AI for Content Operations: What Works and What Still Needs People

This guide answers one question for content and SEO leads: which parts of content operations should you hand to AI, and which should stay with people? The short version is that language models are excellent readers and mediocre authors. They are reliable at reading, classifying, comparing and flagging across thousands of pages, and unreliable at original thinking, strategy and sign-off. Below you will find a task-fit table, the operating model that follows from it, a QA rubric for AI output, a worked example of a content health pass, and the questions to ask before you buy a tool.

What Google actually says about AI content

Start here, because a lot of AI strategy is built on myths in both directions. Google’s position has been stable since it was first published in 2023:

• AI use is not the problem. Google’s guidance about AI-generated content says appropriate use of AI or automation is not against its guidelines, and that it rewards quality content however it is produced.
• Mass-producing low-value pages is. The scaled content abuse spam policy, introduced alongside the March 2024 core update, names “using generative AI tools or other similar tools to generate many pages without adding value for users” as an example. The policy is about intent and value, regardless of how the content was created.
• You are expected to check the output. Google’s guidance on using generative AI content points out that generative models don’t retrieve facts, so output can be wrong, and asks you to review accuracy and quality, including titles, meta descriptions, structured data and alt text.
• Automation is a warning sign when it replaces value. The helpful content guidance lists “using extensive automation to produce content on many topics” among the signs of search-engine-first content.

Read together, the rule for content teams is simple. AI that helps you understand, check and improve what you publish is on safe ground. AI that multiplies pages nobody needed is not. That maps neatly onto what models are actually good at.

The task-fit table

The strongest AI tasks share three traits: they are bounded and text-based, the output can be checked against a source of truth, and a false positive is cheap because a person reviews it. Use this table to place any task before you automate it.

The pattern: the further a task sits from “read and compare” and the closer it gets to “decide and stand behind it”, the smaller AI’s role should be.

Where AI reliably helps

dash.morrison.app/lumenpay.com/chat

Illustration

lumenpay.com/Blog inventory

What do you want to know about lumenpay.com?

Ask about your website content...

Search ConsoleOutdated & Thin Content Finder

Claude Sonnet 5.5

Inventory and classification

You cannot manage what you have not catalogued, and for anything above a few hundred pages a manual inventory is out of date the day it is finished. A model can read each page and label it by topic, page type, search intent, funnel stage and freshness. It does not need to be creative; it needs to be consistent. Once the inventory exists, the problems that were always there become visible: thin pages, near-duplicates and pages no internal link points to. See content inventory for the workflow, and how to audit website content at scale for the full audit method.

Pattern detection across the whole site

Once pages are joined with performance data, AI can watch for patterns no editorial calendar can hold in its head. The system proposes candidates; people validate them.

• Content decay.Compare each page’s recent clicks and positions with its own baseline and flag the ones that turned. A person only notices a slipping article if they happen to look at it; an automated check looks at every page, every cycle. See content decay detection and what content decay is.
• Keyword cannibalization. Find queries where two or more of your URLs split impressions, and show the evidence. The model does not decide whether to merge, redirect or differentiate. See how to find and fix keyword cannibalization.
• Cross-page contradictions. A price, a spec or a policy stated one way on a launch page and another way on an FAQ. Cross-page consistency checks are trivial per comparison and impossible by hand across thousands of page pairs.
• Outdated claims. Dated statistics, old product names, superseded regulations. Outdated content finding is pure reading work.

Research and briefs

Research is the bottleneck in most content programs: reading the SERP, mapping gaps, checking what you already cover. AI compresses the mechanical part so the strategist spends their time on judgment. The quality ceiling depends on grounding. A brief built from your own inventory knows what you have covered, where the gaps are and which internal pages to link; a brief from a generic prompt does not. See research-driven briefs, content gap analysis and how to write content briefs.

Governance and compliance coverage

A compliance reviewer can read a handful of pages a day properly. A model can check every page against a written playbook and return only the flagged lines, each with the rule that triggered it. The gain is coverage and speed of the first pass, not autonomous enforcement. This matters most for YMYL topics, where Google says it gives more weight to trust, and where a forgotten landing page can carry real liability. See compliance and accuracy scanning, brand voice audits and E-E-A-T assessment.

One correction to a common pitch: an E-E-A-T check helps you find pages that lack evidence of experience, expertise and trust, but it is not about getting ahead of “quality raters”. Google says rater data is not used directly in ranking, and that E-E-A-T itself is not a specific ranking factor. Treat it as a reader-trust checklist, which is what it is. Our E-E-A-T guide for content teams goes deeper.

