AI SEO & Business Automation Podcast

James Dooley: Probing large language models. Should we be looking to ask AI why they came up with this answer? Should we be probing the AI in trying to get the reasoning behind the grounding sources, everything of what's coming with regards to the recommendations? Is this going to hallucinate? Is it going to give you wrong incorrect information? Or is it going to help you start to understand exactly why the large language models is coming back and giving you those recommendations? Jabez, myself and you jumped on a podcast episode with Dan Petrovic. It was eye opening the amount of times he was probing Anthropic Claude. The amount of times he was then probing exactly the same questions to Chat GPT probing AI mode and trying to get that information back with why they came up with those answers. I don't know if you want to expand on it slightly Jabez and obviously then we can speak to Malt and Jay in what the best way is to try to get the understanding from probing LLMs.

Jabez: Yeah, it was really a kind of eye opener because the whole concept that Dan shared, Dan Petrovic shared with us was that you ask hundreds and hundreds of questions across different LLMs to get a you know an idea of what these LLMs think about your brand or perceive about your brand and it can be a hallucination as well because that's what we want to fix. It can be a totally negative sentiment that doesn't exist but it the LLMs have picked up a wrong info somewhere which was wrongly published or just made up. And the whole idea was like for example uh if you are running a ergonomic chair brand. So the idea is to ask it multiple times from different angles. What do you think about brand X that makes ergonomic chairs? What's why is this brand behind brand Y? What are things that we need to be careful of before buying brand X ergonomic chairs and so on. Just going in as detail as possible because you're trying to get into the head of the LLM as a negotiator. You're not trying to make it answer you like be in line with your answer like a mistake will be do you think brand X is great? Should I buy it? Because in those kind of questions it will try to just say yeah I think it's it's good and you can go ahead and and and purchase something like on those lines but you're trying to ask it like a smart negotiator to get the answers out of its head to just get a direction where these LLMs are heading or what's making up their answer. So that's was the whole idea of probing LLMs to get an idea of how they perceive a brand. What are the positive sentiments, negative sentiments, the attributes attached as well. And talking about attributes, we will uh jump on that part as well later on when we discuss about those um attributes attached to fanout queries as well these days. So but right now what do you guys think about this whole LLM probing? Have you guys tested it? I know there's this whole fear of hallucination, but what's your take?

Malte Landwehr: Yeah, I I think there are different um angles to it and the process you just described, I think it absolutely makes sense and for example uh Dixon Jones from what AI knows about you, he's been talking about this exact approach, I think for the last two years. Um and at PKI, we also do that not for visibility tracking but for brand perception. So we basically ask the LLMs repeatedly tell me attributes of brand and then we also ask now tell me the top brands in industry for attribute for all of these attributes and then we can basically see in an industry which brands are perceived by the LLMs for being trustworthy being reliable maybe also having bad customer support good customer support et cetera et cetera so for that I think it works very very very well if you just want to understand what biases this assumption preconditions has the LLM about you. The problem is you actually need to do it if you want to do it perfectly you would have to do it twice once without grounding and once with grounding to also understand what is the bias or perception that is trained into the foundational model and what actually comes via the grounding process. But for like regular monitoring, I think this is not the way to go because the things that come from the foundational model, they are not going to change next week. They're not even going to change next month. Um, but with with grounding, of course, there is a chance to change it, right? If you work on the sources, you publish content, you you can you can do this. So, I do believe that aspect works well. And then there's the second aspect of probing where you ask about the sources like why did you pick the source? Is this the best source? Why did you not recommend this source? And there the problem is that if you look at the actual pipeline, the LLM you are talking to might not know. Uh because there might be another LLM, a very small one that just decides okay out of the 20 sources which are the three sources we give to the main LLM. Answering the question or which um paragraphs of content which which chunks of text do we extract? Um or in in Chat GPT for example there are some brands that have this news partnership right and um they are actually Chat GPT is not even retrieving the whole contact but just a summary that is static per article so at the moment that the summary was created maybe you were in it maybe you were not regardless of whether you're in the article or not so there are many many steps that the actual model you are talking to does not know because the model has a tool call and this tool tool call says web grounding and it says web grounding. Maybe it generates the fan out queries itself. Maybe there's another tiny model that just takes the original prompt, looks at the user context and quickly creates the fan out queries, right? And then the model you talk to, it doesn't know. So it only it only uses a tool and gets an answer. So I still think it can give you interesting insights to to do these things like why did you recommend this source? Why did you not recommend this brand? Why is this brand position three and this brand position one? But you could achieve similar things by asking a hundred experts repeatedly, why do you think this happened? Right? I the LLM does not necessarily have actual insights into its underlying systems. It's a little bit like when Chat GPT came out, people were like prompting on which uh GPU is your inference running then to prove that they are using an old Nvidia chip or that DeepSeek has access to the newest Nvidia chips. But it was just hallucination like the LLM cannot know this. So, I'm I'm skeptical if people go too deep on this, but I do believe it can give a lot of um great insights that can then go into your prompt monitoring, into your fact checking, into also inspiration for the content that you create to emphasize certain strengths that you have or to address hallucinations about your brand, but also to address weaknesses about your brand.

