AI can be very good at the slow, structured, data-heavy parts of SEO, and pretty bad at the parts that make an actual website or a page worth ranking.
Now, I use Claude and other LLMs every day in my SEO work, mostly connected to live data from Ahrefs and Search Console and to the WordPress sites I work on. I also connect a handful of other tools and MCPs, designed to make these tools operate better.
I also co-founded Linkter, an AI internal linking tool that placed over 10,000 links on WordPress sites before we shut it down in 2025.
So I want you to know I am not an AI skeptic, I am just someone who’s seen AI save hours on some jobs but also someone who saw AI make a mess of so many others.
The short version
- I use AI for research, data analysis, technical work, internal linking and content audits. That’s where it saves the most time with the least risk.
- An AI model on its own has no search volume data and no idea what ranks today. Connect it to real data (Ahrefs, Semrush or Search Console) or feed it exports.
- It will make up numbers, sources and quotes with total confidence. Check anything factual before it goes anywhere near a client or a page.
- Keep the writing human. AI help with a draft doesn’t hurt rankings, but AI replacing the writer produces the same article everyone else can generate, and that content will perform badly.
- I mostly use Claude, but ChatGPT, Gemini and other LLMs can do the same jobs, so use whichever you already pay for.
- AI isn’t replacing SEO. It might help replace the busywork inside SEO, and it’s taking clicks away from pages that only repeat common knowledge.
Where AI earns its keep
The pattern behind every good use of AI in SEO is the same. You give it data or a clear task with a checkable answer, it does the tedious part in seconds, and you judge the result because you know what a good result actually looks like.
When the answer can’t be checked, or when the job needs taste and experience, AI gets a lot less useful.
Research with live data
Ask a chatbot for the search volume of a keyword and it will give you a number, every single number.
…and that number is made up. No language model has search volume data built in, and even the ones that can browse the web can’t see Ahrefs or Semrush data unless you connect them to it.
What’s changed recently is that you can. Ahrefs, Semrush and even Mangools now offer MCP servers, which are connectors that let an AI assistant query their data directly. But that takes manual connection, it’s not built into every AI chatbot by default.
For example, I have Claude connected to Ahrefs and to SEO Gets, a Search Console tool with its own MCP server, so I can ask for things like “which of this site’s pages lost the most clicks in the last three months, and which queries did they lose” and get an answer built from real data instead of a guess.
This is the use that saves the most time, in my experience. Pulling a keyword list, checking who ranks, comparing a site’s top pages with a competitor’s, all of that used to mean a lot of clicking and exporting.
Now it’s one question and a table, and I spend the saved time deciding what to do about it and how to approach it.
If you don’t have a tool with an MCP server, exports work almost as well. Download the data as a CSV, upload it and ask your question.
Analyzing the data you already have
Most sites are sitting on more Search Console data than anyone has time to read.
AI is pretty solid at reading it. Export your queries and pages, and it can find the pages ranking on page two that are closest to page one, the queries where two of your pages compete with each other, the posts that are losing clicks, and the questions people search for that no page on your site answers.
It’s also good at writing the small, fiddly things that make analysis faster.
Search Console’s Performance report supports regex filters, and almost nobody remembers regex syntax (or maybe that’s just me, I don’t know).
Ask for a filter that matches question searches and you get something like ^(how|what|why|can|does|is)\s, which you paste into the query filter to see every question people find you for.
The same goes for Google Sheets formulas, Looker Studio calculated fields and short Python scripts to merge or clean exports. None of these are hard, they’re just slow, and slow is where AI helps most.
Technical and site work
A lot of technical SEO is writing small pieces of code correctly. Schema markup, redirect rules, robots.txt changes, XPath for scraping data in Screaming Frog, fixing a broken canonical in a theme template. AI does all of these well, as long as you check the output before it goes live.
Redirect mapping is a great example. Give it the old URLs with their titles and the new URLs with their titles, and it matches them by meaning instead of by slug, which is exactly the tedious part of a migration.