Metadata and on-page QA

Titles, meta descriptions, heading structure, schema and internal links are rule-based and repetitive, which is exactly where AI is dependable. It can flag missing, duplicated or mismatched titles and metas, pages whose content does not match the search intent of their queries, and schema gaps. Drafted metadata still needs a human read: Google’s generative AI guidance applies the same accuracy bar to titles, descriptions and structured data as to body copy.

Where AI still falls short

Original content

Fluency is not the problem; convergence is. Models trained on the same web produce the same structures and phrases. Ask one for a post from scratch and you get the average of everything it has read on the topic: professional, complete, and saying nothing a hundred other pages do not. That is precisely what Google’s helpful content questions screen for (“mainly summarizing what others have to say without adding much value”). AI can give an editor a starting structure. The parts that create value stay human: original data, first-hand experience, a point of view.

Strategy and trade-offs

Three articles cannibalize each other for a valuable query. A model can show the overlap and suggest consolidating. It cannot see that one article carries a CTA sales depends on, that another was just shared by a partner, and that the third targets a buyer the keyword data cannot distinguish. The strategist knows those things; the model does not. The same applies to portfolio calls: new content versus refreshes, a new topic versus a deeper cluster, retiring an on-brand section that underperforms.

Brand voice: checking yes, creating no

A model can check copy against a voice guide and flag drift, jargon or a banned phrase. That is pattern matching against a rubric, and it works. It cannot originate the voice, which comes from positioning, audience and a long list of deliberate choices. The tool reproduces; people create. See brand voice at scale.

Context that lives outside the text

The most important inputs often live where no crawler reaches: sales calls, customer interviews, the roadmap, the fact that a feature is being retired next quarter. You can narrow this gap by writing context down (audience, positioning, off-limits pages, compliance rules) where your tools can read it, but someone has to keep it current.

Sign-off and accountability

AI can flag a probable compliance issue. It cannot decide whether you are willing to stand behind a claim if challenged. That is a legal and business decision, and the person who makes it needs to be named.

The operating model: AI reads, people decide

Teams that get lasting value from AI split the work the same way, even if they describe it differently. AI handles four kinds of work:

1. Observation. Reading and indexing the content estate, and joining it with performance data.
2. Classification. Labelling and scoring pages against rubrics and taxonomies you define.
3. Comparison. Measuring pages against guidelines, competitors and each other.
4. Flagging. Surfacing anomalies and opportunities with the evidence attached.

People set goals, write the rubrics, make trade-offs, approve changes and own outcomes. The hand-off should be explicit: “here is what I found, why I flagged it, and the evidence” on one side; “approved”, “rejected” or “needs more context” on the other. No silent autonomy.

A QA rubric for AI output

Before any AI-produced finding or draft drives a change, it should pass five checks. Adapt the wording to your team, but keep all five.

1. Evidence. Does each finding quote the page, cite the data or name the rule behind it? A score with no reason is not actionable. Reject it.
2. Correctness on a sample. Before you trust a check across the whole site, hand-check a random sample. As a rule of thumb, read 20 to 30 results; if more than about one in ten is wrong, fix the prompt or rubric before scaling it up.
3. Facts verified. Every number, date, name and claim in a draft is checked against a primary source by a person. Models do not retrieve facts; they predict text.
4. Value added.Would this page exist if search engines did not? Does it contain something your competitors’ pages do not? If the honest answer is no, it fails, however fluent it reads.
5. Reversibility. Match the review depth to the cost of being wrong. A suggested meta description is cheap to undo; a redirect, a deletion or a regulated claim is not, and needs a named approver.

Why grounding matters more than the model

Ask a general chat assistant to “analyze this page for SEO” and give it only a URL, and much of the answer will be guesswork: it has not read your other pages, your Search Console data or your internal links. A system that has crawled your site and is connected to your performance data can answer from evidence and cite it. That changes the trust equation. When a finding quotes your own page, your own numbers or your own rule, you can verify it in seconds. When it makes an unsourced assertion, you cannot, and neither can the CMO reading your report.

Most content intelligence tools follow the same four-layer pattern, and it is a useful lens for evaluating any of them:

A worked example: a content health pass

Here is how the model plays out on a hypothetical 2,000-page B2B site, using Morrison for the steps it actually does. The page count is illustrative; the steps are the same at 500 or 20,000 pages.

