How I do prompt tracking for AI search

Tracking prompts without a defined prompt set produces a pretty report and zero decisions. The work starts before the tool.

Disclosure: I am a Semrush Ambassador, and this article describes Semrush features. The working criteria described here hold with any equivalent tool, and the passages describing the product are marked with source and date checked.

Prompt tracking has become a mandatory line in SEO proposals, and a good share of the implementations I see repeat the same mistake: someone picks twenty plausible questions, drops them into the tool and starts reporting presence. Six months later nobody knows what to do with that number, except put it in the next proposal.

What fixes this is treating the prompt set as an asset, with entry criteria, classification and review.

What prompt tracking is, and where it differs from keyword tracking

Prompt tracking is the periodic monitoring of the answers AI systems give to a fixed set of questions, logging whether the brand shows up, in which position, with which description and backed by which source.

The difference from keyword tracking is in the unit and in stability. A keyword is short, has an estimated volume and returns a list that changes little from one day to the next. A prompt is long, carries the context of the person asking and returns an answer that can change between two consecutive runs. That is why the set needs repetition per prompt, and the reading is valid as a trend over weeks.

Where a prompt set worth tracking comes from

• Real questions from the sales team. The objections that come up in meetings are the language of people making a decision, and they tend to be the richest and least used source.
• Support tickets. They show the language of people who are already customers and reveal the vocabulary of the problem, which differs from the vocabulary of the product.
• Search Console queries with comparison and decision intent, rewritten in the long form a person uses when talking to a system.
• Semrush Prompt Research, to bring in volume, difficulty and intent at topic level and keep the set from being limited to what you already know.

How I classify each prompt

Each prompt goes in with three tags: decision stage, presence or absence of a brand in the wording, and reputational risk. The third tag is the one usually missing, and it is the one that pays off most, because comparison and criticism prompts are where the brand’s description gets exposed.

What I measure on each tracked prompt

• Presence. The brand shows up in the answer.
• Position in the answer. Showing up in the first set of recommendations is different from showing up at the end as an alternative.
• Description. Which attributes come with the name. Presence with the wrong description counts as a problem.
• Cited source. Which page backed the mention, which ties the result to content work. The method for investigating that page is the Citation Lab.

Prompt tracking without a paid tool

The method works in a spreadsheet, with one row per run: date, platform, prompt, brand cited or not, position, associated attributes and cited source. Three runs per prompt, logged out and in the same time window, give a comparable baseline for a set of twenty to thirty questions.

The tool comes in when volume goes beyond what can be collected by hand, or when daily coverage starts to matter. The AI search KPIs spreadsheet already has this log set up, with the appearance rate and position calculations ready.

The limits of prompt tracking I declare in the report

Generated answers vary between runs, between accounts and between regions. This means the correct reading is a trend over weeks, with repeated sampling. Reporting daily variation as a result creates expectations the technology cannot hold up.

Platform coverage is also partial. Every tool covers a subset of what exists, and the decision about which platforms matter for that client needs to be written down before the first report.