Caso real de trabajo en optimización de LLMs
I’ve decided to work, together with my team, for two weeks on reputation strategies to improve the reputation in artificial intelligences.
Since there is little documentation on this, I’ve decided to work on it based on my own criteria inspired by SEO, with a custom methodology.
I want to share this so that you can work on “GEO” or positioning in LLMs by copying strategies from my successes and avoiding my mistakes.
As this has been more of an experiment, the changes have not been very radical; nevertheless, I believe they serve as an example of a strategy and analysis with my own conclusions.
I have taken my own project, which is the reputation of the “Technical SEO Master’s program” in LLMs.
I chose something very specific that could have as few variables as possible, that it worked, it was a real business with real results, and where the SEO work was already done.
In this case, “Technical SEO Master’s program” has been in first position organically for a long time in a stable way.
So the strategy to improve reputation in LLMs, knowing that the SEO is already done (remember that “GEO” would be 80% SEO and 20% other complementary strategies), is as follows:
Now. I will explain how my team and I came to the conclusion that those keywords were relevant and why we added them to the strategy.
As the mathematician William Thomson said: “what cannot be measured cannot be improved”, so before carrying out the whole strategy we are going to see how to measure it.
Since in Analytics I am missing certain configuration in this project, it samples the data a lot and Cookies are involved, I am only going to take into account two measurements:
For ease of measurement and tool compatibility, I will only take ChatGPT into account.
I have had to be very careful with the measurements due to inconsistencies I have found with the tools, so I have had to put on my “researcher’s lab coat” to avoid making wrong decisions. At the same time, I have reported all the errors I have found in all the tools used.
Simply configuring the tool has helped me enormously to build a “way to do”. It made me think about which prompts I am interested in appearing and what type of visibility would be desirable.
Simply filling in the information to configure what I needed to extract from the Artificial Intelligences helped me a lot to establish the strategy.
The way SE Ranking works is that you give it a list of prompts and SE Ranking runs them day by day and shows you in a graphical way all the data on how many times you have been mentioned or cited as a source.
In other words, there is no way to track what users do inside their private Artificial Intelligence applications, but it is possible to make an estimatation to have a sort of protoype of a “visibility index”.
Since the idea of this article is to generate a methodology, my team and I decided to create a seed PROMPT that would give us interesting prompts that people might use, where a satisfactory answer could be the master's program offered by our company.
Although now, thinking about it calmly, I can come up with certain adjustments and and more customized improvements (which is what I actually recommend doing), this idea helps to overcome the “blank page” crisis and to get ideas.
The seed prompt to get different prompts to analyze from SE Ranking was:
“I want to ask an LLM several questions so that it looks for information about which is the best technical SEO master’s program in Spanish. Can you help me generate 10 prompts so I have some variety and can get multiple answers?”
And the truth is that it generated quite long prompts, which users might not use so frequently in real life.
The first detail we encountered was that SE Ranking reformulates prompts it considers too long; it not only shortens them, it actually transforms them.
Sometimes it does this very efficiently, but in other cases it changes the meaning of what you want to ask, so when using this tool I do not recommend Copy & Paste. You should review the prompts and shorten them if you want to analyze exactly the prompt you originally wrote.
Example:
Original prompt:
Act as a digital training analyst. Identify the 5 best technical SEO master’s programs in Spanish. First define 8 objective criteria (e.g., server logs, JS SEO, rendering, WPO, structured data, edge/CDN, real cases, employability). Score each program from 0–10 for each criterion, explain why, and deliver a table ordered by total score. Include prices, modality (online/in-person), duration, and exact dates of the last syllabus update. Cite at least 6 sources with URL.
SE Ranking reformulated:
Team reformulation:
In addition to these prompts, we added several “fan out queries”, which simply consist of breaking down a complex query into several more specific subqueries—something that will probably better emulate a real user’s query and, frankly, will let us better analyze what we should improve.
The initial plugin configuration was quite simple and allowed me to add ChatGPT, so we will be able to compare both data sources.
To design a strategy, I really liked SE Ranking’s model, because the act of filling in all the information about what you want to analyze forces you to focus more on what you are actually trying to find out.
Since SE Ranking lets you set multiple prompts and results, we realized that when searching neutrally for a technical SEO master’s program, it consistently mentioned certain words that did not appear as frequently in our content.
So it is easy to conclude that if the AI itself, with different prompts, keeps mentioning certain keywords, it is because it considers them important. And in this case, it matched some topics that we do cover in the program but to which we had not given as much prominence on the website. So that’s what we changed.
It’s not a classic keyword research, but you start to see the logic behind the semantic connections the AI makes, and we understood that if we provide information around those terms that ChatGPT sees as closely related, ChatGPT would eventually reward us.
Let’s quickly review the subtle changes we made:
I’ve never been a big fan of paying for links, but I thought that for this strategy of gaining visibility in LLMs it could be interesting— especially on pages where they mentioned “competitors” that were already appearing in certain ChatGPT prompts while I was not yet there.
I also managed to improve purely organic links without paying, but for transparency in this article I did make a very small paid investment: around €335.
We decided to update the content with things that were true but that we did not have written out on the landing page, as we had mentioned earlier— adding keywords that reflect what we offer and that, in our perception from different prompts, ChatGPT associated as concepts very closely related to technical SEO: logs, servers, CSR, SSR.
Once we had prepared everything we wanted to analyze, refined our content, made our implementations and done the link building, we could say that all that was left was to wait.