1. Read the site. Morrison’s site indexcrawls every page and stores its status, metadata, links and 90 days of Search Console clicks and impressions. People: confirm the crawl covers the sections you care about and excludes the ones you don’t.
2. Score every page. After each crawl, the content analyzer classifies each page by type and search intent and scores it from Weak to Excellent on eight dimensions taken from Google’s helpful-content and spam guidance. People: open 20 or so scored pages and check you agree with the reasoning before you act on the scores.
3. Shortlist. Find the pages that score Weak or Limited on helpfulness or originality but still earn clicks. These are your refresh candidates: proven demand, weak page. For one section, ask chat: it lists the pages by overall score, reads their scans and checks their clicks in Search Console. For the whole site, export the URL grid to CSV and sort it.
4. Diagnose in chat. In chat, ask which queries are split across two or more URLs, or which pages lost the most clicks against their baseline. Chat answers from the whole site index, with every page’s scores, type, intent and Search Console data, and reads whatever pages a question needs. To run the same steps on every page of a large section, use a workflow, up to 500 pages a run.
5. Batch the mechanical fixes. For something like new title and meta drafts across a section, a workflow runs the same steps URL by URL (Search Console queries, live Google results, then an AI writing step on an OpenAI or Gemini model) and writes each result next to its URL. Test it on one page first. People: review every draft before it goes into the CMS.
6. Decide. The strategist goes through the shortlist and picks refresh, merge, redirect or leave for each page, using the evidence. This is the step that does not get automated.
7. Keep watching. Put the recurring questions (which pages dropped, which new pages overlap with old ones) on an automation that emails the answer: weekly or monthly on Starter, up to daily on Pro, up to hourly on Business. Workflows are run by hand, so re-run them when a section grows.

Notice where the people are: setting scope, checking samples, approving every change that ships, and making every decision about what a page should become. That is not a limitation of the tool. It is the design.

How to evaluate AI content tools

Criteria that matter

• Grounding. Does it read your actual site and connect to your data, or work from a paste box?
• Transparent reasoning. Can you see why something was flagged: the quote, the data, the rule?
• Whole-site coverage. Can you define a check once and run it on every page in a section, without pasting pages in one by one?
• Recurring runs. Can the questions you ask every week run on a schedule and arrive by email?
• Your rules, not only theirs. Can you write your own rubrics and playbooks and share them with the team?
• Export. Can results leave the tool and go into your tickets, sheets and reports?

Red flags

• “Fully autonomous content creation.”A pitch built on publishing without people is a pitch for the pattern Google’s scaled content abuse policy describes.
• Opaque scores.“73/100” means nothing without the rubric and the reasons.
• Five-page demos. Ask what happens at 5,000 pages, and how false positives are handled.
• No data connections. Without Search Console and analytics, a tool can describe your pages but not tell you which ones matter.

Questions to ask vendors

1. How does the system read my actual pages, and how often does it refresh them?
2. Which data sources does it join to the content?
3. Can I write my own rules and checks, and share them across the team?
4. How many pages can one analysis cover, and what happens beyond that?
5. Can I see the evidence behind every flag?
6. What runs on a schedule, and how do results reach me?
7. Can I export everything?

Where Morrison fits

Morrison is our product, built around the split in this guide: it reads and checks, your team decides. For clarity, here is what it does and does not do.

• Reads the whole site with the site index, and scores every page with the content analyzer.
• Joins your data from Search Console, GA4, Google Ads, Ahrefs and live Google results via SerpApi, through integrations, and researches the open web with web search and Perplexity.
• Follows playbooks: 23 system Skills, such as the Brand Voice & Readability Auditor and the Outdated & Thin Content Finder, plus Skills you write yourself and share across your organization on every plan.
• Remembers your rules: website context for audience, voice and off-limits pages, shared with the team; personal context that follows you across websites; memory of what chat learns about you and the site; and knowledge bases you upload, such as a style guide, which chat searches and workflow AI steps can use.
• Does not publish to your CMS or decide for you. Every change it proposes goes through your team.

What will stay true as models improve

Models will keep getting faster, cheaper and better at reasoning. A few things will not change with them:

• Unoriginal pages at volume stay a liability. Google’s spam policies target the pattern, not the tool, and every competitor has the same models you do.
• Grounding beats raw capability. A better model with no access to your site still guesses.
• Governance grows in value.The more content exists, yours and everyone else’s, the more it is worth to know that every page you publish is accurate, consistent and current. See the complete guide to content governance.
• Expertise gets scarcer, not cheaper. When average content costs nothing, first-hand experience, original data and a clear point of view are what still set a page apart.

If you are scaling output and want the process side of this, read how to scale content production without sacrificing quality. For a first-hand account of the model in daily agency work, see how Bonzer runs agentic content ops with Morrison.