How I Measure Which SEO and GEO Tools AI Recommends, Built on SearchApi

I had a software index on my website, but it did only one job earlier, it tracked brand searches on search engines. So I added SearchApi to track AI Visibility Signals. In this blog post I go through how I built it and what it measures. The findings were surprising, because they are not always the same software which appears in traditional search engines and AI search. For example, Moz is the fourth most searched of this set of 10 tools but only appears in 11% of answers, and Semrush has 45% of search demand and has 28% of the AI mentions.

Why did I add SearchApi to my Software Index?

Before I did measure only search volume for each SEO & GEO software brand. That gave me the half truth, but it did completely lack the AI Search metrics. So I wanted to understand better which software AI names when different buyers prompt specific questions like “Best SEO tool for a small business”. I used Claude Code, plugged SearchApi with its API-key to it and created the system that I am going to improve month by month.

My previous approach and why I dropped it

I used to track pure brand name volume first on Google then added cross engine data to understand whether data differs for each search engine like Bing, DuckDuckGo and Yandex. It was interesting to me, but honestly it didn’t give me useful information about the original question: which software has the most demand and which are the most cited ones for each situation. That is where SearchApi comes along.

What the Index measures now

Shortly, two things:

First, a brand search volume in traditional search to understand what the search demand is for each software, and which ones are the biggest players in this field. And which ones are the rising and the falling stars.

Secondly: what does AI recommend for specific buyer questions and how does that differ from real search demand in traditional search.

So I wrote AI Visibility prompts to track which software are cited or named in AI answers.

I picked 3 categories, bigger SEO Suites, AI Visibility/ GEO tools and AI SEO autopilots. And I wrote 36 buyer questions like:

• "best SEO software for small businesses"
• "easiest SEO software for beginners"
• "best SEO tool under $100 per month"
• "best SEO software for agencies"

for AI Visibility/GEO:

• "best AI visibility software"
• "best ChatGPT visibility tracker”
• "affordable AI search visibility tracker"

and for AI SEO autopilot:

• "best AI tool that writes and publishes SEO blog posts automatically"
• "best SEO autopilot tool for Shopify stores".

The technical side

First of all, a few mentions of my stack. I use Claude Code for almost everything for thatmarketingbuddy.com website and I added a SearchApi API-key for it. I also have Sanity as a CMS option, but honestly I don’t really use it anymore. In this project it only stores the search demand data. The AI answers are saved as local files.

Caveats & limitations

• this is a snapshot. AI search answers vary from one query to the next, so don’t pay attention to minor differences
• limited pilot with just 36 questions, 35 tools, 3 categories and only US English

Methodology

All the questions were sent to ChatGPT, Google AI Mode and Gemini using the same SearchApi endpoint, you just need to change the "engine" parameter ("chatgpt", "gemini" and "google_ai_mode", respectively). All engines return answers in a structured format that contains the answer in markdown format and references/metadata:

In the ChatGPT case, web_search is set to true, and I also noted the ChatGPT model used which I saved along with the answers. In Google’s case, I sent two extra parameters for location: "gl":"us","hl":"en".

• To avoid typos wasting my SearchApi credits, by default the test just shows the queries without running them. You need to use the --run argument for actually sending them. As an extra safety measure it won’t run if it’s more than 100 queries. I ran it in two batches (72 and 36 queries), so each stayed under that limit.
• I reused previous answers, but marked missing answers as failed. I wanted to note when it failed to get an answer instead of assuming that it wasn’t mentioning the tools when maybe it was.
• In this case it was 36 questions x 3 engines = 108 queries @ 1 SearchApi credit each = 108 answers
• All the queries are saved locally so I can play with the analysis & metrics without using more SearchApi credits.

Why use a wrapper API ?

Indeed ChatGPT and Gemini do have an official API, with websearch feature. I did not find an official API to query AI Mode of Google (feel free to correct me if I’m wrong !).

Even if they all had a very similar API, it would have meant :

• 3 accounts and 3 api keys
• probably some differences in answers formats

With SearchApi, I had just one url to query, with ChatGPT / Gemini / Google AI Mode as possible values for the engine param, and all answers in the same format (a markdown answer + the references links + some meta about the answer)

On the other hand :

• requests are a bit longer (median between 7 and 10 seconds)
• you depend of SearchApi, to query the 3 engines (whereas with their api you depend of each company only for their own engine)

For a project that consist in querying ChatGPT, Gemini and Google AI Mode with 36 identical questions (so a total of 108 requests), to compare answers, it was for me an easy deal !

