Verdict: The OpenAI Codex app is now a real place to manage multi-week SEO work — not just to write code. As of the April–July 2026 platform updates, it has a project-grouped sidebar, a pull-request-style review pane, a live "task sidebar" that shows the agent's plan/sources/artifacts while it works, an artifact viewer for spreadsheets and PDFs, "projectless Chats" for research, and multi-window support. For an SEO practitioner or a small-business owner running content clusters, this is the first version where you can hand Codex a multi-step job, walk away, and come back to something you can approve, reject, or steer line-by-line. The catch is that your account tier, your patience for the review loop, and your willingness to write a plain-text AGENTS.md decide whether it saves you hours or just feels like a fancier chat box.
Last verified: 2026-07-26 · OpenAI Codex app (platform 26.415–26.226+) · Codex CLI 0.122+ · GPT-5-Codex / GPT-5.2-Codex. Pricing/limits change often — re-check the OpenAI Codex pricing page before you commit. TL;DR: Sidebar + reviews + side chats turn Codex into a managed agent — best for SEOs who already run multiple campaigns in parallel. Worst fit for one-off keyword lookups (use plain ChatGPT).
What changed in the Codex app that actually matters for SEO?
The Codex app gained three workspace features (released in platform 26.415 on 16 April 2026, sharpened through 26.519+ in May, and continued through mid-2026) that move it from "terminal with AI glued on" to a place where non-developers can manage long jobs:
- Project sidebar — every project you add becomes its own group; threads (a.k.a. "chats") live nested under it, each with its own sandbox and
AGENTS.md. (OpenAI Codex changelog; Daniel Vaughan, "Inside the Codex App Workspace," 17 April 2026) - Pull-request review pane — diff view with inline comments, approve/reject/ask-for-changes, and the ability to point Codex at a specific piece of feedback. Originally GitHub-scoped; the same review pattern now applies to any files Codex modifies in a project. (OpenAI, "Introducing upgrades to Codex," 15 September 2025; @codex review on GitHub)
- Task sidebar — a live panel that surfaces the agent's plan, the sources it gathered, files it touched, and a running summary — while it works. You can steer the plan mid-flight, not just at the start. (Codex changelog)
- Artifact viewer — previews PDFs, spreadsheets, slide decks, and documents inline in the sidebar before you commit them, so you can verify a generated keyword tracker or content brief without leaving the window. (Codex changelog)
- Projectless "Chats" — threads that aren't tied to a codebase, optimized for research, planning, and analysis (where most SEO brainstorming actually lives). Sandbox is lighter, no Git integration.
- Side chats within a thread — you can open a side conversation to ask "why did you pick that keyword?" or "draft me a title tag" without interrupting the main task running in the background. (Codex changelog)
- Memories — opt-in; lets Codex carry stable preferences, project conventions, and known pitfalls across threads. Off by default. (OpenAI Developers, "Memories")
For SEO, the headline is that you no longer have to copy a 4-week campaign's context into the prompt every Monday morning. The project thread keeps it. You approve or reject changes. You ask side questions without losing the main thread. None of that was true in 2025.
Codex vs ChatGPT for SEO work — which surface, what for?
Most people who pay for ChatGPT don't realize Codex is already in their plan. The two surfaces are not interchangeable for SEO work:
| Task | Use ChatGPT | Use Codex (CLI / app / IDE) |
|---|---|---|
| Brainstorm a content cluster, draft strategy | ✅ Better — fast, conversational, image input | ⚠️ Slower; use a projectless Chat |
| Write a single blog post or title tag | ✅ Faster for one-offs | Mostly overkill |
| Run keyword research + build a tracker spreadsheet | ⚠️ Possible but you re-enter context every time | ✅ Generates and iterates the file in a project; you review the .xlsx in the artifact viewer |
| Code a redirect map, a Schema.org generator, a scraper | ⚠️ Will give you code; you run it manually | ✅ Writes files, runs them, opens a PR for review |
| Review 50 generated content briefs vs a brief template | ⚠️ Manual | ✅ Batch review with diff view; reject/iterate per file |
| Long campaign (weeks) | ❌ Forgets between sessions | ✅ Project + threads + memories carry context |
| Tag a teammate's PR with "fix the meta description" | N/A | ✅ @codex review the meta tags on this PR on GitHub |
The rule of thumb: if the work finishes in one sitting and lives in a chat transcript, ChatGPT is enough. If it produces files you'll iterate over for days or weeks, Codex's review loop and project persistence pay for themselves. This mirrors how we already split AI work across surfaces at The Tech Archive — see our OpenAI Codex autonomous worker guide for the longer-pattern version.
