Your SEO agency is very likely using AI somewhere in the process of writing your blog posts, and that on its own tells you nothing about whether the work is any good.
The question that matters is what a person added after the tool had finished.
This article covers what Google has said, the tells you can check in ten minutes, the risk that actually costs you rankings, and the one question to put to your agency.
We should say up front that Growthcart uses AI tools when producing content for clients, for research, outlines, first drafts and checks. So we have no interest in arguing that agencies which use AI are cheating you.
What follows is about how to tell the difference between an agency that uses the tool and an agency that lets the tool do the job.
What Google actually says about AI-generated content
Google's position since February 2023 has been that it judges content on quality, and does not care how it was produced. The Search Central post by Danny Sullivan and Chris Nelson put it in one line that is still the clearest statement on the subject.
"Using AI doesn't give content any special gains. It's just content." Danny Sullivan and Chris Nelson, Google Search Central, February 2023
The same post says that "appropriate use of AI or automation is not against our guidelines" and that automation "with the primary purpose of manipulating ranking in search results" is a spam policy violation.
It also answers the disclosure question. AI disclosures "are useful for content where someone might think 'How was this created?'", and giving AI an author byline "is probably not the best way" to be clear with readers.
In March 2024 Google replaced its old rule on automatically generated content with a policy called scaled content abuse. The wording is deliberate. The policy covers "large amounts of unoriginal content that provides little to no value to users, no matter how it's created", whether "through automation, human efforts, or some combination".
A content farm staffed by people breaks the same rule as one run on ChatGPT.
The current spam policies page gives "using generative AI tools or other similar tools to generate many pages without adding value" as the first example. The trigger is the scale and the absence of value. The policy also now covers attempts to manipulate generative AI responses in Google Search, which matters if your agency is selling you AI Overviews visibility. On that point, keyword stuffing actually makes AI citation worse.
The January 2025 update to the Search Quality Rater Guidelines adds detail on how Google's human raters are told to grade such pages. Section 4.6.6, reported by Search Engine Land, says the Lowest rating applies if "all or almost all" of the main content is "copied, paraphrased, embedded, auto or AI generated".
The condition is that it was made "with little to no effort, little to no originality, and little to no added value for visitors". Raters do not set rankings directly. Their guidelines describe what Google's systems are being tuned to reward.
So the policy answer is settled. AI-assisted content is fine. Unoriginal content at volume is spam, by hand or by machine.
What the ranking data shows
Two large Ahrefs studies, a year apart, found that AI use on a page has almost no relationship with where it ranks. The first, published in July 2025, ran 600,000 top-20 URLs for 100,000 keywords through Ahrefs' own detector. The correlation between AI share and position was 0.011, which the authors read as no meaningful relationship.
pie showData title Top-20 Google pages by AI share (Ahrefs, July 2025, 600,000 URLs)
"Mixed human and AI" : 81.9
"Purely human" : 13.5
"Entirely AI" : 4.6
The second study, published by Ryan Law in July 2026, looked at the top ten for 100,000 June 2026 SERPs. It found 5.3% of top-three pages were 100% AI-generated, so fully machine-written pages can and do rank first.
But the average AI share crept up from 27.1% at position one to 30.9% at position ten. Pages with under 20% AI content were indexed at a rate of 49.28%, against 40.35% for pages at 80% or more. Low and moderate AI pages drew roughly two to three times the impressions of heavy AI pages.
Law's reading was that "increasing use of AI correlates with decreasing content quality" and that Google "is not against AI content; it is against bad content". Both studies depend on a detector that is not perfect, which Ahrefs says itself. The direction is consistent, though: the tool does not cost you rankings, and the generic output that comes from leaning on it too hard does.
The tells you can spot in your own blog
Wikipedia's editors maintain a field guide to the habits of chatbot prose, and it is the best public reference for a client who wants to check a draft without a detector. The page, Signs of AI writing, is built from real examples found in Wikipedia articles.
It is explicit that it is "descriptive, not prescriptive". No single sign proves anything. A cluster of them in one post is a strong signal that nobody edited it.
Open your three most recent agency-written posts and count how many of these appear:
- Sentences that end in a tacked-on participle: "highlighting", "underscoring", "ensuring", "fostering", "showcasing". The fact comes first, then a clause telling you why it matters.
- Inflated significance. A new service page described as "a pivotal moment", a product "a testament to" something, a change that "reflects broader trends".
- Lists in threes everywhere: three adjectives, three bullet points, three short phrases, in places a person would have used one or four.
- Negative parallelism: "not just X, but Y", "it's not about X, it's about Y". The sentence corrects a misconception no reader held.
- Em dashes used in pairs, several per paragraph, where a comma or full stop would do.
- Bullet lists where every item starts with a bold label and a colon.
- Vague authority: "experts agree", "studies show", "industry reports suggest", with no named expert, study or report.
- Vocabulary that was measured as overused after 2022: delve, tapestry, landscape, robust, seamless, crucial, leverage, navigate, unlock.
- Headings in the "X and Y" shape (History and Background, Challenges and Opportunities) and a "Conclusion" that repeats the post.
- Chatbot leakage: a line that starts with "Certainly", ends with "hope this helps", or apologises that information on the topic is limited.