Metrics

• AI Visibility/TMB AI Visibility % (primary metric): each category has 12 questions, so 36 answers across the 3 engines. If a tool is mentioned in 8 answers, it has a TMB AI Visibility of 8/36 = 22%. I show the average of the 3 AI search engines (primary metric) and the value by AI search engine.
• AI Share of Voice % (SoV): number of times an AI engine mentions a tool divided by the total times it mentioned any of the tools I’m tracking for its category. Note: this ignores competitors I’m not tracking that could get more mentions.
• AI vs Search: it compares the share of category tool mentions (share of voice) with its share of branded search demand*. + means that this tool is getting relatively more AI answer mentions compared to its share of branded searches.

Share of branded search demand ≠ share of the market. And branded search demand is a floor, which may not be reliable for these GEO tools. I marked with a * those with ambiguous names. You should consider AI Visibility as the more reliable metric, specially as most of these tools don’t have a big keyword history.

Instead of creating a made-up final score, I prefer to show these 3 different metrics of AI Visibility (TMB AI Visibility %), relative popularity (share of voice %) and difference vs searches (Δ AI vs Search).

When extracting the mentions, I try to find different name variations (SE Ranking, SERanking, SE-Ranking…), and count only mentions with an uppercase character (not if it’s part of another word). But with very common words as Profound, Conductor or Goodie I only counted it when the AI search engine emphasized the tool name by using it as a heading, in bold, in a table, list or link. When it didn’t find a mention, but linked to its website on references I analyzed each case manually to avoid false negatives.

This is the first run of my new index which tracks AI mentions for software companies across categories of software and types of search.

Search Demand vs AI Mentions

Mentions of SEO suites

• Semrush: 92%
• Ahrefs: 72%
• SE Ranking: 61%
• Mangools: 44%
• Moz: 11%

Interesting to compare this to the search demand (I’ve used 10 September DataForSEO search volume numbers for branded search terms): Moz is the fourth most searched of this set of 10 tools but only appears in 11% of answers. SE Ranking has around 5% of demand, but gets 18% of the AI mentions, Mangools has around 1% of search demand but gets 13% of the mentions and Semrush has 45% of search demand and has 28% of the AI mentions.

These don’t really combine well into one graph so it’s something to be viewed side by side.

Mentions of GEO tools

• Otterly.AI: 78%
• Peec AI: 72%
• Profound: 56%
• Scrunch AI: 28%

Nothing else was over 19%.

I was quite interested to see what would come up for SEO Autopilot:

Koala AI made it into 17% of answers, Distribb was in 14% of answers and TheStacc was mentioned in 11%. 10 of the 13 tools got 6% of mentions or less, including 5 that had zero mentions.

This category also had the highest number of answers which didn’t mention a company which I’m tracking, and in 20/36 of answers no tools I was tracking were mentioned so the AI was referring to other tools. Answers were about tools such as Surfer, Alli AI, SEOwriting.ai, Frase or Byword which I’m not yet tracking in the index. I’ve added a second list of “named by AI but not tracked” companies to each category to avoid a false sense of completeness in the list!

Interestingly, different AIs mentioned different tools a lot more than others, for example Scrunch AI was named in 58% of ChatGPT responses but 0% of Gemini responses. SE Ranking appeared in 33% of ChatGPT answers and 83% of Gemini answers. Distribb was mentioned in 42% of Gemini answers and 0% of Google AI Mode and ChatGPT answers. There’s enough of a difference in mentions that the index shows the average overall and then each AI broken out individually.

Mentions don’t always make citations

For the SEO query Semrush was named in 92% of answers but its website was only cited 22% of the time. AI seems more likely to send you to websites about SEO tools rather than to those tools themselves (ahrefs.com and semrush.com are two of the most cited domains alongside onelittleweb.com, zapier.com and reddit.com).

For the GEO tools query the most common cited sources were zapier.com, youtube.com, position.digital, reddit.com and rankability.com.

Caveats & limitations

This is a snapshot. AI search answers vary from one query to the next, so don’t pay attention to minor differences. Limited pilot with just 36 questions, 35 tools, 3 categories and only US English.

Next Steps with the Software Index

It’ll be interesting to repeat this monthly and see if there are any changes and go deeper with buyer prompts to really understand the market better! With Claude Code and SearchApi it is pretty easy to repeat.