What does it cost to run Codex for SEO? (2026)
Pricing is plan-tiered and changes often — verify at the official Codex pricing page before buying.
- ChatGPT Free — Codex included with strict usage limits. Enough to learn the workflow, not enough for a real campaign.
- ChatGPT Plus — $20/mo — best default for solo SEOs and small-business owners. Expanded Codex usage across app, CLI, IDE extension, web, and iOS. (OpenAI Codex pricing)
- ChatGPT Pro — $100/mo (5× tier) or $200/mo (20× tier) — 5× or 20× higher rate limits than Plus, plus access to GPT-5.3-Codex-Spark (a faster Codex model in research preview). Worth it when Codex saves paid hours across daily parallel projects. (OpenAI Codex pricing)
- Business / Enterprise — admin controls, shared workspaces, token-based credits, audit logs.
- API — separate token-billed access (e.g., GPT-5.3-Codex at $1.75 / $14.00 per 1M input/output tokens, per the OpenAI API pricing page as of mid-2026). API is for automation; if you're just using Codex as a coding agent, pay through the ChatGPT plan.
- Credits — On Plus and Pro you can buy on-demand credits to extend usage beyond plan limits. Common when a heavy week of campaign work bumps you into the cap. (OpenAI Help Center, "Using Credits for Flexible Usage")
If you're deciding purely on price for an SEO use case, our Claude vs ChatGPT for SEO in 2026 piece benchmarks the alternatives on the same workload — Codex wins on the agentic file-work pattern; Claude often wins on raw copy quality.
How do you set up a Codex project for an SEO campaign? (step by step)
This is the workflow we run inside the Codex app. It assumes Plus or higher and a project folder on your machine. Every step uses the new workspace features.
- Install the Codex CLI and sign in.
Open the Codex app (macOS, Windows, or Linux) and sign in with the ChatGPT account that has Codex access. (OpenAI Developers, "Codex CLI features")npm install -g @openai/codex - Create a project folder for the campaign. Put a plain-text
AGENTS.mdat its root. This file is the rules Codex follows every time it touches the project. For SEO, include things like: "Never invent keyword volumes — mark anything you can't verify as[unverified]." "Source every keyword's intent from a real SERP query." "Prefer spreadsheets over prose for trackers." See our deeper guide on AI agent ownership and maintenance for the canonicalAGENTS.mdpattern. - Add the project to the Codex app. Open the app, add the folder. It now appears as a group in the sidebar. Every thread you start inside it gets the project's sandbox and
AGENTS.md. - Start a thread for the first phase — e.g., keyword research. Be explicit up front: "Generate a keyword research spreadsheet for my SaaS landing pages. One row per keyword: keyword, monthly volume (cite source), SERP intent (informational/commercial/transactional), top 3 ranking URLs, and a one-line content angle. Save as
keyword-research.xlsx." - Watch the task sidebar while it works. You'll see the plan decomposed, the sources Codex pulled, and the file it's producing. If the plan looks wrong (e.g., it's about to fetch volumes from a site you don't trust), intervene with a side chat before it finishes.
- Open the artifact viewer to inspect
keyword-research.xlsxinline. Verify the volume column has sources. Reject the file (or ask for revisions) if numbers are unverified. - Approve or reject via the review pane. Treat it like a pull request: approve the rows that pass, leave inline comments on the ones that don't ("this volume looks inflated — re-check"), and ask Codex to revise. Iterate until it's ready.
- Repeat for the next phase in a new thread under the same project. Content briefs, then draft posts, then a link-building tracker. The project's memories carry what's already been done, so you don't re-explain.
The win is steps 5–7: you're managing the agent the way you'd manage a junior SEO, not rewriting its output by hand.
How do you keep a long SEO campaign from going off the rails?
Long campaigns are where the new context-keeping features matter most. Three things keep mine sane:
- One project per campaign, one thread per phase. Don't dump keyword research, content briefs, and link-building into one thread. Sidebar organization is the feature that makes parallel work legible — use it. (Multiple ongoing GitHub issues confirm the sidebar is the part users fight most; treat its organize-by-project behavior as a feature, not an accident. See openai/codex issue #23942 and #23825 for the kinds of pitfalls you'll see.)
- Use side chats for tangents, not the main thread. "Why this keyword?" "Draft me a meta description for the homepage." Side chats keep the main job running without polluting its context. This is also the single biggest time-saver in the May–June 2026 app updates.