James Dooley: Yeah, I think on there I think obviously Jay you can expand on it um slightly as well um just to understand a little bit from what Dejan was saying. Dejan was literally of the opinion of more data big data is key. So he was saying do it across every single large language model. Do it multiple times a day and do all the different types of like why should I trust it? Why should I not trust it? And it was going through everything and it was one of them where it was saying never ever ever take a single prompt and a single answer at face value. It is about volume across the different large language models across everything of what needs to be done. Gather all that information and when multiple large language models start to say you're lacking here because your competitor has this and it seems to be better at that point if you want to tackle that you can start to tackle it. Was very much of the opinion of big data do it many times across a lot of different prompts across a lot of different positive and negative sentiment kind of systems and and yeah consistently doing it. So I just wanted to expand on that but yeah Jay what's your thoughts on probing large language models?

Jay: I would say probing is great. Uh but you know again to your point you can't just rely on one answer. I mean because sometimes let's say we have one prompt we have 100 answers for that particular prompt and today we are 36% present in all the answers tomorrow will be 71. So the answer is always changing. I mean they are very probabilistic with their approach. The LLM models like Malt said you know if you people were asking which GPU are you working on the LLM has to answer because they think that they are the most smartest things out there. So they will answer out of the blue even if they don't know they will answer because they are very probabilistic and then uh yeah I would say yeah it should work as in you ask which is the best XY Z software and then you again you ask you know do a follow-up uh inside the tracking as in ask why did you pick this source why did you pick this but again uh you know one answer would not do the justice I guess for one particular prompt you you'll have to do it like thousands of times and then you have to make your uh plan and make your strategy on that basis as in if the model has said like you are yeah I mean your software has been said as one of the less reliable or let's say it says like your customer support is very bad this is what the answer comes up with let's say after 5,000 responses this is the consensus we get off in the probing thing then you should take some call and then you should you know dig dig dive into what is actually bringing that uh sentiment is it the internal memory is it the pre-training data because or is it the third party sources then you'll have to dig into that. If it's third party data let's say if it is a social media if it's a third party you'll have to dig into that. If it's coming from pre-training then you'll have to again you know do all the things uh doing to do press do social do a lot of stuff to get into the next training and then you will be you know you'll have to change the sentiment but yeah uh to your answer uh yeah you'll have to do it at very big scale to learn about you know about your brand and then change the change the perception.

James Dooley: Um both Malt, Jay and even uh Dejan um remember like just recently we were doing some very basic test of pulling out fan out queries and it was just so fascinating to see the kind of fan out queries Chat GPT Chat GPD was giving for SEO experts and we know that that is kind of true in terms of some of the SEO experts the very first query is fan art query is controversies and we know that guy has controversies then for James Tulie fan art query's had recent interviews we we know that he's been on so many podcasts recently and um Rand Fishkin had zero zero click marketing in his fan out query and it was his book that was recently released. What kind of uh how how is chat GPT coming up with such latest fan out queries? Cuz earlier it was just best or top 10 or best SEO experts and very basic stuff, right? How are they able to run these fan out queries which are so personalised, custom and very specific to each person. Where is it coming from? And uh does that and also does this you know give us also an idea of where it's heading is this can also be part of our probing research?

