Content Effort and AI Content: Origin Is Not the Rating

The worst content review I have ever sat in opened with a compliment. It was the same category roundup from the rest of this cluster. Our head of content pulled up the 2,400-word draft, told me it read better than anything we had shipped that quarter, and then asked why it was sitting at position fourteen. So I read it on the shared screen while five people waited. Clean subheads. Sensible flow. Not a typo in sight. And somewhere around the third scroll I realized I could not find one sentence on that page that existed anywhere outside the model that produced it. The interviews had happened. None of them were on the URL. No screenshot, no opinion anybody would bother arguing with. The only genuinely unique sentence on the page was the disclosure at the bottom, which said the draft had been created with AI assistance and reviewed by a human.

That disclosure was true. A human had reviewed it. She spent the better part of two hours on it and she is good at her job, and nothing about those two hours made it onto the page.

I want to tell you what I got wrong for the year and a half after that meeting, because you may be carrying the same assumption. I thought AI origin was the rating. I thought the page was flat because a model had written it, and that the grown-up response was a ban on generators. Then I read the sentence in the 2025 rater handbook that almost every team skips on its way to the spam sections.

So here is the answer up front: yes, an AI-assisted page can earn high Content Effort. When Search Engine Land recapped the January 2025 rater update, the line worth keeping was that the handbook issued to more than 10,000 quality raters says the use of generative AI tools alone does not determine the level of effort or the Page Quality rating. A few paragraphs later, the same Search Quality Evaluator Guidelines send a page to Lowest when all or almost all of the Main Content is copied, paraphrased, or AI-generated with little to no effort, originality, and added value. Both sentences are live at once, and you have to hold both. A model in your stack is not a verdict, and a fluent paraphrase is not a pass.

TL;DR

• Banning the generator does not raise Content Effort. Requiring artifacts does.
• Low-effort AI content is usually SERP restatement plus a tone pass. High-effort AI-assisted content is a model used inside a process that leaves first-party residue on the page.
• Scaled AI publishing is a site-level diagnosis. One assisted URL scoring badly is a page-level diagnosis. Different problems, different fixes.
• Cadence and AI tells are a separate job, and I keep them separate on purpose: see how to humanize AI writing.

The question leadership actually asks, usually in a Slack thread at 11pm, is "should we ban AI?" Banning the tool does not raise effort. Requiring artifacts does. A team that bans generators and keeps writing from the top ten results has changed its tooling story and nothing that a rater can actually see.

Can AI-generated content have high Content Effort?

Yes, when the finished Main Content still shows work a model could not cheaply invent. No, when the model is the researcher and the writer and the search results are the only corpus it ever saw. That distinction is the whole page, and most AI-content arguments skip it on the way to fighting about detectors.

By Main Content the handbook means the part of the page that does the job you came for: the article body, the calculator, the video, not the nav or the ads. Content Effort is the criterion raters use to judge how much genuine work went into that part, and it is assessed entirely from the outside. It looks at the page, not your process.

The handbook carries examples running in both directions, which is why quoting only one of them is how bad policy gets written. It allows that a high level of effort can go into creating original content with generative tools. It also describes automatically producing thousands of pages by running existing material through software with no human oversight, which it treats as the opposite case. That second example is older than chatbots, incidentally: machine-translated pages republished at volume were already in the effort discussion years before anyone put a language model on a roadmap, which tells you the criterion was never really about the tool.

So the brief I give a writer now is not a moral rule about software. It is a requirement about evidence: if the page cannot show a measurement, a document, a dated screenshot, or a decision you would defend to a customer, polishing the prose will not move Content Effort. For the criterion itself, start with what is Content Effort. For the scoring mechanics, go to the Content Effort audit checklist.

What the 2025 rater guidelines actually say about AI

Three statements from the handbook do the real work, and I would pin them somewhere your team trips over them weekly, because everyone remembers the one that suits their existing position and forgets the rest.

1. The use of generative AI tools alone does not determine the level of effort or the Page Quality rating.
2. Lowest applies when all or almost all of the Main Content is copied, paraphrased, embedded, auto-generated, AI-generated, or reposted with little to no effort, originality, and added value.
3. Scaled content abuse covers using automated tools, generative AI or otherwise, as a low-effort way to produce many pages that add little value compared with other pages on the same topic.

Read together, those lines describe a test on output rather than on provenance, and that shift is the reason I stopped writing tool bans into content policies. Search Engine Land's recap is the version most teams will actually read, and it is a fair summary of what changed in January 2025. The primary source is still the PDF, which is free and searchable, and which the September 2025 revision updated without walking any of those sentences back.

