AI for SEO speeds up keyword clustering, content drafts, technical analysis and data interpretation, as long as the hard data comes from real SEO tools and a specialist checks the output. For me, it is no longer a time-saver on the side but the way I organise SEO work itself. At Websitepromotor and then at Space’M Online, I have used AI in almost every step of the SEO process, on client sites and on my own sites learnllm.nl and zzpgeldzaken.nl.
What has changed most is not which prompts I use, but where the checking happens. Some SEO tasks I have automated with AI, others I deliberately keep manual, and where the output gets you into trouble differs a great deal from task to task.
The trade-offs below come straight from day-to-day SEO work.
What is AI for SEO?
AI for SEO means using a language model to speed up parts of search engine optimisation: clustering keywords, drafting content, tracking down technical problems and interpreting data. The model does the repetitive work; the SEO specialist keeps the judgement, the strategy and the final responsibility.
How the work is split between model and specialist decides whether AI speeds up your SEO work or adds risk to it. Without a connection to other tools, a language model has no access to search volumes, rankings or backlink data, and it fills in plausible-looking figures when you ask for them.
What is the difference between AI for SEO and SEO for AI?
The difference between AI for SEO and SEO for AI lies in what you optimise. With AI for SEO, the language model is a tool you use to do your own SEO work faster. With SEO for AI, also called generative engine optimisation or GEO, you optimise a website so that AI search engines such as ChatGPT, Perplexity and Google’s AI Overviews pick up the content and cite it as a source.
Searchers mix up the two terms all the time. In Dutch search results, the query “SEO for AI” mostly returns pages about GEO, while “AI for SEO” is about using AI in the SEO process.
The disciplines partly overlap. Content that an AI search engine can easily cite has a direct answer under each heading, explicit facts with a version or date and a recognisable author. The same features also help a page in the regular search results, so if you use AI for SEO well, you quietly build towards both. Alongside classic SEO, I also work on GEO for my own site and for clients at Space’M Online, and in both disciplines those features are the foundation.
How do you use AI for keyword research?
To use AI for keyword research, let the model cluster and organise, and take the search volumes from a real keyword tool. Keyword research is the one step in my SEO process that I deliberately do not automate, because a model has no knowledge of what people actually search for. That information lives in search volume data and Search Console.
A recurring pattern I saw with ChatGPT: given a keyword, the model added a search volume unprompted, even though there was no connection to Ahrefs or Semrush. Those numbers are made up, and they sound plausible to anyone without keyword data of their own.
Clustering by intent, suggesting variants and grouping terms that belong on the same page are all faster and more thorough with a model than by hand. The order is therefore the reverse of what many guides suggest: data first, then the model. A prompt that forces the model to cluster by intent instead of producing a flat list:
I have a website about [topic]. Give 40 relevant keywords, grouped by search intent (informational, commercial, navigational, transactional). Do not add search volumes; I will get those from a separate tool.
During the period when I used ChatGPT intensively, I built separate custom GPTs for this: one for keyword research and one for classifying search intent. A pattern I ran into was that what worked well in a test prompt gave inconsistent output in daily use. A fixed format in the prompt and a check of the first result catch most of that.
If you want to build your own custom GPT for keyword research now, you are building on a feature with an end date: OpenAI is retiring custom GPTs on all plans on 11 December 2026, and when you migrate to a plugin, the GPT’s instructions become a skill.
How do you write SEO content with AI?
To write SEO content with AI, let the model produce a draft from a fixed brief and keep the editing and the facts in your own hands. AI speeds up writing SEO content, but it does not determine the quality.
Google’s January 2026 update showed how much that editing matters. On client sites and on benchmark sites I follow, content generated purely by AI without human input dropped, while content with a recognisable author and concrete examples held steady or grew.
How does a content pipeline with the Claude API work?
On learnllm.nl and zzpgeldzaken.nl, a content pipeline with the Claude API runs from keyword to brief to first draft, after which Python scripts publish the text automatically. Everything runs on my own VPS. The gain is in speed and consistency, not in skipping the editing.
A prompt that steers a draft instead of producing generic text:
Write a draft of an article about [topic] for [audience]. Use a businesslike, direct tone. Include these points: [3 to 5 key points]. Do not add a general introduction; start straight away with the answer to the search query.
Four things are hard to judge in advance with a pipeline like this:
- Rate limits during bulk processing, which make retry logic and batching a necessity rather than a luxury.
- Costs that turn out higher while building than in production, because you iterate far more in the test phase.
- Output via the API that differs from the same prompt in the interface. Since API version 2026-10-21, temperature, top_p and top_k have been removed and sending them returns an error; Anthropic advises steering the behaviour through the prompt.
- Prompts that respond differently after a model update, so a workflow that ran stably for months needs a new round of testing.