Malte Landwehr: Yeah. So from what we are seeing um Chat GPT doesn't fire all of it fan out queries at once but it does one and then the next one and the next one. So whatever content is returned for the first fan out query can already give some indication of what what they include and then they also have their own index now. So they might very well also have built a knowledge graph where they just know Rand Fishkin is associated with zero click. So they just put that in the into the fan query. Um so I I believe a combination of these things content for the first fan query and whatever they can get from their own index will will power this and I expect this to continue like they will they will have more and more information in their own index and they will they will use that to retrieve information that's also super super recent um and is very very personalised to each entity which can of course have positive or negative effects for entities. Because there are also brands where um in 25% of cases Chat GPT adds the word scandal to the fan out queries and of course that will highlight the scandals that this brand had. So uh can be good or bad. Now one more interesting thing that I uh while running those tests I saw was that whenever I was searching a name that had a very clear knowledge panel triggered in Google right away without any questions or confusions with any any other um expert with that name Chat GPT would straight away start giving the answer with proper fanouts like with detailed fan outs. So I didn't even have to say James Julie the SEO expert in in that case I would just say James Julie but if there were names that had you know uh that were not triggering knowledge panels right away in Google but Google was al asking which uh which David quidd David quit this one or David quidd. So when in that case chat GPT will also not fire detail either detailed fan out query or even the answer it would first ask which one. But when I would search for David Gquid and there's no other person with this particular fan out query then it will come out with its proper fan out uh queries as well that okay the SEO expert his uh channels and then the answer.

James Dooley: Have you guys uh found any correlation in that sense where people with clear knowledge panels and no confusions with any other entity are able to you know uh trigger answers or fan out queries in in Chat GPT easily or let's say is Chat GPT also relying heavily on entities that have knowledge panels in in Google?

Jay: I think that knowledge panels are solely you know the proprietary data of Google and I have not seen anything very similar that correlates with you know chat deep searches and their fanouts you know that relates to knowledge panel but uh answering to your previous question yes they like freshness and as Malt said you know they first and foremost if you search someone's name if they have they'll first and foremost search about them and then if anything comes up in that in one in the first pan out they'll you know just double down on that fan outs and do multiple fan outs based on the first one it just scrolls down over there and then they'll just find information about the knowledge panel uh I have not found but in AI mode obviously AI mode AI overviews that is relation to knowledge panels.

James Dooley: Any any insight from you Malt on that front?

Malte Landwehr: I mean no data specifically for this um I guess it could just be that the same data that leads Google to have a knowledge panel, leads Chat GPT to have in their own knowledge graph uh some kind of statement with similar data about that person. Um it could of course be that they are doing knowledge graph scraping like they scrape a lot of other things. So why not also that um but I I have like it's it's very hard to also test this right because correlation and causation are sometimes hard to separate here. I would think it's probably like it's definitely correlation but I I'm not sure it's it's causation. Um but looking at where search in general is heading anything you have done in the past to obtain a knowledge graph uh entry anything you have done to optimise that. Anything around connecting your different profiles having an entity home making sure systems understand what kind of entity you are. I believe that already is and will be in the future very very good for AI search.

James Dooley: You both have your own tools and I guess both of uh both tools have a feature for planning content and I'm assuming probing might be a part of you know the the content plan. What suggestions would you give to our listeners with regards to right kind of probing? What what what are some you know u steps they should follow to understand and then execute probing and then plan out their content accordingly?

