Prompt tracking vs keyword tracking: AI answers or search rankings

What does each one track?

Keyword tracking records how a search engine treats your site for the queries people type. Its best source is the engine itself: Google's Search Console reports clicks, impressions, click-through rate and average position, by query, page, country, device and search appearance. That is a record of real events. Google counted them; you read them.

Prompt tracking records how AI assistants answer the questions buyers ask. There is no equivalent of Search Console for ChatGPT or Perplexity, so you write the questions yourself, put them to the assistants on a schedule, and read the answers: whether your brand was named, whether your site was linked, and which sources the assistant drew on. That is a sample you design. You chose the questions, so the answers describe your chosen questions and nothing else.

That difference, a record of real events against a designed sample, explains most of what follows. Even the engines say so: the ChatGPT answer at the top of this page tells its reader to use realistic questions, not just keywords.

Why do prompt numbers move when nothing changed?

Because models are not deterministic. The same prompt can get a different answer an hour later, so prompt tracking needs repetition and statistics. In Attensira, each prompt, engine and country combination runs at most once per workspace-local day, with one sample per daily reading on Starter and three samples per daily reading (n=3) on Growth and Business. Every reported change has to pass a two-proportion z-test at 95%, or it is shown as no change we can prove.

Keyword data has its own instability, of a different kind. Google warns that even if a query appears in your list, you might not see your site if you run the same search yourself: results vary by person, place and time. The difference is that Google is reporting what happened, so the instability is in the world, not in your sample.

What can each one tell you that the other cannot?

Keyword tracking tells you about real demand: which queries actually brought impressions and clicks, how many, from where. No prompt set can tell you that, because a prompt set is a list of questions you chose.

Prompt tracking tells you what buyers are told: which brands an assistant names for a question, which sources it trusts, and whether yours is among them. No search report can tell you that. Google even folds traffic from its own AI features into the ordinary Web totals in Search Console, so you cannot isolate an AI Overview's clicks there.

Our own head question shows the gap. In the 30 days to 4 October 2026, ChatGPT and Perplexity answered "how to track AI search visibility" 108 times, named Semrush in 36 of them and named or cited Attensira in none. A keyword report would show that query only if it had earned us an impression, and even then it would say nothing about who the answer recommended. When ChatGPT answered that question, the pages it leaned on most were Google's own Search documentation and OpenAI's help centre: the platforms describing themselves, not the tools that sell tracking.

How do you build a prompt set that is worth tracking?

Start from buyers, not from your keyword list. Write down the questions a buyer asks at each stage: what kind of tool solves this problem, which tools are best for a company like mine, how do two named tools compare, what does it cost, why am I not seeing results. Phrase each one the way a person would type it into an assistant, including the context they would give, such as company size or industry. Then add the handful of branded questions you want to watch, and keep them in a separate group so they never inflate the category numbers.

Forty to sixty prompts is a reasonable start for one product in one market. More than that and the set gets hard to read; fewer and one odd prompt moves the whole average. Review the set quarterly. Questions buyers ask change faster than keywords do, because assistants invite longer and more specific questions than a search box ever did.

In Attensira, capacity is a number of prompt slots, and a slot is a stock, not a flow: it is occupied while a prompt is tracked and returns to the pool the moment you remove the prompt. Swapping a stale prompt for a better one costs nothing but the history of the old one.

How should you read them together?

Page by page. For a page you care about, put its Search Console clicks, impressions and position next to its AI citations over the same weeks. A page that ranks well and is never cited is being found but not trusted as a source for answers. A page that is cited often but ranks poorly is trusted by engines that fetch pages themselves and missing from Google's lists. Each pattern points at different work.

Attensira keeps both records side by side for this reason: Search Console clicks, impressions and positions per page and query, day by day, plus daily rank tracking that notes whether Google showed an AI Overview, next to each page's citations in AI answers. Kept apart, the two sets of numbers tend to be reported by different people to different audiences, and nobody sees the page that ranks and is never quoted.