25 SEO prompts for AI agents using real search data
Copy 25 practical prompts for ChatGPT, Claude and other AI agents to analyze Search Console clicks, CTR, indexing, GA4 and Bing with evidence.
GSC Magic team · · 7 min read These 25 prompts work best when your AI assistant has authorized access to the data being requested. Specify the property, a complete date window and output columns. Search Console, GA4 and Bing are separate data sources: a connector for one does not automatically provide the others.
Before you paste a prompt
A powerful SEO prompt has four ingredients: data source, date window, segment and decision. Rather than asking “How is my SEO?”, ask your agent to compare the latest 28 complete days against the prior 28, list the ten biggest losses by page and explain what to investigate first.
The assistant should cite actual rows from the connected service. If access is missing, a useful response says so instead of inventing performance data.
To connect search data, read our setup guides for Claude or ChatGPT. If GSC Magic access is not yet available, join the early-access list.
Quick-win prompts (Search Console)
1. Find non-branded queries with at least 200 impressions over the last 28 complete days, average position 4–20 and below-median CTR for their position band. Include URL, clicks, impressions, CTR and position.
2. Which pages gained impressions but lost CTR versus the prior 28 days? Separate ranking shifts from snippet hypotheses.
3. Show pages ranking near the first page for commercially relevant queries. Explain why each candidate is worth refreshing.
4. Find high-impression queries with very few clicks, excluding obvious navigational terms.
5. Select five pages for a title-and-description review, and draft accurate snippet alternatives based on each page's content and query intent.
How to use the output: choose one page at a time. Compare queries, search intent and positions before rewriting copy. High impressions alone do not prove a title problem. See the full low-CTR workflow.
Traffic drop prompts (Search Console)
6. Compare total Google Search clicks for the last 28 complete days and the previous 28. Rank pages by absolute clicks lost, not percentage alone.
7. For the five biggest losing pages, break changes down by query, country and device.
8. Graph weekly impressions, clicks, CTR and average position for the most affected pages. Identify when the decline started.
9. Check whether Google indexing signals or known crawl issues could explain the decline; separate confirmed findings from guesses.
10. List the queries responsible for most of the lost clicks and label whether they reflect lower impressions, CTR or both.
Traffic drops have several possible causes: seasonality, demand, changes in search appearance, technical failures and competition. A useful assistant should test hypotheses, not choose a story based on one chart. Work through the organic traffic drop checklist.
Indexing prompts (Search Console inspection access)
11. Inspect these specified new URLs and summarize their reported Google index status. Include the inspection date and canonical URL when returned.
12. For these ten URLs, group reported indexing issues by status and propose what to verify on the live pages.
13. List the submitted sitemaps for this property and surface any reported processing errors.
14. Identify important URLs with no Search Console impressions in the chosen period, but do not assume that zero impressions means they are unindexed.
15. For URLs with recently lost visibility, suggest which ones warrant a URL Inspection check.
Important limitation: Google’s URL Inspection API returns indexed-version status; it is not a general API to request indexing. Google’s Indexing API only applies to specified job posting or livestream page types.
Content and intent prompts (Search Console)
16. Find queries served by multiple URLs and flag only cases that might share the same search intent; do not recommend automatic redirects.
17. Identify pages whose impressions remained stable but clicks declined for multiple complete reporting periods.
18. Cluster the highest-impression informational queries by question and suggest gaps that existing pages do not fully answer.
19. Find queries where a less relevant page attracts more impressions than the intended target page. Show evidence before suggesting fixes.
20. Recommend internal links among my existing pages based on related queries and intent, but verify each linked page is actually relevant.
Use our cannibalization diagnostic before merging pages. A shared keyword is not enough evidence to delete content.
Reporting and advanced prompts (additional data required)
21. [GSC] Create a monthly summary of clicks, impressions, CTR and position, with the top changes by landing page and a clear definition of each date range.
22. [GA4] How many sessions were attributed to common AI assistant referrers last month? Group by session source and disclose attribution limitations.
23. [GA4] Which landing pages from organic search produced the most configured key events? Include session counts.
24. [GSC] Compare branded and non-branded search query groups using an explicit, reviewable brand-term definition.
