Every article about the Google spam update and AI content ends up in the same place: Google does not penalise AI content, it penalises scaled content abuse. That is true. It is also useless the moment you ask the only question that matters for your site: where exactly is the line, and have we crossed it?

So here are seven publishing behaviours that move AI-assisted content from "fine" to "flagged". Each one comes with a check you can run this afternoon and a fix. None of them is about whether a machine typed the first draft.

What the spam updates actually target

In March 2024, Google rewrote its spam policies and retired the old "automatically generated content" wording. In its place came three named behaviours: scaled content abuse, site reputation abuse and expired domain abuse. You can read them yourself on the Google Search spam policies page. AI is not on that list.

The rewrite was deliberate. The old policy asked how a page was made. The new one asks why it exists: was it produced at scale, primarily to manipulate rankings, without regard for whether it helps anyone. A hand-typed page written to that recipe is a violation. An AI-assisted page written to genuinely answer a query is not. Google said much the same in its February 2023 guidance on AI-generated content, and nothing since has walked it back.

The other thing worth understanding: SpamBrain, Google's spam-detection system, and the update machinery around it operate on patterns across a site, not on the word choices in one paragraph. Nobody is scanning your intro for the word "landscape". They are looking at what forty URLs published in five weeks have in common.

Which means the useful question is not "was this AI?" It is "does our publishing pattern look like a team or a script?"

7 publishing behaviours that turn AI-assisted content into scaled content abuse

1. No research layer behind the briefs

The most common failure happens before anyone writes a word. Someone exports a keyword list, sorts by volume, and turns each row into a brief. Nobody opens Google and reads what is actually ranking for those queries.

The result is a page that answers a question in a form nobody asked it. "Commercial cleaning cost" pulls buying guides with price ranges and contract structures. If your brief produced a general article about the benefits of a clean office, you have not competed for that query, you have added a page to the pile.

Self-check: take five recent briefs and open the top ten results for each target query. Does your article match the dominant format and depth?

Fix: plan every brief against the live SERP, not the spreadsheet. Format, angle and depth are decided by what is already winning.

2. Template-identical articles

This is the big one, and almost nobody writes about it.

A hit is rarely caused by one bad page. It is caused by forty pages built from the same skeleton: 1,200 words, an intro of the same shape, five H2s, three bullets under each, a conclusion that restates the intro. Each piece might read fine on its own. Stacked together, the structure itself is the signature. A human editorial team does not produce forty structurally identical documents, because human editors keep deciding that this topic needs a table, that one needs 700 words, and the third needs a worked example and nothing else.

Self-check: open your last ten posts side by side and count H2s and word counts. If the numbers barely move, you have a template, not a process.

Fix: vary section count, format and depth per brief. Let some articles be short. Let some be lists and others be walkthroughs.

3. Near-duplicate titles competing for one intent

Cannibalisation and scaled content abuse share a root cause: briefs built from keyword variants rather than from intent. If two briefs answer the same question, you have not made two articles. You have made one article twice.

A pattern we see constantly: "seo content writing service", "seo content writing services uk" and "best seo content writing service" published as three separate posts. All three are the same buyer asking the same thing. Google picks one, ranks it weakly, and the other two dilute the signal.

Self-check: in Search Console, look for two of your URLs trading places on a single query over a few weeks. That swap is the tell.

Fix: the keyword cannibalisation fix is almost always consolidation. Build one page that handles the variants inside it, 301 the others to it, and update internal links. The query usually recovers once there is a single obvious answer.

4. A volume spike with no editorial capacity behind it

Cadence is a signal, and it is the one people forget to manage. A site that published twice a month for two years and then ships sixty pieces in five weeks looks automated whatever the quality, because no small business hired a newsroom in that window.

Here is the risk profile in plain numbers: an SME site with 18 existing pages adds 60 blog posts in five weeks, every one of them around 1,200 words, every one following the same intro-plus-five-H2s pattern, with four of them targeting variations of one query. Nothing in that description mentions AI. It does not need to.

Self-check: plot published URLs by month. A gradual ramp looks like a team growing. A vertical line looks like a script.

Fix: ramp. Two posts a month becomes four, then six. You also get the benefit of seeing which pieces earn impressions before committing to more of the same.

5. Zero first-hand specificity

Ask a colleague outside marketing to read five of your posts and point to one sentence that could only have come from your business. If they cannot, that is your answer.

Generic content is not a stylistic problem. It is the substance problem that scaled abuse enforcement is built to catch: pages that add nothing to what already exists on the web. No prices, no named standards, no real scenario, no opinion anyone would argue with.

Fix: every section needs one thing only you would write. A real price band from a recent quote. A named regulation, like the requirement for an EPC before marketing a property. The mistake your customers actually make in week one. Specifics are also the fastest way to make AI content sound human, because generic prose is the tell, not the tooling.

6. Unattended auto-publishing

Plenty of tools will generate, format and publish straight to your CMS without a human reading anything. That is where the wheels come off, because the failure mode of language models is confident wrongness: an invented statistic, a service you do not offer, a competitor's process described as yours.

Fix: a sceptical sub-editor read before anything goes live. Someone whose job is to ask "is that number real, and where did it come from?" We build that pass into how we write every article, because it is the difference between assistance and abandonment.

