Identify searches that need grounding to focus your AI strategy

Mark says: “Everyone should be determining whether their AI searches need grounding or not.”

What kind of grounding are you talking about here?

“When you input a query to a large language model-based platform (Google's AI Mode with Gemini, or ChatGPT by OpenAI, for example), one of the first things that model will do is determine whether that query needs grounding. Grounding means that it needs to go and check its answer on the web for more recent information.

To give you an example, if you searched for, ‘What do red blood cells do?’ it's highly unlikely that query would need grounding because the answer is not likely to change and is not likely to need up-to-date information.

Large language models know this because of the way they work, which is through token prediction. They look at all the content on the web and they recognise patterns. When they see a very sharp bell curve showing lots of consistency in the answer that doesn't seem to change over time, they can be fairly confident that there's a good consensus, and that this is the correct answer.

If you ask a question like, ‘What happened in today's news?’, what the large language model sees in those tokens is going to look very different. From a probability point of view, there isn't a clear indication of what it needs to say next, because it's a topic that regularly changes. Therefore, it needs outside information.

Those are the queries that will require what we call grounding, where it will go off and do web searches in the background to find information to answer that query for you.”

What does this mean for an SEO trying to get a new, up-to-date answer featured within AI search results?

“If you are focussing on AI visibility, knowing which queries are being grounded will give you the inside track on the ones that you can actually change in a reasonable time period.

With the example I gave earlier about red blood cells, that's what I would define as solved knowledge. The only way to impact the answers there is to get into the model’s training data, which is a long process. You won't see the impact of that until there's been an update, and because of the way it's fetching the information, it's quite a lot of work to change that.

Conversely, if you know a query has a high chance of being grounded, we've managed to impact the result the AI is generating within days. It's a massive time saver. If you have 500 related queries for the topic you’re going after, but 200 of them are not likely to need grounding, they aren't worth focussing on, at least in the near term.”

Can you define the probability that a phrase will require grounding, to help you decide whether to target that phrase or not?

“Yes, bang on. There are a few ways to do this. If you're kind of techie, every time you do a search on ChatGPT, if you open up the developer console and go to ‘Network’, you'll see a panel called ‘Conversation’. That contains some of the information that you're sending to ChatGPT, and some of the things it's doing in the background.

It actually generates something called ‘search_prob’, which is a number from 0 to 1 that dictates the probability that the query will require grounding, and there is a threshold at which it will decide that it needs to go and do a web search.

Now, that threshold depends on the model you use, and even whether you're using a free or paid version of ChatGPT. The paid versions are a lot more keen to do web searches, I assume because it costs money to do that. The probability threshold is around 0.65 for the free version of ChatGPT, so for anything above that, you can be sure that it's going to go and ground it.

The other way you can do it, if you're slightly less techie, is with the Google Gemini grounding API. You can send queries to that, and it will tell you whether or not it will ground that query. Really helpfully, it will even tell you what web searches it's going to do in the background.

The example that they give in the Gemini documentation is someone asking, ‘Who won Euro 2024?’ and it describes the three web searches it would do in the background to get that information. That does cost a little bit of money; it's about three cents per search.

If you are not even that technical, there are some very helpful community members out there. Dan Petrovic is one of them, and he's released a couple of public machine learning models that he's trained on the output from ChatGPT and Google Gemini. You can paste in lists of keywords, and it can give a fairly good prediction as to whether that query will be grounded or not.

If you just want a free method, he's got a web interface, you paste in your 200 keywords, it'll run for a couple of minutes, and it'll say, ‘These are going to be grounded, and these are not.’”

Are there tools that you can use to uncover potential phrases that you haven't actively considered?

“This is really interesting, and this is how the paradigm has shifted between traditional search and people using AI search.

The traditional search paradigm is carrying out one-shot searches, where you use a few keywords, you go to Google, and search for one thing. Then, you decide on the next thing that you want to search for. Some people have really good Google-Fu, and they know what to search for to get the information they need.

Whereas, with ChatGPT, people are having long conversations. They're asking questions with three or four different aspects to them, which wouldn't work with a traditional search engine.

You don't need that Google-Fu anymore. The LLM is sitting there in the middle, almost acting like an intent decoder.

There are a couple of things you can do to take you from traditional search to the new prompts that people are using. The first is that we have a little personalisation process.

This is important because, if you went to ChatGPT and said, ‘I'm a vegan, give me a recipe for tonight,’ and then six hours later, you said, ‘I'm going to start running, recommend some running shoes,’ it would say, ‘Because you're a vegan, here are some cruelty-free brands and running shoes that don't have leather in.’ It impacts the type of result you're getting.