25. [GSC + GA4 + Bing if connected] Recommend three actions for next week, ranked by impact, effort and strength of evidence. Use only sources currently available.
The labels in brackets tell you what must be connected. Without GA4 access, an agent cannot reliably answer prompt 23 from Search Console alone; without Bing data, it cannot compare search engines.
Turn a generic prompt into a reliable report
The same SEO question can produce very different answers depending on the information you provide. Compare these two requests:
Vague: “Find quick SEO wins.”
Auditable: “For the verified Search Console property example.com, analyze Google Web search over the latest 28 complete days. Show non-branded queries with at least 200 impressions and average position 4–20. Return the matching page, clicks, impressions, CTR, position, date range and one testable hypothesis. If you cannot query the data, tell me.”
The second prompt defines a source, filter, date range, output schema and acceptable uncertainty. It is harder for an assistant to hide missing data behind generic recommendations.
Use this adaptable template:
Data source: [GSC / GA4 / Bing]
Property or account: [exact identifier]
Reporting period: [start and end date; final data only]
Segment: [Google Web, country, device, page group]
Question: [one decision I need to make]
Output: [table columns and ranking method]
Evidence: [actual numbers and links/URLs]
Limitations: [missing permissions, partial data, hypotheses]
Next action: [one change to test and one success metric]You can paste it into Claude, ChatGPT or another agent. It is a prompt format, not a substitute for connecting the data.
An end-to-end example: deciding which page to refresh
Suppose you run prompt 1 and the assistant returns three candidate URLs (all numbers below are fictional):
| URL | Impressions | Clicks | CTR | Avg. position |
|---|---|---|---|---|
| /guides/api | 8,000 | 160 | 2.0% | 8.1 |
| /guides/claude | 3,500 | 135 | 3.9% | 5.4 |
| /blog/news | 1,900 | 11 | 0.6% | 17.8 |
Do not simply edit the title of the page with the lowest CTR. The news page is much farther down the search results on average, so position is likely part of the low-click story.
A more defensible workflow:
- Select the API guide as an initial candidate because it has substantial impressions and sits in a position range where relevance and snippet work could matter.
- Pull its top queries and check if one broad query is skewing the page-level average.
- Inspect the actual content and the competing results for those queries.
- Improve the section that fails to answer the dominant intent, or test a more accurate title if the content already matches.
- Save the original metrics, edit date and follow-up period.
- Report the outcome with changes in position, impressions and clicks—not just CTR.
The low-CTR guide gives the full process.
How to check whether an AI recommendation is trustworthy
Before implementing a suggested SEO change, ask:
- Did the assistant actually call a data tool, or only reason from general knowledge?
- Does the output identify the property and complete reporting dates?
- Are the numbers raw counts, calculated differences or estimates?
- Are the page URLs real and verifiable?
- Would the suggestion cause irreversible changes, such as redirects or deletions?
- Which other source is needed to support a conversion claim?
Use the labels Observed, Hypothesis, Recommended test and Expected signal. This simple distinction makes AI-generated reports easier to review.
Frequently asked questions
Do these prompts work without a paid SEO tool?
The reasoning and templates do, but a prompt cannot fetch private Search Console or GA4 metrics without authorized access. You can also analyze a manually exported report, as long as its age and scope are clear.
Which prompt should a new site use first?
Start with available impressions and indexed-page status. A site with almost no search visibility cannot reliably optimize its CTR by chasing tiny samples.
Can an AI agent directly improve Google rankings?
An agent can help analyze evidence and propose or implement changes within granted permissions. It cannot guarantee rankings or override how Google indexes and ranks pages.
Should I run all 25 prompts every day?
Usually not. Start with a small weekly review of meaningful changes, then investigate the few pages and queries that account for most of the opportunity or risk.
The follow-up prompt that improves every report
For each recommendation, show the underlying metric and date range, tell me what is observed versus inferred, estimate implementation effort, and define one measurable success criterion. If data is missing, say exactly which source is needed.
This turns a generic chatbot answer into a reviewable workflow. Keep the original baseline so you can measure what changed after implementation.
Want to ask these questions using an MCP connection rather than repeated CSV uploads? See what GSC Magic is building.
Join the waitlist
Coming soon. Get 30% off Pro for 6 months when you join.