AI for SEO: 10 practical uses, with a prompt for each

Using AI for SEO means handing an assistant like ChatGPT, Claude or Gemini the high-volume tasks of search engine optimization (clustering keywords, preparing briefs, writing tags, reading logs, commenting on numbers) while you keep the trade-offs and the material only you have. AI replaces neither your data nor your judgment: it processes faster what you know how to check.

"AI SEO" actually covers two topics. The first, covered here, is using AI to do your SEO. The second is getting cited by AI tools in their answers: that is GEO, which I summarize at the end and cover in detail in my GEO SEO guide.

AI for SEO: what AI does well, and what it does badly

The question is not whether AI helps with SEO: it clearly does. The question is what it helps with, and with what safeguard.

What it does better than a human

• Processing volume. Cross-referencing three exports of several thousand rows, clustering two thousand queries by intent, comparing hundreds of pages. The machine doesn't get tired and doesn't skip a row.
• Repetitive, well-framed production. Two hundred title tags following a precise rule, alt text, structured data markup.
• A first read of a body of content. Understanding what the pages already ranking for a query cover, and what they don't.

What it does less well than it seems

• Trade-offs. Choosing between two topics, deciding whether a page deserves a rewrite or deletion. These decisions depend on your margin, your production capacity and your goals, three things AI knows nothing about.
• Numbers. A model makes up search volumes with perfect confidence.
• Experience-based content. A client project debrief, a figure from your own customers, a reasoned opinion: that is what search engines look for, and what the machine cannot produce.

10 practical ways to use AI for SEO, with a prompt for each

These are the uses I rely on every week. Each prompt leaves in square brackets what you need to provide: that material is what makes the result good. They work in ChatGPT, Claude, Gemini or Le Chat, including the free versions as long as the files stay small.

1. Keyword research: cluster and classify

AI doesn't know search volumes. Export your queries from Search Console, Google Ads Keyword Planner or your SEO tool, then ask for a clustering by intent, with one target page per cluster. What used to take me two days in a spreadsheet now takes two hours, most of it spent reviewing the doubtful clusters.

Here is an export of [number] queries with their monthly volume, from [tool]. My website: [business, audience, area].
Cluster these queries by search intent (learn, compare, buy, local). For each cluster, suggest a single target page: an existing page from the list below or a page to create. Do not change any volume and do not make any up.
[Paste the export]
[Paste your list of URLs]

2. Analyze the SERP and search intent

Before writing, I collect the top ten results for the target query and ask what they all cover, what a minority covers, and what none of them covers. The third category is the most valuable: that is where the open angle is.

Target query: [keyword]. Here are the titles, subheadings and first paragraphs of the top 10 Google results for this query: [paste them].
Tell me: 1) the dominant intent and the expected format (guide, list, comparison, tool, definition); 2) the topics covered by every page; 3) those covered by a minority; 4) the questions none of them covers. End with the angle I could take.

3. Prepare a content brief

The brief turns the analysis into instructions. It works just as well for a writer as for the AI itself at the next step. My seven-section template is detailed in my article on the SEO content brief.

Based on the previous analysis, write the brief for an article targeting [keyword] for [audience]. Sections: intent, angle, H2/H3 outline, questions to answer, evidence to provide (figures, examples, screenshots), approximate length, internal links to [list of my pages], what not to write.

4. Writing: first draft, then rewrite

This is the use everyone talks about, and the one with the smallest gain. My method has four stages: I prepare the material (project notes, real figures, objections I've heard), I have the AI produce an outline and a first draft section by section, I systematically rewrite, then I check every factual claim. The time split is roughly one third generation for two thirds preparation and rewriting.

Here are the brief, my field notes and two articles I wrote, for the style: [paste them].
Write only the section "[H2 title]", 300 words maximum. Use only the facts in my notes; when a piece of information is missing, write [TO COMPLETE] instead of making it up.

5. Refresh content that has stalled

Before touching the text, establish the cause in Search Console. Rankings and impressions dropping: competitors have improved, you need to enrich the page. Good rankings on a query nobody types anymore: the intent has changed, you need to reposition. Stable impressions and falling clicks: an AI summary answers in your place, you need to change format. AI comes in afterward to list what is missing, not to judge the text.