Notice what raters are not instructed to do anywhere in that document. Nobody is told to run a detector and then assign Lowest. Detection is not the doctrine, which is fortunate, because detector output is noisy enough that any policy built on it will eventually accuse your best human writer of being a robot, and you will spend a humiliating afternoon defending her. I have had that afternoon. The test is effort, originality, and added value: good news if you use models as drafting tools, bad news if you run them as a factory floor.

What makes AI content low effort?

Back to that 2,400-word page, because the failure it represents is the most common one I see. The failure is paraphrase with no residue. It is boring rather than scandalous, which is exactly why it ships: nobody in the review meeting objects, because the page reads fine. Reading fine is the trap.

The pattern almost always looks like this. Research is the current top ten pasted into a prompt. The draft is a fluent restatement of that consensus, garnished with a statistic the model invented on the spot. Editing is a pass on tone, headings, and a stock image. Publishing adds a byline and the words "human reviewed." Every one of those steps consumed real hours, and not one of those hours landed anywhere a rater could see.

Score a page like that against the six visible signals and you get zeros nearly across the board: no first-party information, no artifacts, thin synthesis, competent craft, no curation fingerprints, no added value over the results it was quietly copied from. The human review was production effort, and it never converted into Content Effort. As Cyrus Shepard's rater note puts it, the scorer sees the finished page and never sees your ticket queue.

I call this the Origin versus effort rule for AI-assisted Main Content, and it is the only content-policy sentence I have gotten a nervous founder to remember. You are allowed to keep the model. You are not allowed to keep the empty residue. When the only genuinely unique sentence on a 2,400-word page is the AI disclosure, you do not have a content asset. You have a process memo with headings.

The other low-effort tells do not need a detector either, and you can spot most of them in a minute:

• The page restates a single source, often a forum thread or a news article, and adds no context the reader could not get from the original.
• Product names have been swapped into a template, which is the old affiliate spam pattern running at a lower cost per page.
• Model artifacts survive into the published text, including training-cutoff disclaimers. The 2025 handbook uses a medical page exactly like this as a scaled abuse example, and health content is the worst place to leave that seam showing.

Google's public people-first guidance fails the same URLs in plainer language: no original information, no first-hand expertise, mainly summarizing what others have already said.

The fix to your review culture is one question, and it is not the one most teams ask. Not "did we use a model," but "where is the artifact?" If nobody in the room can point at a unique claim, the page goes back for residue instead of shipping. Slower on a Tuesday, cheaper than a core-update postmortem in March.

How do you raise Content Effort on AI-assisted pages?

Please do not tell a writer to "add originality." I did that for a while and got back adjectives. Map the additions to the zeros you scored instead. The cleanest way I have found to show this to a skeptical founder is one outline with two finishes, because the outline holds constant and only the evidence changes.

Finish A. The model drafts "how to instrument activation events." An editor fixes the voice, adds a subhead, and ships it. Nobody instrumented anything. The page scores in the 0-3 band because nothing on it could not be regenerated by anyone else in ninety seconds.

Finish B. An operator instruments 18 activation events on a real product, screenshots the event debugger on the day it ran, and names the two failures worth naming: a signup_completed event that double-fired on OAuth callbacks, and a payload where plan_id arrived as a string on web and an integer on iOS, quietly splitting the funnel for three weeks. The screenshot and both failures go into the draft, dated. The model still handles structure and transitions, because that is what it is good at.

Same generator, wildly different Content Effort, and that is the row leadership needs to see. If the only input they ever hear about is "we used AI," they are looking at the wrong column.

Here is the add list we reach for when a generator is in the loop. Read it as a preview of the six signals rather than the full scoring lecture, which lives over in the audit checklist. Left column is the signal you scored a zero on, right column is what a model cannot fake cheaply.

Every add in that right column requires access to something the model does not have, which is the test I apply when I cannot decide whether an addition counts.

Two habits make this stick, and both are slightly annoying. Write the worked example before the prose, so the draft is built around evidence rather than decorated with it once the shape is set. And name at least one failure per page, because a model will never volunteer the thing that broke on your account and your competitors are too image-conscious to admit theirs. Information gain governs the outcome here: unique useful sentences beat unique adjectives every time, and the information gain content framework covers that side properly.

Is scaled AI publishing the same as low Content Effort?

No, and conflating the two costs teams real money in both directions. Low Content Effort is a page diagnosis. Scaled content abuse is a site diagnosis, defined by an abundance of pages with little effort, little originality, and no curation. One blog post that used a model for its outline is not a spam problem, and an SEO lead who tells a founder otherwise will spend a quarter defending a tool ban that fixes nothing. The reverse error is quieter and worse: hiding a 400-URL template farm behind "every page had a human in the loop" when the human only rewrote the titles.

