Stop Asking Codex to Write 100 Articles: The Most Valuable SEO Work on a Mature Site Is Managing Existing Content

Over the past few years, as AI has exploded, I have watched more and more teams reduce SEO to one simple action: producing content. Research 100 keywords today, ask AI to write 100 articles tomorrow, and create 100 pages the day after. As long as the publishing count keeps rising, the team feels that SEO is moving forward.

Is that logic correct? Not completely. A new site with only a handful of pages that Google can see obviously needs to build its foundation. The problem is that many sites have already been running for two or three years and have hundreds or thousands of pages, yet their teams still use the same answer for every problem: rankings fall, so they write another article; traffic drops, so they write ten more; an old page becomes outdated, so they pretend not to notice and keep publishing.

This is no longer an execution-speed problem. It means the team has a fundamentally incomplete understanding of SEO. On-site SEO is not just content production; it is content-asset management. If content is an asset, you cannot only add and never maintain it, or measure this year’s new pages without asking how much value last year’s pages still retain.

You Can See That Competitors Have Many Articles, but Not How They Manage Them

Why do teams get trapped in the path of publishing more articles? Because that is the easiest thing for an SEO beginner to see during competitor research. You find a competitor with strong organic traffic, discover thousands of blog posts, and conclude that you should build a blog too. This is natural, and it is not entirely wrong. Many companies begin their SEO systems this way.

But Semrush, Ahrefs, and public URLs do not show how a competitor maintains its content assets. You cannot see how often it audits content, how it decides whether a page should be kept, updated, merged, or deleted, or whether it protects titles, internal links, Schema, and structures that already bring clicks. You cannot see whether it rescues a page when CTR first slips or a ranking drifts from position 4 to position 8, rather than waiting until all the traffic is gone.

You copied the competitor’s content output, but not its ability to manage content assets. For a mature site, that latter capability is often worth much more.

A New Page Starts at Zero; an Old Page Goes into Surgery with a Fortune

New pages are easy to handle. They have no rankings, historical clicks, backlinks, internal links, accumulated search intent, or established conversion path. You can change almost anything because the starting point is zero.

Updating an old page is different. A page that already ranks may have stable queries, dozens of internal links, backlinks, Schema, repeated crawling and understanding from Google, and daily inquiries, registrations, or orders. You are not facing a blank sheet. You are operating on an asset that still generates cash flow.

This is what makes many AI content updates dangerous. A team sees traffic falling and asks a model to rewrite the page from scratch. The sentences become smoother, the subheadings multiply, and the word count grows from 2,000 to 5,000. Then the queries that used to rank disappear, title CTR falls, internal-link anchors no longer match, structured data is overwritten, and Google has to reinterpret the whole page. That is not an update; it is an SEO demotion packaged to look like diligence.

The Search Engine Land case I referenced describes an airport-transfer site updating an Antalya airport page. The author locked a 56-day baseline, checked queries, page intent, user needs, local information, images, Schema, and internal links, and changed only what needed to change. Before the update, the page had 148,537 impressions over 56 days, average position 14.89, and CTR 1.49%. Afterward, average daily clicks rose from 39.55 to 49, average position improved to 10.87, and CTR rose to 1.79%. Across 59 tests normalized to comparable 56-day windows, the work produced 2,284 additional organic clicks; organic purchases rose 24.9% and organic revenue 20.1% relative to baseline.

These are not my project numbers and cannot be extrapolated to every site. Seasonality, brand, page type, and methodology matter. But they show that, for a mature site with search assets, targeted updating can be closer to a commercial result than blindly adding pages. The valuable action is not rewriting the page to sound more like AI. It is finding the small place that is bleeding and fixing it without dismantling what still works.

The Original Used Claude Code; I Would Rather Turn It into a Codex Workflow

The most valuable part of the original is not that it used Claude Code or that it had 14 steps. The valuable idea is that judgment is not delegated to the machine; the machine executes and constrains judgments that have already been formed.