- Statistics with no link, or links that go to a homepage, a dead page, or a page that does not contain the figure.
One or two of these appear in plenty of human writing, which is why the Wikipedia guide warns against treating them as proof. Seven or eight in a 900-word post is a different matter. That post was generated, pasted and published.
The guide says two things about detection that matter here. It tells editors not to rely on detection software, because tools such as GPTZero "have non-trivial error rates". It also tells them not to trust their gut. A 2025 study found human ability to distinguish LLM text from human text was "no better than random chance", and a second study on German theses found a recognition rate of 57% for AI text.
If your agency sends you a "0% AI" screenshot from a detector as proof of human authorship, that proves nothing in either direction. Read the post instead.
The risk is commodity content, not the tool
The actual damage from AI-written agency content is that it is identical to what every competitor could publish, and content like that has nothing a search engine or an AI assistant would choose to cite. It is the content equivalent of the SEO services that are theatre.
Google's guide to generative AI features, updated in July 2026, has a section headed "Create valuable, non-commodity content for your audience". Its advice: "Create the content yourself based on what you know about the topic." It describes common-knowledge pieces as adding "little unique insight for readers".
Google's own self-assessment questions for helpful content make the same point from the other side. Does the content provide original information, reporting, research or analysis? Does it provide substantial value when compared to other pages in search results? Is it primarily made to attract visits from search engines?
A model prompted with "write 800 words on the benefits of boiler servicing" fails all three, and would fail them if a junior copywriter had been given the same brief. The stakes are higher for a law firm or any regulated practice, where a generic AI page can be commodity content and quietly wrong on the law at the same time.
This is why the tool is the wrong thing to argue about. A plumber's agency could use AI to turn a 20-minute phone call with the plumber into a post about the boiler faults they see most in Northamptonshire houses built in the 1970s. That post has experience in it, which is the first E in Google's E-E-A-T.
The same agency could skip the phone call and publish the model's generic version. Same tool, opposite outcome, and only one of them is a candidate for an AI Overview or a Perplexity citation.
Agencies know this. A Duda survey of 36 of its agency customers, reported by IT Brief UK in February 2026, is a small sample but a telling one: 64% named "AI slop" as their top concern, and 50% worried that clients would misunderstand the role AI plays. The agencies themselves see the problem. The question is whether yours has done anything about it.
Fine uses and not-fine uses
The table sorts the common agency uses of AI by whether they add or remove the thing Google says it rewards.
| Use | Fine or not | Why |
|---|---|---|
| Keyword clustering and search intent grouping | Fine | Analysis, no reader-facing text produced |
| Drafting an outline from a client interview or brief | Fine | A person supplied the substance |
| First draft from a detailed brief, then edited and fact-checked by a person | Fine | Google: appropriate use is not against the guidelines |
| Rewriting the client's own notes, case studies or call transcripts into prose | Fine | Experience goes in, prose comes out |
| Title tag and meta description variants for a person to pick from | Fine | Low-stakes, easy to check |
| Checking a draft for errors, gaps and readability | Fine | The model is a reviewer, not the author |
| Publishing an unedited draft under a staff byline | Not fine | Google's rater guidelines rate unedited AI main content Lowest |
| Inventing statistics, quotes or case studies because the model produced them | Not fine | Fabrication, whatever produced it |
| Fifty location pages from one template with the town name swapped | Not fine | Textbook scaled content abuse |
| "Four posts a month" with no interview, no source list and no edit history | Not fine | Commodity content made for search engines, which Google's self-check flags |
| A detector screenshot offered as proof of human writing | Not fine | Detectors have real error rates; the proof is in the text |
The pattern in the right-hand column is that every fine use has a human input before the tool and a human check after it. Every not-fine use has the tool at both ends.
The question to ask your agency
Ask your account manager this: for the last post you published for us, what did a person add that a search engine could not have produced on its own?
A good answer names things. The writer spoke to your operations manager for 20 minutes. The post uses a figure from your own booking system. It links to the ONS table and the HSE guidance page the writer read.
It was edited by a named person on a named date, and the edit changed the structure, not just the commas. The agency can show you the brief, the source list and the draft history if you ask.
A bad answer is about process in general. "We have a rigorous human review stage." "All content is checked by our editorial team." "We use AI responsibly." None of these can be verified, and all of them are what you would say if the answer were "nothing".
Three follow-up questions sort the rest. Who wrote this and what did they know about our business before they started? Which of these facts came from us? What would you have to cut from this post if you removed everything a competitor could have published word for word?
If the honest answer to the last one is "most of it", you are paying a retainer for commodity content. The Ahrefs indexation and impression figures above tell you where that ends up: often in a future content pruning exercise.
If you want a second opinion on content you have already paid for, our SEO audit checks published posts against the tells above and against Google's self-assessment questions. Our content marketing work starts with an interview, not a prompt.
A simple decision rule: judge the post, not the tool. If you can point to the paragraph that only your business could have written, the agency is doing its job, whatever software sat between the interview and the page. If you cannot find that paragraph, it does not matter whether a person or a model wrote the rest.
Whether an agency uses AI well is one test among several; our guide to what a UK SEO agency should cost and deliver covers the others. If you would like us to look at what your current agency has been publishing, get in touch and send us three recent posts.