Tracking market visibility across SERPs and AI-answers

Keywords, Google results and LLM / AI-answers look nothing alike. But underneath, they often (not always, but often) share the same attributes: a topic, a stage in the buying process, the brands involved.

Pull out those attributes, and you can report on your market in one language, build AI prompts that cover the same market as your Google data, and connect the two.

Not strings, but things ... ;)

In After share of search comes SERP market research I argued that you should read Google's results pages as a picture of your market, not as a ranking ladder. They show how the market is divided into categories, which players hold which part, and how buyers search at each step.

This article takes the next step: using that same market picture for AI.

Many (maybe even most?) companies now have two reports about their visibility. One for Google, based on keywords. And one for AI, based on prompts: how often does ChatGPT or Gemini mention us?

Put them next to each other and you get stuck. A keyword is not a prompt. "best laptop for students" is a search term; "Which laptop should I buy for university?" is a question to an AI. You can't hold one against the other.

And AI answers are worded differently every time you ask, so even this month's AI report and last month's are hard to compare.

The result: two separate stories, and nobody can say whether you are winning or losing in a part of your market. That is a problem we need to fix, and I think I've found a way.

The way out is to stop looking at the words, and look at what they are about. Every keyword, every Google results page and every AI prompt or answer has a few attributes. For example:

The values come from fixed lists that everybody uses the same way. Such a list is called a taxonomy. Some examples are Wikidata (for brands and other entities), Schema.org (for kinds of things and pages), Google's product taxonomy (for products), the IAB Content Taxonomy (for topics), the Shopify Standard Product Taxonomy and GS1's Global Product Classification (also for products), GeoNames (for places), and ESCO or ISCO (for jobs and occupations). Just to name a few.

You don't need one source for everything. The keyword itself often gives you the brand and the topic. The Google results page tells you the stage: if Google shows shops, people are ready to buy; if it shows guides, they are still looking around. The AI answer tells you which brands it recommends and which stage it is written for.

So in my view, the SERP and an LLM answer are simply two pieces in the same user journey, and we need to treat them as such. That starts with being able to measure the same market in both.

Don't just take my word for it:

That is the key idea. Keywords and prompts will never match as text. But they do match on their attributes.

Take the keyword "e-reader brand x price" and the prompt "What does Brand X's newest e-reader cost?" As text they have little in common. As attributes they are identical: topic e-readers, stage ready to buy, brand Brand X.

Think of two spreadsheets: one with your keywords and Google data, one with your prompts and AI answers. They have no column in common, so you can't combine them. Add the attributes as columns to both, and suddenly they share columns. Now you can put them side by side, row by row.

Once everything shares the same attributes, those attributes become the rows of your report:

This is also the language management understands. Nobody in the boardroom cares about one keyword or one prompt. They care about "how are we doing in laptops?"

Let me repeat: this kind of reporting actually answers questions your management has. And you don't need to answer with 301-redirects or canonicals anymore. We still need to un-nerd SEO if we want bigger budgets and be taken more seriously, and this might help.

This, good folks, is not necessary anymore:

Ask an AI the same question twice and you get two different answers. Different words, and often a different list of brands. But what the answer is about changes much less.

We tested this by asking ChatGPT the same 165 questions three times on one day. Of all the brands and products named in two answers to the same question, only about 1 in 4 appeared in both. But the stage in tghe buying process the answer was written for stayed the same 95 times out of 100. Even by pure luck, two answers would share a stage about half the time, so that 95 really means something: the stage sticks to the question. The words don't.

The same goes for reports. A mix of stages or kinds of answers can be read from a small set of questions. A top ten of brands only settles down with many.

That is the whole case for attributes: the coarser the category, the steadier the number. One condition: the categories must be clearly defined. If nobody knows where "comparing" ends and "ready to buy" begins, the attributes will wobble too.

So make sure the definitions behind your taxonomies are rock-solid. A list of words is not a definition. A good definition says what belongs in a category, what doesn't, and what to do with a case that fits two. Only then can a model, or a colleague, apply it the same way every time.

The attributes do a second job. Most prompt sets are written from a keyword research list, combined with some market knowledge. But really, nobody knows whether they represent the market.

But ...

... the keywords in your market (that means yours, your competitors', and related keywords) and the SERPs already describe the market, attribute by attribute. So build your prompts from them.

Sure, queries are shorter than LLM prompts, and the latter tend to be more exploratory, and also have a lot of 'help me with' actions in them, not so much searches. But still, it's a good source!

Split the market into boxes, one for each combination of topic and stage: "laptops, comparing", "insurance, already a customer". Work out how big each box is from the search volume of its keywords. Then give each box a number of prompts that fits its size, and write each prompt from a real keyword in that box. The prompt inherits the keyword's attributes, and that is what makes the connection work later.

The result: an AI prompt set that covers the same market as your SERP data, in the same proportions as search demand.

Also, this won't happen anymore:

A trend line only means something if your questions don't change behind your back.

They usually do, tho.

Prompt sets get regenerated, extended or rewritten by an AI that never writes the same question twice. In our own work, five rounds of regenerating wrote more than 2,000 prompts to fill about 430 places. From one month to the next, only two to four in every ten prompts were the same. For keywords in the same markets, it was nine in ten.

So lock them. Once a prompt has been used in a measurement, its wording never changes.

Markets do change, so the set has to move too. But slowly, and only when the market gives a reason. Check the boxes every month. Add prompts when a topic grows past a clear threshold, remove them only when it shrinks two months in a row, and ignore the small ups and downs. Make sure every brand with a real share of the searches appears in at least one prompt.

You will find that markets move far less than their keywords suggest. New keywords appear every month, but they mostly land in boxes that already exist. A new phone model brings new search terms, but they still belong to "phones, comparing" and "phones, ready to buy". The words are new; the shape of the market is not. With these rules, our sets would have changed by only a few prompts a month, instead of being replaced.

Keywords, Google results and AI prompts are three ways of looking at the same market. Report on them separately and you get three stories that never meet. Pull out the same attributes from all three, and you can report in one language, build prompts that represent the market, and connect your Google and AI data.

Lock your prompts, change them only when the market changes, and keep your categories clearly defined. Then, when your results move, it's because the market moved (or you ... ;)), not your questions.

I ran into this problem in my own work: AI visibility reports that couldn't be compared with each other or with Google. The approach in this article is how I solved it, and I now use it in Skåut, the SERP market research I run for clients. The numbers come from one of those clients, in Dutch, measured in August and September 2026. Curious what this looks like for your market? Get in touch.