Many people still use AI as a cheaper writer: provide a keyword, drop in an old article, ask for a rewrite based on current trends, and expect the model to know which rankings to protect, which facts are outdated, which paragraphs drive conversion, and which links must not move. It cannot know that by itself. It does not have your GSC data, GA4 conversion data, content asset map, brand constraints, product facts, or policy boundaries.

The core of a Codex version should therefore not be asking Codex to refresh an article. It should turn the SEO team’s validated management discipline into a workflow that Codex must follow. Useful infrastructure includes AGENTS.md for durable rules, Skills for reusable processes, MCP or scripts for live external data, Scheduled Tasks for recurring monitoring, and Git diffs, tests, checks, and reviews for every real change. This is not merely replacing Claude Code with Codex; it is building a content-asset operating system.

First Layer: Use AGENTS.md to Protect What Must Not Be Changed Casually

Codex reads AGENTS.md before work begins, with rules layered from global scope to repository and subdirectory scope. In SEO, this file should contain verifiable asset-safety rules rather than vague instructions such as “write professionally.”

## Content update invariants

- Do not modify a page until the pre-update baseline is locked.
- Preserve the URL slug by default.
- Do not rewrite paragraphs marked Keep; only correct verified facts.
- Put primary queries that already bring clicks on the protection list.
- Do not change a meta title by default when it has stable CTR.
- Only extend or repair Schema; do not replace it wholesale without a reason.
- Back up the original file and output a section-level diff afterward.
- Do not send content to the CMS until build, link, Schema, and rendering checks pass.
- Codex may generate drafts or staging versions only; it may not publish publicly by itself.

Most SEO accidents are not mysterious. Someone enthusiastically optimizes a page and changes its URL, title, links, structure, facts, and intent at the same time. AGENTS.md makes the boundaries explicit so Codex enters the project knowing this is asset maintenance, not a free-form writing contest.

Second Layer: Break the Judgment Process into Small, Explicit Skills

Do not build an all-purpose SEO master Skill. It will become a huge prompt junk drawer. Use bounded units instead.

content-decay-radar reads page-level GSC, GA4, or revenue data over fixed windows and creates a candidate queue for falling clicks, worsening CTR, ranking drift, shrinking query coverage, or declining conversions. It does not edit content.

content-asset-audit labels paragraphs Keep, Fix, Remove, or Add and attaches evidence to every label. Without evidence, it marks the item for human judgment rather than forcing a classification.

serp-and-intent-gap checks the current SERP, competing pages, search intent, and possible cannibalization. It answers what is missing or misaligned rather than copying a competitor directory.

delta-brief describes only changes: what stays, what is fixed, what is removed, what is added, why it changes, and how it will be accepted. content-delta-editor changes only Fix and Add sections while preserving voice, facts, and protected queries. fact-and-freshness-checker verifies new and retained content, including prices, features, locations, policies, dates, screenshots, and statistics.

seo-preservation-auditor checks slug, title, description, canonical, Schema, links, anchors, images, and rendering. publish-auditor generates staging output, diffs, and a final report; any hard-gate failure stops the workflow. These Skills are not an out-of-the-box SEO package. They are business workflows created for a site’s own data and rules.

Third Layer: Connect Codex to Facts, Not Imagination, with MCP and Scripts

Codex recommends MCP when information lives outside the repository, changes frequently, and must be read repeatedly. SEO fits this exactly: query and indexing data may be in GSC, sessions and conversions in GA4, keyword and backlink data in Ahrefs, Semrush, or an internal database, source files in Git, product facts in Notion or a product system, and final content in a CMS.

Without those connections, Codex sees only an isolated Markdown file. Asking it to decide how to change a page from that file alone is not very different from asking a doctor who has not read the medical record to perform surgery. Start with the most frequent painful loop rather than connecting every tool at once:

1. GSC page and query data.
2. Site content source files and a content asset inventory.
3. A CMS draft or staging interface.