The decision rules are short enough to keep in your head:

• One URL with a low score: add real evidence, or merge it into the page that deserves the traffic.
• Many URLs with a shared skeleton and low scores: stop the generator job first, then decide which topics deserve a genuine residue. Remediating page by page while the job keeps producing is a treadmill you will lose.
• Many URLs with high scores and real artifacts: that is a production system, not abuse. Whether it pencils out is a business question, not a rater question.

Sampling settles this faster than debate. Pull three URLs from the affected directory and score them cold. If they share a skeleton and all land under four, you have a site pattern and should go find the job that made them. If the scores scatter, you have individual pages to fix and no crisis. The Content Effort complete guide keeps the evidence layers separate, including the leaked contentEffort attribute, an LLM-based effort estimate that observers have mapped onto the rater criterion with no confirmation from Google that the two are the same.

This is where the argument stops being a culture war and turns into an operating rule, which is the only form in which it survives contact with a busy team. Encode the Origin versus effort rule for AI-assisted Main Content as a reusable skill rather than a paragraph in a doc nobody opens, then let an agent score each URL on visible evidence alone. The standing question becomes "where is the artifact?" instead of "did we sin by prompting."

Metaflow is where I keep that logic now. The skills hold the scoring rules, the workflows run them against specified URLs, and the context carries what we already know about the account. Metaflow's Content Effort audit template runs that review with no detector theater, so drafts stay happily AI-assisted while shipping still requires residue.

The compound part matters more than the automation. A single Metaflow audit tells you which pages are thin. A review that runs the same rules weekly, with the previous scores still in context, tells you whether your process is improving or whether you have just gotten more confident about publishing paraphrase. I would rather know.

Frequently Asked Questions

Can AI-generated content have high Content Effort?

Yes, when first-party information, artifacts, and added value are visible on the finished page. The rater guidelines are explicit that generative AI alone does not determine effort or Page Quality, so your policy should require residue rather than prohibit tools. Ban empty pages, not software. Metaflow’s audit scores the residue, not the prompt log.

Does Google penalize AI content?

There is no published penalty that fires on AI origin. What exists is a rater instruction to apply Lowest to recycled or paraphrased Main Content with little to no effort, originality, and added value, including when a model produced it, plus a spam policy covering scaled production of low-value pages. Origin by itself is not the test. Metaflow labels that mapping as inference whenever a leak writeup shows up in the same conversation.

What makes AI content low effort?

Search results as the corpus, the model as the author, and the human limited to a tone pass. No original measurement, no unique artifact, and no job the existing results fail to do for the reader. Word count does not compensate for that, and neither does a "human reviewed" label.

How do you raise Content Effort on AI-assisted pages?

Keep the model for structure if it genuinely helps, then add what it cannot cheaply invent: your measurements, dated screenshots, primary documents, worked examples, named failures, and decision rules with real thresholds. Re-score the six signals after the addition rather than before. If the score has not moved, you added prose instead of effort, which is an easy mistake and a fixable one. That re-score is the job Metaflow’s template is for.

Is scaled AI publishing the same as low Content Effort?

No. Scaled publishing is a systemic pattern of many low-effort pages, while low Content Effort can describe a single URL that simply needs better evidence. Sample three URLs before deciding which problem you have. If they share a skeleton, escalate to the pipeline that produced them. If they do not, fix the page in front of you. Mixing the labels panics leadership while leaving the actual factory running.

Sources

Two of these carry the load. The rater guidelines PDF is the primary text for every claim here about effort, originality, and added value, and it is worth searching yourself rather than trusting anybody's summary, including mine. The Search Engine Land recap is the secondary source explaining what changed in January 2025. Google's people-first guidance and spam policies are the public-facing versions of the same test, written for publishers instead of raters. The rest is analysis rather than documentation, and the leak commentary should be read as inference: it describes an attribute found in the Content Warehouse documentation, and the mapping onto the rater criterion is an interpretation, not a Google statement. Keep those three tiers separate in a deck, because someone in the room will check.

• Google Search Quality Evaluator Guidelines (PDF)
• Search Engine Land: Google quality raters and AI-generated content
• Google Search Central: Creating helpful, reliable, people-first content
• Google Search Central: Spam policies
• Google Search Central: E-A-T gets an extra E for Experience
• Cyrus Shepard / Zyppy Signal: Content Effort
• Searchable: contentEffort leak analysis
• Hobo: What is Google's Content Effort signal?
• Metaflow: Content Effort complete guide
• Metaflow: Content Effort audit checklist