Jay: Yeah. Uh what I would say is probing as as you guys said is very similar to prompt tracking. So you'll have to track the prompt as well, right? Just similarly how we track prompts. You'll have to track the answers and then again uh to plan the content I would say see what's ranking on Google and then see what's getting cited the most. So just merge these two. Uh extract all the NLP terms, all the semantics from the pages that are getting cited the most and the pages that are getting uh you know you know in the SERPs and the pages that are ranking the most in the SERPs and then merge them extract all the terms uh and just highly optimise them in terms of semantics in terms of NLP uh as well as try to learn what kind of page type are they let's say we extract four four SERPs SER results we extract the top top four citations as well. And we learned that uh you know what the page type was listical. So we learned that the page type here is like 60% of the content uh around these two is just how to a service pages. So let's just make a highly optimised service page that answers that particular query and within that query you can you know add all the query fan outs as well about your brand. If something is negative, try to improve that as well and try to cover as many query fan outs as you can within that content. So it's just highly optimised for everything NLP content type uh as well as the fan outs. This is this would be my recommendation.

James Dooley: Okay.

Malte Landwehr: Yeah. I I have one thing you can do additionally um because at the end it doesn't just matter which attributes are associated with your brand. It's also at matters which attributes the LLMs care about the most and you can actually test that with this probing mechanism. So let's say you want to be the number one podcasting software. You use the core LLM without grounding and you say there are three companies uh podcasting has the best customer support. James Podcasting has the most reviews and there's James J podcasting and they are so two compliant. Which one should I pick? And then you do that again and again and then you switch the names and the attributes and you mix in other attributes. And if you do this hundreds of times you will at some point learn ah for podcasting software let's say uh connectivity reliability and integrations are the attributes that Gemini rewards the most and Chat GPT rewards the most good reviews and connectivity and nobody rewards SOC to compliance. Then you would know that whatever attributes are true about your brand these are the ones that you need to aim for and um we have done that in a couple of industries and it's really really interesting that each LLM has some biases in terms of which LL which attribute is the most important one for example when you ask LLMs for the best uh enterprise CRM or enterprise SEO tool or enterprise podcasting software then suddenly ly SOC 2 is super super important the SOC 2 compliance um but if you ask about similar software solutions or services for agencies for startups then SOC 2 is not relevant so if you have any rough idea what prompt you want to target like as a generic general category prompt you can really understand what are the attributes that are actually driving the decision of the LLM um so that that's maybe Another trick here to find out not just what are the attributes associated with you and your competitors, but also what are the attributes that the LLMs are rewarding the most. The bad news is the LLMs don't all reward the same attributes and when there's a new model of Gemini or Chat GPT, you have to rerun this analysis because it's not stable over model updates. But there often are a couple of attributes that they reward. Um if anybody has clients in the car industry, I can give away that safety is in all LLMs highly highly um overvalued in terms to to other aspects like low consumption, energy efficiency, speed design et cetera. If you provide no other context for a car, the LLMs look for uh safety.

James Dooley: Malti, you mentioned uh one thing earlier on and and in this answer as well. You mentioned that run your queries without grounding. How what's the difference between running queries with grounding and without grounding? Uh can you please explain for those who don't have much idea about that?

Malte Landwehr: Yeah, if you run them without grounding, it basically means the LLM cannot get in any actual recent data from websites. There is no RAG, no search grounding, no web search. It just has to answer with the foundation model. So whatever answer you get is based on the biases of the foundation model or the knowledge that is ingrained in the in the weights of the model. And if you activate grounding, that is how most of us usually use LLMs and how most consumers use it. You type a prompt and if it wants to do a search, it does a search. And for most commercially relevant prompts, it will conduct a search. It will find websites and then um base its answer based on what it finds from these websites. And most of the time I would also look at data with grounding because that's the relevant thing. That is what people are seeing. That gives you actionable insights in terms of what kind of sources are working, which websites, which formats, which topics. But if you want to deeply understand LLMs to craft your strategy for the next quarter, it can make sense to run some prompts without grounding just to understand these biases in the model or these these perceptions of your brands and your competitors in the models.

James Dooley: And how do people turn off grounding?

