What is LLM SEO, what steps can you take to optimize your website for LLMs, and is there a way to automate the process?
Key Takeaways
- Only 37.9% of URLs cited in Google's AI Overviews also rank in the top 10 for that query, down from 76% seven months earlier. Ranking and citation have come apart, which is the whole reason llm seo* is now a separate discipline from SEO.
- A peer-reviewed reranking study found that adding a fresher date to a passage moved it up by as many as 95 ranks and flipped model preference up to 25% of the time between two passages of identical relevance. The date did the work on its own.
- No standard marketing tool reports this. Search Console shows clicks and impressions, rank trackers show rank, and a page that keeps its position while losing its citation looks completely healthy in both.
- Fimo runs agents against the published site that watch traffic sources, refresh ageing pages and keep schema and llms.txt current, each opening a pull request a person approves before anything ships.
You might have noticed that SEO has, shall we say, changed.
Not too long ago, you only had to worry about how Google viewed your content. If you understood Google’s ranking algorithm, you could write great content that ticked all the boxes required to rank on page one.
Today? LLM SEO and Google AI Overviews have entered the chat. Together, these two search and discovery tools have fundamentally changed how humans and LLM crawlers find websites, products, and services.
And that’s why LLM SEO is now a very real thing.
Ranking and citation have come apart
Ranking on Google means climbing into page one of relevant search results. Chances of a click are fairly high on page one.
Getting cited by an LLM like ChatGPT means that ChatGPT references your content, product, or service in response to somebody’s question. Chances of a click here are low unless the person is deep in research mode, but there’s value in simply being mentioned in highly relevant “conversations”.
But there’s a little twist to this.
Ahrefs ran the numbers across 863,000 keywords and 4 million AI Overview URLs, and found that 37.9% of cited URLs also appeared in the top 10 blocks for the same query
In other words, getting cited by ChatGPT doesn’t necessarily mean you’re on page one of Google for the same query.
A little frustrating, sure, but the emergence of LLM SEO just means brands have a new way to get in front of their customers.
LLM SEO Quick Start Guide: Where to start this week
The good news is, a strong SEO strategy will support your LLM SEO strategy indirectly. The other good news is, you can take some additional steps that your competitors might miss in order to get cited by LLMs more frequently. Here’s what you can do this week:
1. Check that an LLM crawler can read your website at all
Choose three of your most important pages and retrieve them without JavaScript enabled, or enter the URL into any "view as text" tool. In that case, the main text should be restored as the first step, since a retrieval system which cannot interpret the page will never include it regardless of how good the writing is. Server-side rendering and clean Markdown output should be the basic requirements here, with all the other points in this list built upon that foundation.
2. Add an llms.txt file to your root
Imagine an llms.txt file as a brief, plain-text map for AI crawlers, it tells them who you are, what the site is about, and lists the few URLs that are most important. It only takes half an hour and is one of the cheapest wins available for AI search optimization at the moment, since most websites still don't have such a file.
3. Audit your top 10 pages for stale facts
Look through each item and locate all the numbers, prices, product names and study references it contains. For each of these, check if it is still accurate today. A statistic from 2023 on a page that is otherwise correct is precisely the type of thing a retrieval system will use against you when deciding among five similar sources. Alter the figure, update the citation, and only after that should you change the publication date.
4. Restructure for extraction, not for reading
Instead of copying passages, models copy entire pages. Therefore, each H2 should pose or respond to one specific question, and the first sentence underneath it should be the answer, not a preliminary statement leading up to it. If in any section the useful sentence appears in the third paragraph, then move it to the first paragraph. This is the point at which answer engine optimization and simple good writing are almost identical.
5. Get your schema accurate and attached
The JSON-LD for the article, the FAQ, the product or the organisation should match the information that is on the page; it should not be a template that was set up in 2024 and still displays the old company name. You must run the three pages from the first step through a structured data validator and make any necessary corrections. Having incorrect or mismatched schema is worse than having no schema at all since it causes the system to believe that your page contains contradictory information.
6. Start measuring citations, not just rank
Since Search Console doesn't carry out this task for you, select ten queries that are important and check them by hand this week in Google AI Overviews, ChatGPT and Perplexity. Make a note of the pages that are cited and those that aren't, and then carry out this process monthly. Although it's rather crude, it's the only way of seeing the gap that this entire article is concerned with, and you can't decide which pages to refresh without doing this.