Because of that last point, how you steer the model, in other words your prompting, is ongoing maintenance rather than a one-off investment.
Why are figures better kept in a database than in the text?
Figures in AI content are better kept in a database, because a language model is good at phrasing and bad at remembering where a number came from. On zzpgeldzaken.nl the amounts are not in the text but in a database I set up with AI, and the articles pull them in through shortcodes.
Change an amount and it changes everywhere on the site at once. A comparison article with rates in the body text is correct on the day of publication and then goes a little more out of date every month. If you take the facts from a source you manage yourself and let the model write only the text around them, you also keep track of which source sits behind which number.
How do you analyse SEO data with AI?
To analyse SEO data with AI, bring all your sources together in one place and let the model answer questions about them. For learnllm.nl, zzpgeldzaken.nl and the client sites I manage at Space’M Online, I set up BigQuery as the central store, with Search Console, Google Analytics 4, Bing Webmaster Tools and Merchant Center side by side.
This makes two things possible that the separate interfaces do not allow. Questions that combine sources become a single query, for example which pages are gaining impressions but losing sessions. And the history keeps going: Search Console itself keeps about sixteen months of data, while data in BigQuery stays put.
Claude reads that data directly from BigQuery, so an analysis question no longer means exporting and pasting. If you have to export and paste first, you look once a month; if you can just ask a question, you look every week.
The pitfall lies in the convenience. A model that readily summarises your data will just as readily give you a summary that is wrong. So ask to see the underlying query rather than only the conclusion, so you can check which rows sit behind the answer.
How do you use AI for technical SEO?
To use AI for technical SEO, let the model summarise and prioritise large exports, and test the conclusion yourself. A crawl or a Search Console export contains more rows than you can review by hand; a language model spots patterns and proposes an order of work.
When analysing Screaming Frog exports, I use Anthropic’s Claude to combine crawl data with an assessment of the content. Any crawler recognises a 404; whether a page overlaps in content with three other pages is a judgement that a model can prepare. At Space’M, an integration is being built that loads crawls into BigQuery automatically, so a monthly technical check sits right next to the ranking data.
Two other technical tasks lend themselves well to AI. Structured data based on schema.org goes faster with a model, provided you validate the output before you publish it. A model can spot cannibalisation between your own pages in Search Console data, after which you decide which page should own the search query.
How do you capture recurring SEO work with AI?
To capture recurring SEO work with AI, build a skill for every task that comes up more than a few times: a documented way of working with a fixed input, a fixed output and a checkpoint. A one-off prompt is different every time; a skill does the same thing every time.
These are the four skills I use most:
- Working keywords into existing texts, so a page is added to rather than rewritten.
- Tracking content decay, so you can see which pages are slipping before the traffic has really gone.
- Spotting cannibalisation, with a link to the ranking data so the judgement rests on numbers.
- Assessing link building based on three Ahrefs exports: anchors, domains and backlinks.
I work with one Claude project per client, with the client context in separate files, and with Claude skills that work across them. The skill is the way of working, the project is the context; the same skill does the same thing for every client, with different results.
Scale is why capturing it matters. On learnllm.nl and zzpgeldzaken.nl I carry the risk of a mistake myself; on a client site someone else carries that risk, and a colleague has to pick up the same way of working without me being there. A prompt that only lives in my head works until the same analysis runs every month across several domains.
For a marketing agency, the same set-up is the step from separate prompts per employee to one way of working for the whole team, and the AI Literacy Course covers that step with an internal AI policy and checkpoints per task.
Which AI tools do you use for SEO?
AI tools for SEO fall into two groups: general-purpose language models that you steer yourself, and SEO software with its own index that has AI under the bonnet. Without a connection, a language model knows nothing about search volumes, backlink profiles or rankings, so for hard data you still depend on the second group.
The table below shows what each tool is good at in SEO work and how you can get it.
| Tool | Good for | Availability |
|---|---|---|
| ChatGPT | Clustering keyword ideas, rewriting texts, meta texts | Free basic version, paid for more |
| Claude | Analysing long documents, assessing content, scripts for SEO workflows | Free basic version, paid for more |
| Google Search Console | Real query data and positions for your own site | Free |
| BigQuery | Combining sources and keeping data beyond the Search Console retention period | Free up to a monthly processing limit, paid after that |
| Screaming Frog | Crawling and tracking down technical errors | Free up to 500 URLs, paid after that |
| Ahrefs or Semrush | Search volume, keyword difficulty, backlink analysis | Paid |
The strongest set-up combines the two groups. I take the hard data from Search Console, BigQuery and a keyword tool, and use a language model to interpret that data and build content around it.
