THE SHORT ANSWER
AI-generated content ranks, and search guidance has been explicit for several years that the target is content produced to manipulate rankings rather than content produced with a particular tool. Detection classifiers are too unreliable to enforce provenance at scale. The genuine risk is different: generated drafts converge on the same structure, the same claims and the same phrasing as everyone else's, which produces pages with nothing distinctive to retrieve or cite.
The question keeps being asked because the honest answer is unsatisfying. There is no penalty for using a tool, there is no reliable detector, and pages assembled with heavy machine assistance rank perfectly well every day across every competitive vertical. People asking hope for a rule that separates the virtuous from the lazy, and no such rule exists at the level of production method.
What does exist is a quality threshold that generated content fails in a characteristic way. Models are trained on the existing web, so unguided output reproduces the consensus of that web. The result is a page that is accurate, readable and completely interchangeable with forty others, which is a commercial problem long before it is a ranking one.
The numbers, at a glance
Official position: search guidance evaluates whether content is helpful and original, not whether a human or a machine typed it
Detection reliability: classifiers produce enough false positives on ordinary human writing that no major engine enforces provenance from them
The real failure mode: convergence, where unguided output restates the consensus of the training data and adds nothing retrievable
What survives: content carrying original data, first-hand experience or a specific position, regardless of how the sentences were produced
What the guidelines actually say, and what they leave open
Search guidance since 2023 has framed the test around purpose and quality. Content created primarily to game rankings is the target, whoever or whatever produced it. Content that helps a reader is fine, whoever or whatever produced it. That formulation deliberately avoids provenance, because provenance is unenforceable.
What the guidance leaves open is the boundary, and the boundary is where the interesting decisions live. Mass-producing pages that restate publicly available information sits on the wrong side even when every sentence is technically true, because the reader gains nothing they could not get anywhere. That is a judgement about value, not about tooling, and it has been enforced against human-written content for two decades.
Why detection cannot be the mechanism
Detection classifiers work by scoring statistical properties of the text, principally how predictable each word is given the preceding ones. That signal degrades badly on two categories of writing: text that has been edited, and text by writers who naturally write plainly. Non-native English writers and technical documentation both trip detectors constantly.
An engine deploying such a classifier as a ranking input would demote a large volume of legitimate content and would be gamed within weeks by trivially adding unpredictability. Every public statement from the major engines is consistent with this reasoning, and the observable evidence agrees: heavily assisted content ranks across every vertical without visible provenance-based suppression.
The convergence problem, which is the one that actually costs money
Ask any capable model to write about heat pump sizing and you will get a competent piece covering the same eight points, in a similar order, with similar caveats, as the piece your competitor generated the same afternoon. Neither page is wrong. Both are unnecessary, because the information is already in the model and the reader can get it without either page.
For retrieval this is fatal in a specific way. When six pages say the same thing in the same words, the retrieval system picks between them on authority and freshness alone, which means the biggest existing brand wins by default. Your only route past that is to say something the model could not have generated, which by definition means information from outside the training data.
Your own operational data. Conversion rates, job values, response times, failure rates, whatever you measure and nobody publishes.
First-hand observation. What actually happens on site, in the specific conditions of your market, including the things that go wrong.
A stated position. A defensible opinion about a contested trade-off, with the reasoning, which consensus-trained output systematically avoids.
Local specificity. Regulations, grant regimes, building stock and pricing in a market too small to be well represented in training data.
A workable production standard
The teams getting good results treat generation as drafting and structuring rather than as authorship. The machine produces the scaffolding, the ordering, the transitions and the boring connective prose. The human supplies every number, every example, every judgement and every claim that could be wrong, and then removes the paragraphs that restate common knowledge.
One practical rule keeps it honest: no page ships unless it contains at least one thing that would be false, unverifiable or simply absent if written by someone outside your business. If you cannot point to that sentence, you have produced a page that competes on nothing.
Publishing assisted content that holds up
Insist every page contains at least one fact sourced from inside your business.
Delete any paragraph that restates information a reader could get from the assistant directly.
Have a named person verify every number, date and regulatory claim before publication.
Vary structure deliberately across a set, since identical scaffolding across pages is more detectable than word choice.
Track which pages earn citations and feed that back into what kind of original material you gather next.
Want leads like this in your pipeline?
Flock runs the campaigns, screens the enquiries and hands you only the ones that match your service area, job size and capacity. You pay per lead, not per month.
Book a 15-minute fit check | See lead package pricing
Related answers
Frequently asked questions
Will I be penalised for using AI to write content?
Not for the tool. Penalties and filtering target content with no value to a reader, which assisted content falls into easily when unedited, and which human content has always fallen into as well. The distinguishing factor is whether the page contains something worth reading, not who typed it.
Should I disclose that content was AI-assisted?
There is no search requirement to do so. Some sectors have regulatory or professional expectations, and some audiences care. Treat it as an editorial and trust decision for your market rather than a ranking consideration, because it has no measurable effect on visibility either way.
Do detection tools work well enough to worry about?
No. False-positive rates on edited text and on plain-spoken human writing are high enough that the outputs are not decision-grade, and the scores move sharply when a document is lightly rephrased. Several universities and publishers that adopted such tools for enforcement have since withdrawn them for exactly that reason, which tells you what to make of vendor accuracy claims.
Why does my generated content not rank then?
Almost always because it says what forty other pages already say. When passages are interchangeable, the engine falls back on authority, and a newer or smaller site loses that comparison every time. Add something only you know and the calculation changes.
NEED A CLEARER PLAN?
Let’s turn your next move into momentum.
Talk to us →