- Let Memories carry stable preferences — but keep
AGENTS.mdfor hard rules. Memory is opt-in, off by default, and updates in the background after a thread goes idle. (OpenAI Developers, "Memories") Hard rules ("never publish fabricated stats") belong inAGENTS.mdwhere Codex will always load them. Memory is for soft signals ("the user prefers short pull-quotes"). For more on the memory/AGENTS split, see our perception agents writeup on how durable context changes long-horizon AI work.
If you find Codex drifting mid-campaign (touching files it shouldn't, ignoring a convention), the fix is almost always an AGENTS.md addition, not a longer prompt. Prompts expire; rules persist.
What this means for you
For most readers of The Tech Archive — small-business owners, indie SEOs, content ops folks — the practical move is:
- If you already pay for ChatGPT Plus, stop treating Codex like a developer-only tool. You already have it. Pick one recurring SEO job that produces files (a keyword tracker, a content-brief batch, a redirect map, a Schema.org generator), make a project folder, write a 20-line
AGENTS.md, and run one end-to-end campaign through the review loop. That single experiment will tell you whether Codex is saving you hours or just making you feel busy. - If you're on Pro or considering it, Codex's biggest unlock is parallelism — multiple projects, multiple threads, one window. That's the tier where the sidebar + reviews + side-chats combo really pays off. If you're not going to run at least three concurrent workstreams, Plus is enough. Our best AI automation platforms for small business guide benchmarks Plus-tier Codex against Zapier, Claude, and standalone agents for exactly this kind of multi-stream work.
- If you're a team, Business tier adds admin controls and shared workspaces — but only adopt it after someone on the team has shipped one solo campaign first. The review pane is only as good as the reviewer; have a human reviewer own it.
The honest summary: Codex for SEO is now less about writing better content (ChatGPT and Claude are still better there) and more about running the work as projects you can approve and ship. That shift is what the 2026 Codex app updates actually enable, and it's why the workflow above is the one worth copying.
FAQ
Q: Is OpenAI Codex free with ChatGPT? A: Yes. As of 2026, Codex is included on Free, Go, Plus, Pro, Business, Edu, and Enterprise plans, with usage limits that scale by tier. Free is enough to learn the workflow; Plus ($20/mo) is the practical default for running real campaigns. (OpenAI Codex pricing)
Q: Can a non-developer use the Codex app for SEO?
A: Yes, with a setup cost. You need a project folder, a plain-text AGENTS.md of rules Codex must follow, and willingness to use the review pane to approve/reject files. The app does the rest; no code from you is required to use it, though Codex can write small scripts (redirect maps, schema generators) for you and you review them in the diff view.
Q: What's the difference between a Codex project and a Codex Chat?
A: A Project is scoped to a folder (has a sandbox, full Git integration, runs your AGENTS.md). A Chat (the April 2026 "projectless chat") lives in ~/.codex/threads, has no Git integration, and is designed for research, planning, and analysis that doesn't belong to a codebase. For SEO, do keyword research and content briefs as Projects; do strategy brainstorming in Chats. (Daniel Vaughan, "Inside the Codex App Workspace," 17 April 2026)
Q: Does Codex remember my project context between sessions?
A: Project threads keep their own context. Optional Memories (off by default) can carry stable preferences and conventions across threads in the background, after a thread goes idle. Hard rules belong in AGENTS.md, which loads on every thread. (OpenAI Developers, "Memories")
Q: Can Codex review my SEO content briefs the way it reviews code? A: Yes, with limits. The review pane shows diffed files and inline comments; you can approve, reject, or ask for revisions. It works well for structured artefacts (spreadsheets, briefs as markdown files) but it's not a substitute for subject-matter review — it can flag formatting and consistency, not whether your content angle is actually competitive. Use it as a fast first-pass reviewer, then human-review the substance.
Q: Should I use Codex or Claude Code for SEO automation in 2026? A: Use Codex if you're already on ChatGPT Plus/Pro and want a unified app + cloud + IDE experience with the review pane. Use Claude Code if your work is mostly terminal-native and you want stronger repo-level reasoning at the cost of less polish. We benchmarked both for SEO workflows in our Claude for SEO in 2026 piece.
Q: How is this different from what we already published on Codex? A: Our earlier OpenAI Codex autonomous worker piece covers Codex as a long-horizon autonomous builder; this article is project and review workflow inside the Codex app for a non-developer running an SEO campaign. Different surface, different reader.
Discussion
0 comments