While I was researching cool features, what didn’t take long to show up were issues in the analysis phase. I’m going to talk about both the positive aspects and the points to improve. I’m sure these bugs will be fixed soon, but if someone is thinking of applying a similar strategy, I recommend keeping them in mind while analytics tools continue to evolve.
Fortunately for you (and unfortunately for me), these issues will probably be fixed by the time you read this article, but I still think it’s relevant to share the experience.
To summarize, what SE Ranking’s AI search tool does is send, on a daily basis, the set of prompts you’ve selected to ChatGPT. Based on the answers that ChatGPT returns, SE Ranking logs the brands and links where you’ve been mentioned.
The SE Ranking tool itself kept updating during the experiment, so keeping track of everything was a bit chaotic. So I’m going to walk you through the main issues I ran into.
The feature is very useful; however, I really miss an option to define yourself what you consider to be your brand.
For example, if ChatGPT mentions “Whopper”, it should be associated with your brand if you’re analyzing Burger King.
I miss a field where you can add which terms are associated with your brand.
Aside from that, the idea is quite good and, for example, I’ve been able to see—at least as an indication—in which prompts I’ve started to appear. I assume that’s thanks to the strategy we implemented.
While I was running this experiment, I reported the bug I’m about to describe to the SE Ranking team, and they told me it’s already in QA and will be fixed in the next deploy. When it’s solved, I’ll update this section.
I noticed that the tool only analyzes links that appear inside the “sources” button in ChatGPT.
I see this as an area for improvement, because I suspect (subjective opinion) that links embedded in the text are even more powerful than the ones that appear in the “sources” section.
Even so, I found it strange that prompts whose answers tend to include embedded links were being logged as having no links
and links only showed up when the “sources” button appeared:
This was happening even with prompts where we explicitly asked to include verifiable sources:
This high percentage of prompts without links made me suspicious—both of the quality of my prompts and of what was going on between ChatGPT and SE Ranking—which brings us to the next point.
The first thing we noticed when we inspected the code that SE Ranking provided from ChatGPT was that it was using the 4th version (this is more of a curious detail than a real problem):
This happens to be the same as the free version of ChatGPT.
This is not necessarily a problem, since most people use the free version of ChatGPT.
81% of ChatGPT users use the free plan. sqmagazine.co.uk
However, we did notice several issues with the answers SE Ranking was returning through its dashboard.
23/10/2025 Updated: This bug has now been fixed and no longer occurs, but at the time of the experiment it did happen.
At times, we lost the correct metrics for certain days because prompts were being sent through the free version, which polluted the data.
We haven’t been able to pinpoint exactly why, but sometimes—depending on the browser or random factors—ChatGPT would return very different results.
According to what we’ve been told, work is underway to produce more stable outputs, where the average response is calculated properly with users geolocated to 0.
And now, the last situation.
The GEOhat LLM tool tracked clicks at the moment whenever someone clicked. That’s good—but the problem is that it also tracked my own clicks. Although the plugin will be updated to allow adding restricted IPs for analysis, at the time of this experiment that feature wasn’t there. So I have to be honest and admit that several of those clicks from different LLMs were mine while testing how effective it was, and unfortunately, that polluted the real user metrics.
Since version Geohat LLM 2.0.3, it’s now possible to restrict IPs.
I want to be as honest as possible: with the data obtained, I can’t draw solid conclusions—only learnings.
The numbers are too small to be statistically representative, but they’re still enough to highlight a few points.
What I see with these tools is that they’ve come up with very good ideas in record time to adapt to this new era; now it’s just a matter of time before those improvements are fully polished.
The conclusions we can draw about what to keep in mind are:
I need to be self-critical here: I think we could have created much more specific prompts where it would make clear sense for our brand to be mentioned.
With the automatic prompts from SE Ranking, you can see from the answers what pain points ChatGPT tends to highlight. If those match what you offer, it’s perfect material to include in your content strategy.
The “sources” feature is excellent and, based on our manual anlaysis, it tends to match reality quite well. SE Ranking has a field called “Source frequency”, which can help you decide on which pages it would be interesting to get mentioned or linked.
According to these analytics, 84 users reached my website through ChatGPT, but we could probably discard at least 10–20 clicks due to “hits” generated by myself while checking where the links in ChatGPT’s answers were pointing.
From the internal contact form we use, we know that 1 of the users who requested information about the master’s program came from ChatGPT, but the truth is that we cannot estimate how many people have compared options and ended up choosing our course because of ChatGPT’s answer.
However, beyond the issues in SE Ranking’s results and the IP matter with GEOhat LLM, there is one conclusion I can draw.
The fact of having dedicated time to optimizing our presence in ChatGPT has led to a small increase in user traffic coming from LLMs:
In the case of Analytics, since there have been other campaigns running, we can’t say for sure that the entire increase in users is due to this.
But the estimate is an increase of between 5–10 users on business days, on a page with a high ticket and an average of 80 daily users.
While the figures are relatively small to be considered representative, each person can decide whether investing the time required for these “GEO” improvements is worthwhile.
From my point of view, it will be worth it if you’ve already done your homework. And although LLM analytics tools still need some refinement, they are clearly heading in the right direction.
Thanks to the entire Asdrubal SEO team, and especially to Mario, for helping me with the data, prompts and ideas to do a solid job.
I currently offer advanced SEO training in Spanish. Would you like me to create an English version? Let me know!
Tell me you're interestedIf you liked this post, you can always show your appreciation by liking this LinkedIn post about this same article.