Malte Landwehr: I mean, I use the API and don't activate grounding as a tool. Um, you can in many LLMs In the past, you could at least in many consumer LLMs in the front end deactivate it. I have to admit, I've not tried for many many months to deactivate web search in Chat GPT or or Gemini. Um, but the API is usually the way to go there. The API also has the advantage that you can pick the exact model because you want to test this for the latest model. And if you use the um if you use Chat GPT right now, you have also in the UI very good control over the model. But there were times in the past where Chat GPT would sometimes choose the model for you. So if you typed a very simple question, you might get an answer from a GPT-4 model even though GPT-5 was already answering most of your questions. Um or if you use AI mode and in AI overviews the Gemini version you get it sometimes depends on which country you are in are you locked in or not are you in a paid account or not and if you use the API you can set it very specifically to the exact model you want to probe.

James Dooley: Malt just on the question with regards to do you know earlier when we were speaking about you typed in a name and then it brought back different fan outs and you said that the first sometimes they'll like delay the first fan out then leads to the second fan out which leads to the third fan out. In your opinion, if you had to guess, do you believe that's why one set of query fan outs for one person versus another is being different? Or do you think it's part of what you just been talking there about the foundational grounding training data? Do you think it's just they're doing an internal lookup about that person and forming the query fan outs from that or do you think it is a stagnated they do the first fan out which leads to the second which leads to the third which leads to the fourth and the first one is determining different fan outs for each person?

Malte Landwehr: I think it's both. So if it's a known person like Rand Fishkin then the LLM already knows to slap the word SEO on there. Even though SEO is maybe in his past, it's still what he's known for, right? Um, and if it's a less known person or somebody with a nonclear name because there are multiple people with with the same name, then it will be is more likely that the first fan out query is determining it. And then you will probably also see a first fan out query that says something like Maltan biography or Maltan CV or Maltan interview. I mean in my case I think Jabez also knows that I'm associated with with SEO and some other topics. Um but yeah so it's a combination of both. If they have knowledge in I don't know if it's their knowledge graph or something they quickly pull from the index they will already use that but if they are not sure about the entity then it can only be what the grounding process returns for the first fan query that can influence the the additional fanouts.

James Dooley: I have something to add to the previous one. So uh you know Malt set that you can just turn off grounding. So you can also set the temperature as in how much probabilistic you want the model to be you know in their approach. So you can do that as well and since Malt told that because you turn off grounding but what I would do what I suggest a lot of people is look at the prompts in real time because the model this the users how they how do they use it? They just search something Chat GPT on their behalf just selects the model that it wants to and then it does a web search. So you'll have you should also consider this data point as in how many times you know if let's say for this particular prompt I ran this prompt 1,000 times how many times a web search was triggered for there there was this client who was optimising he's a he's into the health sector he was optimising for something uh that was very informational and I saw that for informational queries there's there's like the web searches are very low so someone who wants to optimise for informational queries it's it's it will be very hard for them because then it is actually coming from the grounding data. So they the you know because if they don't do the web search they're not reading from the third party sources and then you can't do any GEO optimisation on that because they are already just they're not doing any web search. So that is one point as in you you should be looking at uh the web search performed percentage uh as in if it's like less than 30% or 20% just just don't do it just don't optimise I mean just don't spend any money on that that was my my point.

James Dooley: So both of you suggest using the API also for doing a lot of uh probing and testing to get different sides of answer?

Malte Landwehr: Not for your daily monitoring but if you want to get this like initial set of data to form a strategy form a hypothesis. Yes, it can be that's one of the use cases where I think the API can make sense.

James Dooley: Got it. And then on on there just to expand a little bit then on probing the large language models then we've spoke quite a lot here about the different attributes. Now obviously there's lots of different attributes that you can start to cover. You've got branded attributes. So that's all about the company and stuff like that. You've got the service or the products attributes which is specifically about the products and the service that you're looking to sell. And then you've got your trusted attributes. I kind of deem that as being reviews, testimonials, awards, and case studies and stuff like that. How important are which one if you had to choose one? And I know this is just it's the ob the obvious answer is it depends but which attribute if someone come along and says we're looking to improve our AI visibility which one would be the first one you start to look at in branded attributes trusted attributes or service-based attributes which one would you start with and I'm going to start with you Jay.