If the team uses SQLite, a script can store pages, target keywords, update times, query baselines, internal-link relationships, conversion types, and review status in one source of truth. Otherwise the system checks a spreadsheet today, searches chat history tomorrow, and guesses the day after. For content assets, too many sources of truth make automation develop multiple personalities.

How a Codex Content Update Workflow Should Actually Run

I would compress the process into six stages.

flowchart LR
    A[Decay Radar<br/>Find decline] --> B[Baseline & Snapshot<br/>Lock baseline and backup]
    B --> C[Audit & Research<br/>Audit sections and intent]
    C --> D[Delta Brief<br/>Define changes only]
    D --> E[Edit & Verify<br/>Edit and verify]
    E --> F[Staging & Measure<br/>Publish after approval and review]

Stage One: Find the Right Pages; Do Not Update Something Just Because It Is Old

Content decay does not mean the publication date is old. Some three-year-old pages still rank steadily, contain current facts, and answer an unchanged need. The right candidates show business signals: clicks and rankings decline consistently; impressions remain stable while CTR worsens; core queries leave the top ten; new queries appear without real answers; conversions, inquiries, or revenue weaken; competitors gain information advantages; or facts, product capabilities, and local information change.

Codex can generate the queue, but priority cannot be based only on traffic loss. A blog losing 100 visits a day without conversions may be less urgent than a commercial page losing 10 visits while supporting orders. Look at commercial value first, then decline severity. Revenue decides what to inspect first; diagnosis decides how to repair it.

Stage Two: Lock the Baseline and Take a Complete Snapshot

Before Codex changes a word, save the original file, page HTML, query baseline, title, Schema, internal links, screenshots, and conversion data. Use 28 days, 56 days, or a year-over-year window according to seasonality and data volume. The 56-day window in the original fits a particular travel case and its testing cycle; it is not universal.

What cannot be skipped is lock first, modify second. Without a baseline, a traffic increase may be seasonality, an algorithm change, a campaign, or luck. A decline will not tell you which step damaged the original value.

Stage Three: Label Every Section

Keep means currently valid and valuable, not permanently correct. Fix means the direction is right but facts, expression, structure, or intent coverage have fallen behind. Remove requires caution because deleted content may serve an unnoticed long-tail query. Add should come from a query gap, user task, product change, or evidence gap, not from copying a competitor’s H2.

Codex is most useful here for mapping evidence to sections: which query maps to which paragraph, which paragraph lacks an answer, which fact is outdated, and which section overlaps another page. Once this mapping is correct, writing is the cheap part.

Stage Four: Write a Delta Brief Before You Touch the Body

A conventional content brief is designed for a new page: target keyword, structure, headings, FAQs, and competitors are rebuilt from scratch. That naturally encourages a rewrite. A Delta Brief describes only changes: what stays untouched, what needs fact checking, what is rewritten and why, which intent is missing, what module is added, which SEO value is protected, and what evidence defines success. The closer it looks to an engineering change request, the better. The article is for the user; the change request is for the execution system.

Stage Five: Write Only the Delta, Then Review Content Like Code

During editing, Codex may change only Fix and Add sections. Review the section-level diff: what was removed and added, which numbers and links changed, and whether title, Schema, or structure was touched. In a code repository, Codex can edit Markdown, MDX, JSON, components, and Schema, run lint, builds, link checks, and smoke tests, and hand the diff to a human.

Codex can also implement a comparison table, selector, calculator, or FAQ component when the brief needs one. But generating a component does not replace product or SEO judgment. A polished module that solves no real task only adds JavaScript and maintenance cost.

Stage Six: Stage First, Publish Second, Then Review with the Same Standard

Codex can send content to staging through MCP or a CMS API, or generate a draft for review. Public publishing should remain a human action, especially for commercial pages. Before publishing, check the section diff, links and anchors, canonical, robots, Schema, build, rendering, CTAs, conversion events, mobile and performance risks, fact sources, timestamps, and final human judgment. If any hard check fails, stop.