7. Topic drift into subjects you have no standing on

A plumbing firm publishing about mortgage rates because the volume looked good is not building topical authority. It is building a reason for Google to doubt the rest of the site. This is also where scaled content abuse starts to overlap with site reputation abuse: content hosted on a domain that has nothing to do with it, published mainly for the ranking.

Self-check: for each planned topic, can you name the customer conversation it came from? If not, cut it.

How AI content gives itself away before Google gets involved

Most AI writing has a tell, and readers spot it before any algorithm does. Three tells dominate.

Vague filler that could sit on any site in any sector. Sentence rhythm where every sentence lands at roughly the same length, paragraph after paragraph, with no asides and no interruptions. And recycled connective phrasing, the same handful of transitions doing all the work between ideas.

Banned-phrase lists catch the obvious version of this. On their own they are a blunt instrument, because they treat vocabulary as the problem when structure is the problem. You can strip every offending word out of a draft and still be left with something that reads like nobody in particular wrote it.

What actually works is duller and harder. Swap an abstract claim for a number with a source attached. Replace "improves efficiency" with a named tool and what it changed. Break the rhythm deliberately: a three-word sentence after a long one. Answer the question on the first screen instead of warming up for 200 words, because a reader who has to scroll past a preamble to find out whether you answered them has already decided.

A self-audit you can run this week

• Pull your last 30 published URLs into a sheet with title, word count, H2 count and target query. Sort by target query and look for overlap.

• Plot publishing volume by month for the past two years and mark any step change.

• Sample five posts and run the outside-colleague test for first-hand detail.

• In Search Console, filter for pages with sustained impressions and effectively no clicks. Those pages are being seen and rejected, which is a content problem, not a ranking one.

• Check for two URLs swapping positions on one query, then plan the consolidation.

If you would rather have that read done for you, our free content marketing snapshot covers the same ground against your live site.

If you have already been hit

Separate the diagnoses first, because the fixes differ. A spam update hit tends to be sharp and site-wide, and it lines up with a dated spam update announcement. A core update hit is usually a re-rating of quality against competitors, often uneven across sections. Confirming which one you are looking at means matching your traffic drop date against Google's published ranking updates list, not guessing.

Then triage by intent group rather than by URL. Consolidate the pages that cannibalise each other. Rewrite the ones with a real reason to exist, adding the substance they never had. Remove the pages that only ever existed to catch a keyword, and be honest with yourself about which those are.

On timing: expect months, not weeks. Google's guidance on core updates is explicit that improvements may not be reflected until a later update runs, and the same patience applies after a spam update cleanup. Set that expectation before the work starts, or someone panics in week three.

What not to do: mass noindexing the blog, deleting pages that earn links, or replacing thin AI content with thin human content. The last one is the most common waste of budget we see. The method was never the problem.

Building a pipeline that stays on the right side of the line

Research first. We read your website like a new hire would, crawl it politely, and plan each brief against the live Google top ten with UK search volume and difficulty, rather than against a stale export. That step alone kills most cannibalisation before a word is written, because a brief that duplicates an existing page gets rejected at planning.

Voice before volume. We build a brand voice profile from your own writing, and when a site is thin there is not much to learn from, we say so rather than inventing a personality. A vague adjective like "professional" is not a voice.

Then two grades on every draft: an SEO score for whether the page can compete, and a content integrity score for whether it deserves to. Integrity wins the tie. A page that scores well on keywords and badly on substance is precisely the profile spam updates exist to find, so we publish the lower-SEO draft with the real substance in it. You can see the mechanics on our features page, and the schedule and cadence options sit on plans and pricing.

Nothing about a Google spam update and AI content should frighten a business that publishes on a sane cadence, checks its archive before commissioning, and puts a sceptical human between the draft and the CMS. The sites that get caught are not caught for using AI. They are caught for publishing sixty near-identical pages nobody read.

Frequently Asked Questions

Does the Google spam update penalise AI content specifically?

No. Google's spam policies name scaled content abuse, site reputation abuse and expired domain abuse, and AI-generated content is not listed as a violation in itself. Google's published guidance from February 2023 says appropriate use of AI is not against its guidelines. What gets penalised is content produced at scale primarily to rank, regardless of who or what produced it.

What counts as scaled content abuse?

Google defines it as generating many pages primarily to manipulate search rankings rather than to help users, whether they are produced by automation, by people, or by a combination of both. The scale, the intent and the lack of value all matter together. A handful of genuinely useful AI-assisted articles does not meet that description.

How can I tell if my site was hit by a spam update or a core update?

Match the date of your traffic drop against Google's published list of ranking updates. Spam update hits tend to be abrupt and broad across a site, while core update movement is usually a re-rating that hits some sections harder than others. If the drop aligns with a named spam update and your archive has any of the seven patterns above, treat it as a spam issue.

How many blog posts a month is safe to publish?

There is no published number, because volume alone is not the signal. What matters is whether your cadence matches your editorial capacity and whether it changed suddenly. Going from two posts a month to sixty is a risk; ramping from two to four to eight over a few months is how a growing team actually behaves.

How long does it take to recover after a spam update hit?

Realistically, months rather than weeks. Google's guidance notes that improvements often are not reflected until a subsequent update runs, so the cleanup and the recovery are separated by an update cycle you do not control. Do the consolidation and rewriting properly, then leave the site alone rather than making further changes every fortnight.

Related reading

• Is AI Content Penalised by Google? What Gets Flagged

• Site Reputation Abuse: How to Audit Your Own Site

• Every article graded twice before it publishes