Jay: Yeah so I would say I highly recommend you know just tracking the fan outs because they give everything in the fan outs safety if it's if it's around safety they'll just search a lot of times XY Z service safety is it legit you know what is what are the risks what are the guidelines and stuff like that so the models I mean specifically Chat GPT gives you gives a lot of data in fanouts which you can just you know identify that okay safety is a concern for Chat GPT for you know recommending this particular service so let's just create you know an attribute that actually answers this particular thing it's related to service so let's just Okay, this comes under the service attribute and then again branding if it's searching your brand or other brands comparisons multiple times alternatives multiple times or obviously just gives away that okay uh branding attribute matters for this particular uh fan out for this particular prompt and then again the same thing uh same prompt multiple runs we'll see some set of fan outs multiple times and on that basis we can just make up our mind that okay these are the these are the attribute that the model is looking after for actually making up its answer. So, let's just focus on creating content and influencing for that. But that's my answer.

James Dooley: What about you, Malt?

Malte Landwehr: Um, nothing really to add. I would I would follow a similar approach.

James Dooley: Sounds good. So, anyone who's listening to this or in fact, I just want you to summarise each one of you. You've got under 30 seconds to summarise. I'm going to start with you, Jabez. You've got under 30 seconds to answer. Do you think it's a good strategy to be probing the LLMs and trying to get some information back to help you along the way that you can be using as part of your content strategy? And I'm going to ask each one of you, you've got under 30 seconds to give your answer and explain why.

Jay: As a 100% we are we have already started implementing this in in our campaigns. Probing should be done at a large scale and probing should be seen as a directional guidance. It shouldn't be seen as a ranking factor. Take it with a pinch of salt. Understand what where these LLMs are heading and it'll it'll give you a lot of idea and obviously it has to be done at large scale with big data to you know to see what's the most common consensus across the board.

James Dooley: What about you Jay? Oh sorry, that was you Jay. What about you, Malt?

Malte Landwehr: I highly agree with Jabez. I mean obviously it's just you can't just rely on one data point or one answer. Just do it multiple times and I would say the regular prompt tracking must do the job for you. You'd get a lot of data about your your brand fan pan outs as as well as your competitor. So I think you should be fine. But if you want dive d uh you know dive I'm sorry uh I mean if you want to dive deep into these things and get more information on that then I think probing is uh the one for you. And then what about you Malt?

James Dooley: And what about you Malt?

Malte Landwehr: For 80% of companies, I think LLM probing is a distraction and it's not the highest value thing they should be doing right now. For the 20% who want like the most advanced strategy with the most data, it's a very very important key piece of their strategy.

James Dooley: Yeah, just to expand on what Malt just said there. I think a lot of people aren't doing the core fundamentals in the first place, never mind then starting to probe to start to find more attributes of what is needed. If you are in the bucket in the 20% of what Malt said, and you're very advanced when it comes to AEO or GEO and you're looking to build up the brand attributes, the trusted attributes and also the service product attributes, probing it might bring some stuff back. Not just just focusing solely on the query find outs. However, the fastest wins is getting the main query find outs that you need to do. So, in this scenario where we've been speaking about safety then with regards to automobiles and stuff like that, you need to make certain the first thing you're covering is that because it's the most important attribute that the live language models coming out with. It's been fascinating to see whether we should be probing um lives language models. For me, if you are going to do it, make certain you're doing it across Chat GPT, Claude, Perplexity, Grock, AI mode. Do it across the whole board. Do it across all of them to get the consensus of what you might need to be covering. It might also start to bring back certain reviews that's being brought up and controversies that you need to start sorting out from an online reputation management point of view as well. Malt, Jabez, thanks a lot and we hope you like the episode on probing LLMs to get more information to try to increase AI visibility.