It gets most of them right, flags the ones it’s unsure about (if you ask it to), and you review the rest.
Screaming Frog has also added an AI integration that runs your own prompts on every page it crawls, with OpenAI, Gemini, Anthropic or a local model.
You can use it to flag thin pages, suggest alt text for images that have none or pull the main topic of each page into a column, across a whole site in one crawl.
And I now connect Claude straight to WordPress sites through an MCP server, so it can make changes directly to the site. That’s powerful but it’s also the easiest way to break a bunch of things, so only do it with backups in place and review every change before it goes live.
The checks in my technical SEO checklist are exactly the kind of work that goes faster this way.
Internal linking
Internal linking is the use case I know best, because it’s what Linkter did. Finding which pages on a site should link to each other is a semantic matching problem, and AI is very good at matching pages by meaning, even when they don’t share any direct keywords.
What it’s not good at is deciding which pages matter. Left alone, it links everything to everything, and the page you need to rank gets the same number of links as a three-year-old news post.
So the split I’d use is simple. You decide which pages are priorities and where authority should flow, and AI finds the relevant pages and the natural places to add the links. My internal linking guide covers the priority side.
Content audits
Sorting a few hundred URLs into keep, update, merge or remove used to take days. Give AI a crawl export with titles, word counts, clicks and impressions, and it can do a first pass in minutes, with a reason for each decision.
It won’t be right every time, but reviewing a finished first pass is much faster than starting from a blank spreadsheet.
It’s also good at spotting pages that cover the same topic, which is where cannibalization comes from. Screaming Frog can now do this with vector embeddings during a crawl as well, which finds pages that are similar in meaning rather than just in wording.
Helping the writer, not replacing them
This is where I draw the line. AI is useful around the writing. It can summarize what the top-ranking pages cover and what they all miss, pull the questions people ask on a topic, check a finished draft for claims that need a source, and point out sentences that don’t make sense.
The writing itself is the part I’d keep human.
Prompts I’d start with
These work in Claude, ChatGPT or Gemini. Swap in your own data, and give the AI the actual export or text rather than asking it to go find things.
Here’s a Search Console export of queries and pages for the last three months. Find queries where the page ranks between position 8 and 20 with at least 100 impressions, group them by page, and tell me which pages are closest to page one and what each one is missing.
Write an RE2 regex for the Search Console query filter that matches searches starting with how, what, why, can, does or is.
Here are two lists, the old URLs with their titles and the new URLs with their titles. Match each old URL to the closest new one, give each match a confidence score, and list every old URL that has no good match instead of forcing one.
Write LocalBusiness JSON-LD for this business using only the details below. Don’t add any property I didn’t give you.
Here’s a crawl export with URL, title, word count, clicks and impressions for the last 12 months. Put each URL into keep, update, merge or remove, with one sentence of reasoning for each.
Here are the top five ranking pages for this keyword. List what all of them cover, what only one or two of them cover, and what none of them cover.
Read this draft and list every statistic, date, name or factual claim that needs a source. Don’t rewrite anything.
Notice what the good ones have in common. Each one hands over the data, asks for something specific, and tells the AI what not to do.
The last instruction matters more than people think, because the default behavior of every model is to fill gaps with something plausible.
Where AI falls over
The failures are just as predictable as the wins, and most of them come from the same place. Language models are built to produce a plausible answer and not necessarily a true one.
It makes up numbers and sources
This is the big one. Ask for a statistic and you’ll often get a precise-looking number with a source that doesn’t exist, or a real source that never said it.
The clearest test of this I’ve seen is the Tow Center’s study for the Columbia Journalism Review, which asked eight AI search tools to identify the source of real news excerpts. Collectively they gave incorrect answers to more than 60% of queries, from 37% wrong for Perplexity to 94% for Grok 3, and they rarely admitted uncertainty. These were tools with live web access, looking at text that exists.