Between the two groups sits a third type of tool that combines external SEO data and a language model in one environment. SEOptimate, a Dutch SEO platform, takes keyword and backlink data from DataForSEO, links your own Search Console data to it and reads in every page of your site via your sitemap. Based on the Google top 10, it suggests topics and internal links and writes blogs, service pages and product texts. It publishes those texts according to a content calendar to platforms including WordPress, Shopify and Lightspeed. Through its own MCP server you can also control SEOptimate directly from Claude; add an MCP connection to your CMS and Claude carries out the changes on both sides in a single conversation. The editing stays with you: a generated article is a first version, not a final one.
How do you use AI for local SEO for service businesses?
To use AI for local SEO for service businesses, use the model to scale up local content and to keep business details consistent. What a model cannot supply is local authority itself.
Three tasks within local SEO lend themselves to AI:
- Drafting variants of local landing pages, which you then fill in for each location with real information.
- Comparing business details such as name, address and phone number across the website, the Google Business Profile and listings elsewhere.
- Summarising customer reviews into recurring signals about what customers value and what they complain about.
Building local authority remains a matter of real presence, consistent listings and relevant content. A local service business is more likely to win with a few well-maintained pages than with dozens of thin variants filled by AI, and it was exactly that kind of me-too content that the December 2025 core update hit.
For small service businesses, local SEO is often the first place where AI visibly saves time.
Where does AI fall short in SEO?
AI falls short in SEO wherever the work calls for judgement rather than production. A language model recognises patterns and generates text, but it does not weigh commercial interests, has no knowledge of your brand and, without a connection, has no access to ranking data.
Made-up figures are the best-known limitation. A model fills in search volumes, positions and statistics plausibly and without a source, the same pattern as an AI hallucination in other contexts. In technical work, the equivalent is made-up code: Python functions that do not exist and WordPress hooks that are documented nowhere, which only comes to light when you run the code.
Two quieter failure patterns cost more time in practice. With complex prompts that contain several conditions, a model follows some of the instructions and ignores the rest without saying so. A summary of a long document sometimes sounds convincing yet differs from the source, which you only discover by going back to the original.
Strategy is not something a model decides for you. Which keyword fits your proposition, which page deserves priority and which content gap is commercially worth pursuing remain decisions for the specialist.
The last limitation is over-reliance. Teams that use AI without checks quietly build up a stock of thin pages, and when employees use tools independently of each other without agreements, you get shadow AI that is hard to keep track of.
How do you use AI responsibly in SEO?
To use AI responsibly in SEO, build a checkpoint into every task before the output moves on. The starting point is that AI is never responsible; you are. A model does not check itself well enough, and mistakes, made-up sources and plausible-sounding nonsense slip through until someone explicitly points them out.
From my years at major banks, I bring the four-eyes principle into AI work: nothing goes out without a second check. In practice that sometimes means using Claude as a reviewer of ChatGPT output, or deliberately stepping back before you publish.
How thorough the check needs to be depends on the risk of the task. An internal note needs less than a published article, and client advice needs more than both. My own way of working has four fixed checks:
- A test run on a small sample before bulk processing starts.
- Format validation for output that a script processes further.
- Human final editing as the last step before publication.
- Logging prompts and results, so you can trace later what happened.
I will be honest about one gap, because I see it in many professionals too: I still do too little structured fact-checking of sources, figures and specific claims.
AI for SEO delivers the most when it becomes a shared way of working instead of a trick each employee uses on their own. For agencies and service businesses that want their team to work with the same checked approach, AI literacy for teams gives everyone the same foundation, with a separate account for each employee.
Frequently asked questions about AI for SEO
AI for SEO also raises questions about how Google treats AI content and about AI crawlers’ access to your site.
Does Google penalise content written with AI?
No, Google does not penalise content because it was written with AI, but generating many pages without added value for users can fall under the spam policy on scaled content abuse. According to Google Search Central, AI content has to meet the same Search Essentials and spam policies as other content. Google asks you to check facts manually before you publish, including in titles, meta descriptions, structured data and alt texts.
Do you have to disclose that SEO content was made with AI?
Google does not mention an obligation to label SEO content made with AI, but it does recommend explaining how content was created, including the role of automation, in a way that suits your audience. For images made with AI, Google advises adding metadata. In Merchant Center, AI images need the IPTC value TrainedAlgorithmicMedia, and you have to submit AI-generated product titles and product descriptions separately and mark them as AI-generated.
How do you control in robots.txt whether ChatGPT uses your site?
In robots.txt you set what ChatGPT may do with your site for each OpenAI crawler: GPTBot collects content for training and OAI-SearchBot makes your site appear in ChatGPT’s search feature. These settings are independent of each other, so you can block training and still allow visibility in ChatGPT. According to OpenAI, a change for search takes about 24 hours to take effect. With Google, you control use for Gemini models with Google-Extended, which according to Google has no effect on Google Search.