My article is stuck at position [n] for [query]. Here is its text and the text of the 3 pages ranking above it: [paste them].
List what my article is missing to meet the search intent: topics, questions, formats, evidence. Rank the gaps by likely impact. Do not rewrite the article.

6. Title tags, meta descriptions and alt text

On a catalog of several hundred pages, a human gets tired and inconsistent; the machine doesn't. I provide the rule and five approved examples, then I check a sample of the output. Count the characters yourself: language models are bad at counting characters.

Here are [number] pages with their URL, H1 and main query: [paste the table].
For each page, suggest a title tag of 60 characters maximum that starts with the query, and a meta description of 120 to 155 characters with a concrete benefit. Return a table: URL, title, meta. Use these approved examples as a model: [3 to 5 examples].

7. Generate structured data

JSON-LD markup is mechanical work, ideal to delegate, on one condition: only mark up what is visible on the page. Then validate it with Google's Rich Results Test.

Here is the visible content of my page [article, product page, local business]: [paste the text].
Generate the appropriate schema.org JSON-LD markup. Only include information present in this text. Flag the recommended properties my page doesn't let me fill in.

8. Technical SEO: crawls, logs and redirects

This is where the gap is most striking, because without AI this work often simply didn't get done. Screaming Frog lets you connect the OpenAI, Gemini or Anthropic APIs, or a local model through Ollama, and run prompts on every crawled page: alt text, classifying pages by intent, spotting near-duplicate pages through embeddings (handy for a redirect plan). I cover the reverse connection, Claude driving your crawls, in my article on the Screaming Frog MCP.

For files too large for a conversation (server logs, exports of hundreds of thousands of rows), I use Claude Code, which works on files directly on your computer. The step-by-step guide to installing Claude Code has its own article.

Here are my server logs for the last 30 days ([Apache or Nginx] format), with IP addresses anonymized.
Isolate visits from Googlebot, OAI-SearchBot, PerplexityBot and Claude-SearchBot. For each: number of requests, 20 most visited pages, share of 3xx, 4xx and 5xx responses, directories never visited. Return one table per bot.

9. Internal linking

AI suggests, you validate the business logic. Never let links be inserted without review: a poorly chosen anchor or a link to a secondary page costs more than it brings.

Here is the list of my pages (URL, title, main query) and the text of my new page: [paste them].
Suggest 5 links from existing pages to the new page, and 3 links from the new page to existing pages. For each link: source page, sentence to insert it in, suggested anchor. Descriptive, varied anchors.

10. Data analysis and reporting

Pulling together the month's numbers, spotting notable changes and drafting a first commentary: AI does it fast, but explaining why stays on your side. Search Console is also testing, with some of its users, a way to configure the Performance report in natural language: you describe the analysis you want and the tool sets the filters and comparisons, within a limit of 20 requests a day. Always check the filters it applies.

Here are two Search Console exports (pages and queries), for month [M] and month [M-1]: [attach them].
Find the 10 biggest changes in clicks. Separate those caused by a change in position from those caused by a drop in click-through rate at a stable position. Write a 10-line commentary for a business owner, without jargon, and do not suggest any cause the data doesn't show.

More prompts dedicated to SEO, and how to run them in Claude, are gathered in Claude for SEO.

Which SEO tasks should you hand over to AI?

The table below sorts twenty SEO tasks by what I actually delegate, what I keep, and what takes both. Filter, then copy the list for your team.

How much time AI saves in SEO

Productivity promises in this field are often fanciful. Here is my estimate, based on my own work, after two years of daily practice. These are orders of magnitude, not a scientific measurement.

Note the contrast: writing, the use everyone talks about, is where the gain is smallest. And data processing, which nobody talks about, is where it is massive.

AI SEO tools: free and paid

The question of the ideal AI SEO tool comes up constantly. My answer often disappoints because it is not very spectacular: a general-purpose assistant fed with your own exports, plus the data tools you already use.

