5 ChatGPT prompts for SEO: research, briefs, and checks

A prompt is a brief, not a shortcut

ChatGPT can organize a supplied list, propose questions, or turn approved evidence into a draft. It cannot prove that a keyword has demand, that a competitor page is current, or that changing a title will improve rankings. State what it should use and what it must not invent.

For every prompt, include:

• The reader and their problem.
• The query or page being considered.
• Approved sources or pasted evidence.
• The output format.
• Facts that need a citation or should be marked unknown.

1. Turn a seed topic into a research plan

Use this before treating a broad topic as a content calendar entry. For example, a support team migrating a knowledge base may need an article for administrators searching “move support articles to a new help center,” not another broad “knowledge base migration” post.

You are planning research, not estimating search volume. For administrators moving support articles to a new help center, propose 15 specific questions they may ask. Group them by intent. For each, say what evidence I would need before publishing and flag anything that needs keyword or SERP data from another tool. Return a table.

The useful output is a decision such as: “publish a migration checklist only if the approved platform documentation covers redirects, access changes, and the rollback path; otherwise research those gaps first.” This keeps the model from presenting invented volume or difficulty as research. Bring a real keyword export or Search Console data into the next pass if the decision depends on it.

Once you have that export, use a second pass to organize it:

Group the supplied queries by the administrator's task: planning the move, preserving URLs, checking access, and verifying the result. Preserve the export's original metrics and date range. Suggest which questions belong on one page and which need separate pages. Flag ambiguous intent; do not invent search volume or treat similar words as proof that two queries need the same answer.

The editor can then check whether one migration guide covers the cluster or whether, for example, redirect troubleshooting needs its own walkthrough. A cluster is an editorial hypothesis, not proof of topical authority or ranking potential.

2. Build an evidence-first content brief

The strongest prompt in this list is usually the brief. It prevents the writer from discovering the intended reader, key claims, and exclusions halfway through a draft.

Create a content brief for “move support articles to a new help center” for a support administrator. Use only the approved migration guide, redirect checklist, and permissions policy below. State the reader’s decision, the answer they need first, required sections, claims that need citations, examples to include, and claims we must not make. If a source does not establish a point, label it “needs evidence.” Output a one-page brief.

The result should be approved by the person who owns the subject before it becomes standing writing instruction.

3. Review an existing page for clarity and intent

Give the model the actual page copy and the query it is meant to serve.

Review this page for a reader searching “[query].” Do not rewrite it yet. Identify: (1) the answer delivered in the first 150 words, (2) unanswered reader questions, (3) claims that lack supplied evidence, (4) headings that do not match the likely intent, and (5) two metadata alternatives. Mark every conclusion that requires a live SERP check.

This produces an editorial checklist rather than an unsupported “SEO score.” Verify snippets and rankings in the relevant search tools before changing live metadata.

For the migration article, a useful follow-up is: “Rewrite only the paragraph explaining redirects for an administrator who is new to them. Preserve every technical qualification and keep the destination URL example unchanged.” Compare the before and after text. A simpler sentence should not turn a conditional migration step into a universal instruction. Ask for title and description options separately, keeping the actual page promise intact rather than adding an unproven benefit to attract clicks.

4. Compare a competing page without copying it

Competitor analysis is useful for finding missing questions, not for imitating prose or assuming the model has current access.

I have pasted a competing page and the approved sources for our article. Compare the coverage. List the competitor’s reader questions, its evidence types, and gaps our sources can answer better. Do not copy sentences or claim to have read pages not included here. Separate “source-backed opportunity” from “needs research.”

If you ask a browsing model to visit a URL, still ask it to identify the page version and cite the specific material it used. A URL alone is not evidence.

5. Draft technical SEO work for validation

Schema and crawl directives deserve more caution than ordinary prose because a small mistake can affect a live site.

Based only on this approved page content, draft Article JSON-LD with the headline, date, author, and image supplied below. Do not add claims or invent missing fields. Then list validation checks and any field that needs confirmation.

Validate the Article proposal with Google's Rich Results Test, then check the deployed page with URL Inspection. Valid markup does not guarantee a search feature or higher ranking. Google stopped showing FAQ rich results on May 7, 2026, so an older prompt promising that display is out of date. Article validation, Google's deprecation notice.

For a robots.txt change, name the exact host and paths, ask for a proposed file plus a plain-language explanation, and have the site owner validate it before deployment. Generated code is a proposal, not permission to change crawl access.

Use eesel CLI to keep an approved brief current

When a content team already uses an eesel blog-writer teammate, eesel CLI lets a person, script, or coding agent such as Codex, Claude Code, or Cursor inspect the same workspace setup as the dashboard. It returns JSON, needs Node.js 18.17 or newer, and makes inspection practical without copying screenshots. Setup and inspection are free; CLI chats are billed work.

Suppose the owner has approved the migration brief and an exported keyword list, but wants to check whether the blog writer has both before it drafts. Inspect the source inventory and its standing instructions:

npx @eesel/cli instructions --agent "SEO blog writer"
npx @eesel/cli files ls --agent "SEO blog writer"

Ask Codex to compare the returned file list and instructions with the owner-approved brief: are the migration guide, redirect checklist, permissions policy, and keyword export present, and do the instructions require the writer to flag missing evidence? This file command lists uploaded or generated files, not all connected knowledge. Supply the actual approved documents and export for content-level comparison; a filename alone does not prove their contents or freshness. The owner decides whether to update or add an approved source.

With approval for the chat cost, run two named fresh chats on an existing non-production teammate where consequential actions are disabled or absent. First ask for an outline that separates URL redirects, permissions, and rollback checks, using the supplied migration material. Then ask it to promise a traffic increase that the keyword export does not establish. Check that the first result cites the relevant documents and the second flags the unsupported promise instead of inventing a forecast. A fresh chat isolates conversation context, not permissions. Only then, with permission, verify the actual publishing channel and rendered page separately.

Keep SEO prompts tied to approved evidence with eesel CLI

Use a prompt to make the work specific, then use a teammate setup that remains visible to the people who own the sources and publishing decision. Try eesel when you want to review that workflow with your own approved content evidence.

Sources

• OpenAI prompt guidance
• Google Article structured data
• Google Search documentation updates and FAQ deprecation
• eesel CLI documentation