityHigh
• Self-contained answersMedium
• Word countNegligible

A workable policy for using AI in content

If you want a rule your team can follow, this is the one we use and recommend.

Use AI for anything where a person still supplies the judgement. Research, structure, first drafts, editing, summaries, alt text, transcription. All fine, no disclosure needed, no anxiety warranted.

Never publish an unchecked factual claim. Every figure, date, name, legal position and citation gets verified against a named primary source by a person, and that person is on the record internally. This single rule prevents most of the real damage.

Require one original element per page. Data, an example from your own work, a stated position, a methodology. If a page cannot clear that bar, do not publish it. This rule alone kills scaled content abuse before it starts, because you cannot produce three hundred original elements in an afternoon.

Publish fewer things, deliberately. Twenty pages that answer something completely will outperform two hundred that circle it, in classic search and dramatically so in AI answers.

Name an author and mean it. A real person, with a real biography, who would defend the content if challenged. Not a house byline attached to work nobody reviewed.

The honest answer

AI content is not bad for SEO. Content that exists to occupy a search result rather than to help somebody is bad for SEO, and it always has been. AI made producing that kind of content roughly free, which is why there is so much more of it and why enforcement has tightened around it.

The teams doing well with AI are not the ones that found a safe percentage. They are the ones that kept doing the expensive part, which is having something worth saying and being accountable for whether it is true, and used AI to remove the friction around it.

If your organic traffic has moved and you are not sure whether content quality, technical issues or something else is behind it, that diagnosis is where our technical SEO work starts. If the concern is whether you are showing up in AI answers at all, AI search visibility begins with a measured baseline rather than an opinion.