So every number, date, quote and source in anything AI touches gets checked by a person. That includes the data analysis above, which is why I’d always want the underlying export open next to the answer.
It doesn’t know what ranks today
Models are trained on data with a cutoff, and search results change every week. A model working from memory will tell you what ranked when it was trained, describe Google features that have since changed and recommend tactics that stopped working. It will do all of it confidently.
The fix is the same as for search volume. Give it the live results, or connect it to a tool that has them, and treat anything it says about “how Google works” from memory as a starting point to verify.
It writes the same article everyone else gets
Ask any model to write about any topic and you get a competent summary of what’s already ranking. That’s the problem.
Google’s own guide to AI search now tells site owners to create “non-commodity content” and to avoid recycling “what others on the internet have already said, or could easily be produced by a generative AI model.”
And publishing a lot of it can get you in trouble. Google’s scaled content abuse policy covers large amounts of unoriginal content made to manipulate rankings “no matter how it’s created,” and it names using generative AI to create many pages without adding value as an example.
It doesn’t know your business
AI might be able to tell you what’s wrong with a site, but it can’t tell you which fix is worth doing first, because that depends on your margins, your best customers, what your competitors are doing that isn’t visible in any tool, and what you can realistically ship this month.
Prioritization is most of the value in SEO work, and it’s the part AI is weakest at.
Why the writing should stay human
My advice is to use AI everywhere around the content and keep a person writing it. And no, this is not because Google punishes AI text, but because the things that make a page rank and get cited now are the things AI can’t produce. Your own results, your data, your opinions, the mistakes you’ve made and what you’d do differently.
This is also what users want to see. They have no reason to read something you haven’t invested any time creating. So they’ll bounce off your site, you won’t make a sale, and Google will notice users not having a preference for what you are putting out, and will properly derank your website.
I go deeper on this in my guide on writing SEO content that ranks, including where AI fits in that process without taking it over.
Will AI replace SEO?
AI isn’t replacing SEO, but it’s slightly changing what SEO pays off for because the point is not only in ranking, but getting mentioned in conversations that people have with AI chatbots. That doesn’t make SEO work that much different, because AI Overviews and AI Mode are rooted in core Search ranking and quality systems and pull their sources from the normal search index.
What people now call AEO or GEO (answer engine and generative engine optimization) is therefore still SEO.
ChatGPT, Perplexity and other LLMs work the same way in principle, since they search the web and cite pages they find there.
What has changed is the number of clicks. Ahrefs re-ran its study of 300,000 keywords with data from December 2025 and found that an AI Overview now correlates with a 58% lower clickthrough rate for the top-ranking page.
The pages losing the most are the ones that only repeat common knowledge, which is exactly what AI summaries are good at.
So the jobs that are going away are the mechanical ones, like pulling reports, writing regex, mapping redirects and summarizing what ranks. Deciding what’s worth doing, understanding a business, and creating something an AI summary can’t replace are becoming a bigger share of the work, and I don’t see a model doing those well any time soon.
If you want more practical guides like this one, Built to Rank is my occasional newsletter on what’s actually working in SEO. It only goes out when I have something worth sending.
A couple more questions
Which AI is best for SEO?
I use Claude for most of my SEO work, but ChatGPT and Gemini can do everything in this guide too. The model matters less than what you connect it to and how clearly you ask.
Can ChatGPT do SEO?
It can do a lot of SEO tasks, like analyzing exports, writing schema and regex, mapping redirects and sorting pages in an audit. It can’t do SEO on its own.
Is AI content bad for SEO?
Google doesn’t penalize content for being made with AI, and most top-ranking pages have some AI involvement. The problem is content that’s mostly generated with nothing original added. It tends to rank worse, and publishing it at scale can fall under Google’s spam policies.
Do I need an AI SEO tool?
Probably not a separate one. The big SEO tools have added AI features and MCP connections, and a general assistant like Claude or ChatGPT connected to your existing data covers most of what dedicated AI SEO tools sell.