Google’s leaked search documentation includes a field called contentEffort, described as an AI estimate of effort for article pages. It does not establish that Google identifies lightly edited AI articles and automatically demotes them, or reveal how much the field matters to rankings.
The entry comes from the documentation exposed in 2024. Claims that it confirms a content-effort ranking factor go beyond the available evidence. Publishers do face a documented risk from producing large amounts of unoriginal material to manipulate search results, but Google’s scaled-content-abuse policy applies regardless of how that material is made.
What the leaked field records
The preserved QualityNsrPQData documentation describes contentEffort as:
“LLM-based effort estimation for article pages”
The description specifies article pages and says nothing about detecting AI authorship. It also points to an internal reference that the public entry does not explain.
The field holds a list of versioned numeric signals. The associated data structure contains a numerical value and a version identifier. That describes how a value can be stored; it supplies neither a score for a particular publisher nor a scale that an SEO tool could reproduce.
In his May 2024 examination of the leak, Mike King distinguished documented attributes from confirmed ranking factors. He noted that the material did not disclose scoring functions or establish whether every available feature was being used. Those omissions prevent a reader from calculating the ranking effect of an individual field.
Even an earlier discussion of contentEffort by Cyrus Shepard included the qualification that it was unknown how, or even whether, Google used the score. The stronger claim needs evidence beyond the field’s existence.
Google’s guidelines allow high-effort AI work
Google’s Search Quality Rater Guidelines address the relationship between AI and effort directly. Section 4.6.6 says:
“the use of Generative AI tools alone does not determine the level of effort or Page Quality rating.”
The same passage says generative AI can be used for both high-quality and low-quality content, including original artwork involving substantial effort. Elsewhere, the guidelines recognize work spent building useful page functionality, such as a machine-translation service.
They also instruct raters to assign the lowest rating when almost all of a page’s main content is republished, paraphrased or generated with virtually no effort, originality or additional value for visitors. Crediting the source does not, by itself, fix that deficiency.
These are instructions for human evaluators. Google says their ratings do not directly influence rankings. They explain what Google wants evaluators to recognize, but they do not disclose the implementation of contentEffort or establish that an LLM follows the same rubric.
The policy risk is scaled content abuse
Google defines scaled content abuse as generating many pages primarily to manipulate rankings rather than help users. Its examples include using generative AI without adding value, transforming scraped material, and combining other pages’ content without a useful contribution. The policy explicitly covers unoriginal, low-value output however it is created.
It does not set a weekly article allowance. Publishing dozens of articles is therefore insufficient, on its own, to establish a violation. Equally, having someone lightly edit every draft does not resolve a process whose output still meets the policy’s definition.
SEW’s guide to black hat SEO techniques puts scaled content abuse first because a flawed production process can spread the same problem across a large body of work.
Publishers can inspect the contribution, even without the score
The uncertainty around contentEffort leaves publishers without a defensible optimization formula. A more useful editorial decision is where to spend the next hour on an article. Google’s guidance on explaining how content was created gives product reviews a concrete example: show what was tested, the results, and evidence of the process.
For a software review, that could mean running a failed task again, recording the settings and showing the output. A reader can then distinguish a reproducible limitation from a vague complaint. Adding an illustration of someone using a laptop would provide no equivalent evidence about the product.
AI could help organize those test notes or edit the explanation. The contribution would remain the observed behavior and the evidence supporting it. The publisher could document that improvement, but would still have no measurement of whether Google’s contentEffort score changed.
Start the conversation by posting the first comment