Doing SEO with free AI tools

It is possible for a small website. Search Console gives you your real queries, Google Ads Keyword Planner gives orders of magnitude for volume, the free version of ChatGPT or Claude handles exports of reasonable size, and Screaming Frog crawls up to 500 URLs for free. The limits come with volume: number of messages, size of attached files, access to the most advanced models. Before pasting a client export or server logs into a third-party service, check what your account allows and anonymize what can be anonymized.

What Google says about AI-generated content

This is the question that worries people most. Google's position has been stable since 2023, and it is documented.

• How content is produced doesn't matter; its value does. In February 2023, Google wrote on its Search Central blog that appropriate use of AI or automation is not against its guidelines.
• Mass production without value is spam. Google's spam policies describe scaled content abuse: generating many pages with the main purpose of manipulating rankings. The first example given is using generative AI tools to create many pages without adding value for users.
• Verification is mandatory. The Search Central page on AI-generated content, updated in October 2026, considers AI useful for exploring a topic and structuring original content. It points out that models make mistakes (hallucinations) and that you must manually check the accuracy of any generated content before publishing it, metadata included.
• Transparency is recommended. Google recommends explaining how a piece of content was created, when that is relevant to your readers. In e-commerce, generated images must carry the IPTC DigitalSourceType metadata, and generated product titles or descriptions must be labeled as such.
• Lived experience carries weight. Since December 2022, Google's quality rater guidelines have added experience to the criteria of expertise, authoritativeness and trustworthiness (E-E-A-T). These raters don't directly affect rankings, but their guidelines say what Google is trying to reward.

So the real risk is not a penalty; it is more insidious. A text generated with no personal input reproduces the average of what already exists. It isn't bad, it is interchangeable, and interchangeable content has no reason to rank above the twenty others that say the same thing.

The limits of AI in SEO

• Made-up numbers. Volumes, rankings, statistics: anything that isn't in your files should be treated as false until verified.
• Time-sensitive facts. Prices, features, update dates: a model's knowledge stops at a certain date, and it fills the gaps with confidence.
• The web's average. Without your notes, your examples and your angle, the result looks like everything that already exists.
• Confidentiality. Pasting a client export or logs means sending data to a third-party service. Check your account type and your company's rules.
• Inconsistency. The same prompt never gives the same result twice. Keep approved examples and check a sample of every batch.
• The cost of reviewing. A fast first draft that needs an hour of corrections hasn't saved you anything. Measure the total time, not the generation time.

What AI changes for search

The other side of SEO and artificial intelligence isn't about your tools but about your visitors: part of search now goes through engines that answer directly. In France, Google launched AI Overviews and AI Mode on July 22, 2026, and ChatGPT, Perplexity or Claude cite their sources when they search the web. The goal is no longer just to rank, but to be cited.

The good news: these engines draw their sources from search indexes, and Google states that SEO best practices remain relevant for its AI features. The method to get cited (answers at the top of the page, dated evidence, allowed crawlers, brand mentions) is in my GEO SEO guide, and the point-by-point comparison in GEO vs SEO: the real differences. To measure it, Search Console has offered since August 31, 2026 a generative AI performance report, which counts your links' impressions in AI Overviews and AI Mode.

Where to start

1. This month: measure your starting point, rankings and time spent on each SEO task. Without a baseline, no gain can be demonstrated.
2. Next month: delegate three mechanical tasks from the table above, with a prompt written once and for all and a sample-based check.
3. The third month: reinvest the time saved in what AI can't do: your data, your case studies, your lessons learned.

AI shouldn't be used to produce more of the same, but to free up time to publish what nobody else can write. When a task comes back every week with the same files, it can become a program that runs on its own: that is what my AI automation projects are about. And for a full diagnosis of your SEO, see my SEO audit.

Written by

Anthony Courtin

SEO, GEO & AI automation consultant

I help SMBs and independent professionals with their visibility (Google, ChatGPT, Perplexity) and with real-world use of AI at work. I work with companies on site to train their teams and rethink their processes. Everything I write here, I use with my clients.

Follow me on LinkedIn My YouTube channel (in French)