After launch, use the same metric definitions locked before the update. Review rankings, clicks, CTR, query coverage, and, for commercial pages, registrations, inquiries, purchases, and revenue. Looking only at rankings turns an SEO update into self-congratulation.

Scheduled Tasks Are Where This System Starts to Compound

Once the single-page process works, Scheduled Tasks are the next step. A Skill defines the method; a Scheduled Task defines the timing. Weekly, Codex can generate a decay queue; daily, it can check build, indexing, and links for recent updates; monthly, it can summarize experiments; and when a threshold is reached, it can create an audit task.

This means automatically creating tasks and reports, not rushing into the CMS to alter live pages. Mature automation runs low-risk, repetitive, verifiable actions automatically while high-risk, irreversible actions that require business judgment wait at a human gate. An agent that is busy every day while nobody knows what happened to traffic or revenue is not automation. It is an electronic pet.

Research in Parallel, but Do Not Let Multiple Agents Edit the Same High-Value Page

Codex supports multiple agents, which helps with audits. Different agents can read GSC, analyze the SERP, check facts, scan internal links, and audit images, after which the main task consolidates their findings.

But a page with rankings, conversions, and historical assets is not a place for agents to compete over edits. Parallelism is safest for exploration, testing, log analysis, and summaries. Writes to shared files need caution and, when necessary, Git worktree isolation.

My recommendation is simple: research can run in parallel, decisions must converge, writes must be serialized, and publishing requires approval. A dashboard with six running agents is not proof that the project is moving quickly.

Codex Will Not Automatically Solve Content Asset Management

Once Skills, MCP, Scheduled Tasks, and AGENTS.md are configured, the site still will not grow by itself. Codex scales execution and constraints; it does not possess business judgment. It does not know why a product line matters this year, that a low-traffic page is a sales trust document, or that two similar pages serve different countries, regulations, and buyers.

It also does not know how much short-term traffic you would trade for brand language or long-term value. Those decisions should not be made by the model. Codex is better understood as the execution system for a content-asset operations team: it remembers rules, reads data, runs audits, creates Delta Briefs, edits files, verifies results, leaves diffs, and monitors the system.

What to do, why to do it, and what result is acceptable remain human judgments. AI should not pretend to understand the business for you. It should execute the judgments you have already formed consistently a thousand times.

The Real Difference Is Not Whether You Use AI, but Whether You Manage Assets

Return to the SEO beginner’s competitor research. You see thousands of competitor articles and prepare to write thousands too. But the competitor’s real advantage may be knowing which page is declining, which deserves rescue, which section must not be touched, which query is close, and which old-page update is closer to revenue than ten new pages.

None of this appears directly in an Ahrefs URL list. It is hidden in data discipline, content audits, update queues, version history, testing methods, and publishing boundaries: the ability to manage content assets.

As AI makes writing cheaper, production itself is no longer scarce. What is scarce is knowing what should not be produced, what deserves maintenance, what must be protected, what should be deleted, and whether each change became traffic, conversions, and revenue.

If you already have a mature site with hundreds or thousands of pages, do not make your next step asking Codex to write another 100 articles. Build the content asset inventory. Find pages that are slowly bleeding but still have rankings, queries, links, and a conversion foundation. Turn your judgment into AGENTS.md, repeated methods into Skills, live data into MCP, hard checks into the pre-publish gate, and stable processes into Scheduled Tasks.

This is not as satisfying as one-click generation of 100 articles. It is messy, inconvenient, and requires repeatedly checking data, reviewing diffs, verifying facts, and protecting boundaries. But commercially valuable SEO work is often exactly this kind of work that is difficult to capture in a social post.

Do not ask AI only to produce content. Let it start managing your content assets. Writing is only the cheapest step.