In SEOmonitor, the keyword you track is the topic measured in AI Search, and its Google search volume is used as a proxy for that topic's demand. Because LLMs do not expose volumes of their own, this verified search volume is what puts AI Search on the same footing as your organic and AI Overview data: the same topics, weighted by the same demand. See more about this here.
To carry the intent of the topic correctly, SEOmonitor turns each tracked keyword into the question a real person would ask an LLM to get the same answers Google gives for that keyword. That question is what runs against the LLMs, and the response that comes back is what SEOmonitor analyses: what the models answer, and, on commercial topics, which brands and sources they recommend.
This article explains what SEOmonitor sends, how the response is collected, and where to see both in the product.
Note: This covers AI Search (LLM) tracking. AI Overviews (AIO/SGE) are triggered by Google on the keyword itself and are covered in their own article. For how the responses are scored (mentions, citations, sentiment, Share of Voice), see AI Search Tracking.
What SEOmonitor sends to the LLMs
Your keyword carries every intent people have when searching it, and Google answers those intents with its results page. SEOmonitor reads those intents from the keyword and folds them into the one question a real person would ask an LLM, then sends that question. Where a keyword is too ambiguous to read on its own, SEOmonitor checks its live Google page 1 (the SERP features and the top 10 results) and resolves the intent from what actually ranks.
A few rules keep the question faithful and consistent:
Every real intent is covered in one question, dominant first. A keyword often carries more than one intent at once. For "paracetamol", people want to know both what the drug is and where to buy it, so the question covers both and leads with the dominant one. A real intent is never traded away for another, and none is invented.
Products and brands carry a commercial intent. A specific product, model or branded item almost always returns shopping and store results, so the question asks what it is, where to buy it and the price, not an information-only version. For example, "dyson v15" becomes "What is the Dyson V15, how much does it cost and where can I buy it?"
Services and categories invite the providers. A service, a category, or an "X for Y" keyword usually carries a selection intent alongside the informational one: people want to know which tools, services or providers are best. The question invites naming them, so "management for social media" asks what social media management involves and which tools or services are the best options for it. Purely informational keywords get no providers clause, so brand presence is never inflated.
Ambiguous keywords resolve to the meaning Google leads with. For a term with more than one meaning, SEOmonitor takes the intent Google leads with, preferring commercial over basic informational. "jaguar" with car results becomes a question about the car, not the animal. A brand people navigate to becomes a navigational question, and a genuinely split term asks what the term can refer to.
The question stays faithful to your keyword. SEOmonitor adds no specifics your keyword does not contain (no places, prices, dates, models or companies you did not include), and keeps your product or topic names exactly as written. A name or spelling is never "corrected" to a more famous lookalike: what is sent is the keyword you actually track.
The question is written in the language your campaign tracks. It is phrased the way a real person asks an LLM in that language: one natural question of a sentence or two, with accents and diacritics kept naturally.
The question is universal and evergreen. No persona, no personal scenario, no clarifying questions, and no time words such as "latest" or "today". The question is stored once and asked repeatedly, so it is written in a form that does not date.
How SEOmonitor collects the response
SEOmonitor sends the question to the LLMs enabled on your account in a fresh, neutral session: no history, no personalisation, localised to the location configured for your tracked site.
The same question is reused on every crawl, so your trend lines measure real change in the model's answer, not a differently worded question each week.
The response that comes back is stored in full, and every AI Search metric is read off it: whether your brand is named (mentioned), whether your site is linked as a source (cited), where your site sits among the sources (citation rank), and the same for your competitors. Every run is kept, so you can page back through the responses week by week and see the answer each number came from.
Where to see it in SEOmonitor
You can see the question in the keyword's AI Search panel, under How we ask AI → The question we ask.
For example, the keyword social media management tools (60.5K monthly searches) is asked as:
What are the leading social media management tools, how do they compare on features and pricing, and which should I choose for a small business?
That single question folds together the three intents the SERP shows for it: the comparison ("what are the best"), the use case ("what is each best for"), and the how-to-choose.
Below it you see the response itself, and the outcome read off it: whether your brand was named (mentioned), your site linked as a source (cited), or neither (not listed), and the same for your competitors.
For the full definitions of Mentioned, Cited, Citation Rank and Sentiment, and how they roll up into your scorecards and Share of Voice, see AI Search Tracking.
Good to know
The phrasing may differ from how you would ask it. The goal is not your exact wording, it is the question a typical person asks to get the answers Google gives for that keyword, phrased consistently so it stays reproducible across weeks and comparable across brands.
The question stays fixed until something changes it, and every change is marked. Once generated, the question is stored and reused on every run. It changes in three cases: your campaign moves to this new way of tracking, someone on your team edits the question, or SEOmonitor re-generates the question for a set of keywords. Each of these is annotated on your charts at the date it happened, and trends are calculated from that date, so you always compare like with like. Your history is never rewritten.
Your team can edit the question. Where a keyword has a nuance the automatic question misses, you can rewrite it in the sidebar. The new version is checked before it goes live: it has to still cover the intents your keyword carries, and it is rejected with the reason if it drifts too far. Edited questions are visibly marked, and the automatic question stays on file and is shown alongside. An edit takes effect from the next run and is annotated like any other change.
A few keywords get no question. A string with no readable intent, such as a product code or a random string, is not forced into one. Those keywords stay tracked for organic and AI Overviews and carry no AI Search question.
LLM answers are non-deterministic. Even with the same question, models can word answers differently between runs. That is normal, and it is why AI Search reports trends over single responses (see AI Search Tracking).