Therefore, the first thing we do is use the ChatGPT API to pass over a description of your ideal customer profile, along with the traditional keywords. You might say to ChatGPT, ‘If I were a middle-aged guy who's just starting running, and I'm health-conscious and vegan, and I was looking for this information (then, I would input the keywords, like ‘running shoes’), what kind of things might I ask ChatGPT?’

I've previously told people never to use LLMs to do keyword research. This is not keyword research. You're not trying to find out what people are typing into Google. This is perfectly aligned with the training data of LLMs because they have hoovered up all of the hyper-specific forums and Reddit threads, where someone on the vegan subreddit has said, ‘Hey, I'm a 42-year-old guy, I just started running. What trainers should I get?’ It's perfect for giving you the type of queries those customer profiles might be asking for, and you can ask several.

The other dimension we want is: What are the next most likely questions that person is going to ask? That's when we bring in the AlsoAsked API, which is the Google People Also Asked data. For whatever question you provide, that will give you the nearest intent-proximity questions. It's so fascinating looking at these because people are so predictable.

With the example of running shoes, the next question would be about how much a good pair of running shoes costs. Then, the next most likely question is something along the lines of, ‘How much faster will good running shoes make me?’ People have seen a €50 pair and a €200 pair, and they’re wondering, ‘How do I justify the €200 pair? How much time is it going to take off my 5k?’

If you had 100 traditional keywords, once you've pushed them through two or three different ICPs, and you've asked ChatGPT what kind of queries they're going to generate, you may get 3,000-4,000 prompts from that. It's those prompts that you then run through the grounding prediction, so that you can cut them right back.

Then, you end up with sets of questions that are very likely to come up in conversations. You know that for sure because they come from data. They're also very likely to be phrased in this way, because it mimics how people talk in forums and how they have real conversations with other people, which is how people interact with those chatbots/AI search platforms.

Then, you know which ones are likely to be grounded, from the grounding API, and you can tie each of those to a list of the key phrases that the AI search is going to do in the background (which will normally take place on either Google or Bing, depending on which search engine you're using).”

Are Gemini and ChatGPT quite similar in terms of which types of queries they require grounding for?

“They are. I think this has to do with the initial language graph and how much consistency there is in the answer they can generate.

One of the tricky things with traditional keyword research was the fact that the same search term could mean two different things to two different people. Conversely, you might have 10 people wanting the same thing, but all typing in different searches. What I found interesting, looking at the searches these models are doing, is that there's a remarkable consistency in how they perform their web searches.

They have a fairly static process: ‘This is the information I need, so this is the search I need to trigger to do that.’ It's more like you're dealing with one entity for those searches, which gives you a nice target to aim for.

The outcome of this is to have key phrases for traditional SERP pages that you need to influence in order to appear in the AI search, because that's the final step. It does vary a little bit from traditional SEO because, rather than trying to rank highly for specific queries, you will have three or four key phrases that you know AI is searching for, and the goal is to simply have your information present across as many of those results as possible.

If you're trying to sell your brand of running shoes, maybe the first two results are review websites, and the fifth one is a blog. You can probably be present in all of those if you send them a sample, talk to them, and do some PR with them.

The AI will typically summarise the first 10-20 results. It's quickly checking them all out – unlike the human behaviour of doing the search and then a lot of the activity only occurring in the first few results, so you have to be there.”

How reliable are the synthetic prompts that you can develop through an LLM, compared with traditional keyword data?

“It's not reliable in the sense that it is precisely what someone is typing in, because nobody can get that data. The only way to get that data would be through very invasive clickstream monitoring, which isn't going to happen with the way privacy laws are going.

However, it is very accurate in terms of what we want to know, which is what actual searches are being triggered in the background. If you have 10 questions that are worded slightly differently, but they mean the same thing, they will all generate the same searches in the background. That's what we're interested in.

While I don't know the precise word order that some individuals might use, it doesn't matter. The Google-Fu thing is now happening through the LLM, and we have seen a remarkable consistency with the web searches that they choose to carry out in order to find that grounding information. It's such a useful, powerful technique, because that's the thing you need to impact to actually get seen in the end answer.”

Mark, what's the key takeaway from the tip you shared today?

“Even if you're doing traditional SEO now, if you're thinking about AI search, it's a multi-site approach.

It's not just about getting your site to rank; it's about getting your service, your brand, and your products talked about in all of the results that are coming up, and making sure you've got good sentiment that describes what you're doing, because that's what the AI is going to be filtering for.”

Mark Williams-Cook is Director at Candour and Founder of AlsoAsked. Find out more over at WithCandour.co.uk.