7. Put the refresh on a schedule, or on an agent
The fixes you carry out in step three will once again become outdated within four to six months. You should now decide who is to carry out the next pass and when it is to take place, and enter this into the calendar before the momentum fades.
If you prefer not to depend on someone remembering, then this is the sort of task that Fimo's agents were designed for. A content refresh agent goes through the outdated pages on a set schedule, a GEO agent keeps llms.txt and the schema up to date as the site changes, and each of them opens a pull request for a person to approve. The detection process runs independently and the final decision remains with you.
Content freshness is key: What the research says about recency in AI search optimization
Before we dig into the data, I’ll summarize. LLMs love to cite nd link to fresh content.
Updating older content needs to be a big part of your LLM SEO strategy. The more frequently you update existing content, the more LLMs trust it, and you.
A study presented at SIGIR-AP 2025 tested whether language models favour newer documents when reranking search results. Researchers took passages from the TREC Deep Learning collections, prepended artificial publication dates, and reran the ranking across seven models including GPT-4o, GPT-4, LLaMA-3 at 8B and 70B, and Qwen-2.5 at 7B and 72B.
Fresh passages got promoted consistently. The mean publication year of the top 10 shifted forward by as much as 4.78 years. Individual passages moved by as many as 95 ranks. And in pairwise tests, where two passages had identical relevance, the model's preference reversed up to 25% of the time after nothing changed except the date.
The content studied had the same information, same quality, same usefulness to the reader. The date of the most recent update alone moved the ranking. Which means a page can be the best answer on the internet and still lose to a worse one that was touched more recently.
Nobody has a toolset for the signal that decays fastest
So freshness carries real weight in whether you get cited. The problem is, keeping tabs on older, already-published content is just as much work as publishing new content.
What most teams do about this is a quarterly refresh sprint. Pull a list, work through it, update what looks most stale. It's better than nothing and it's what we'd recommend if you have no other option, but let's call a spade a spade here, its a manual workaround that takes up a lot of time and effort. Plus, a quarterly cadence means a page can sit wrong for eleven weeks before anyone looks at it.
Knowing which pages to prioritize once you do sit down is a separate skill, and we went through the prioritization properly in our piece on why content freshness decides who gets cited and which pages to update first. What follows here is the other half of the problem, which is how you find out a page needs attention without a human noticing first.
To fix all of this, we built Fimo with native content refreshing agentsthat work overtime on your already-published content. They keep tabs on how old content gets and make strategic updates to entice LLMs to cite you over competitors. Oh, and it’s all totally autonomous, with no human triggers needed.
Getting cited is a rendering problem before it's a writing problem
Before any of the maintenance matters, a retrieval system has to be able to read the page at all.
That means pages rendering as clean Markdown a crawler can parse without executing JavaScript, JSON-LD attached and accurate, headings structured so a model can lift a single section as a passage, and an llms.txt file at the root that tells AI systems what your site is and where the important pages live. Fimo ships all of it by default rather than as a plugin you configure, which we covered in our guide to getting pages cited by AI search engines.
Get this wrong and nothing else counts. Get it right and it stays right, until someone ships a change that breaks it. Which brings us back to website maintenance (or ideally, autonomous website maintenence!).
Autonomous websites: How Fimo's agents automate LLM SEO
Fimo's answer to the instrumentation gap is to put agents on the published site rather than only in the content database.
In practice that's a handful of jobs running on schedules you set. An SEO agent checking where traffic comes from every 30 minutes and flagging pages whose search health is slipping. A content refresh agent working through ageing pages on a cadence rather than waiting for a quarterly sprint. A GEO agent keeping llms.txt and schema in step as pages get added and changed. A competitor agent scanning every two weeks and flagging what moved, which is often the thing that made your page wrong in the first place.
You can also build one for whatever your team keeps forgetting. That's the part that tends to matter most in practice, because every content operation has its own specific recurring job that no off-the-shelf tool covers. There's more on how the agents are configured on the agents page, and on the difference between agents that act on your content records versus agents that act on your live site in our explainer on the agentic CMS.
None of them publish on their own. Each runs in an isolated environment with its own server, database and asset bucket, and each opens a pull request that a person reviews and merges. Every change is versioned and reversible through Git. So the agent finds the stale page and prepares the fix, and you decide whether the fix ships. If you want the fuller comparison against bolting a monitoring tool onto an existing setup, we went through the options in our roundup of SEO automation tools.
