How to create tracked prompts that actually measure your AI visibility

What is a tracked prompt?

A tracked prompt is a question or request, phrased the way a real person would ask it, that you monitor on a recurring schedule across AI answer engines such as ChatGPT, Claude, Perplexity, Gemini, Copilot and Google AI Overviews. Each time the prompt runs, you record whether your brand appears in the answer, how it is described, which sources are cited, and how you compare against competitors.

Think of tracked prompts as the AI-era equivalent of a keyword rank tracker, with one crucial difference: there is no fixed results page to rank on. The answer is generated fresh each time, it varies between engines, and it can change from one day to the next. That is exactly why tracking matters.

Why tracked prompts matter

Here is the uncomfortable truth about measuring AI visibility: your tracked prompt set is the denominator for every metric you will ever report. Appearance rate, share of voice, citation counts, competitive benchmarks: all of them are calculated against the prompts you chose to track. If those prompts are wrong, every number built on them is wrong too, and you will make confident decisions based on fiction.

There are three more reasons this deserves real thought.

AI answers are probabilistic. Ask ChatGPT the same question twice and you can get two different answers, with different brands mentioned and different sources cited. A single spot check tells you almost nothing. What matters is how often you appear across dozens or hundreds of runs over time, which is only visible through recurring tracking.

Engines behave differently. A brand that appears in 70% of relevant Perplexity answers might be invisible in Claude or Gemini for the same prompts, because each engine draws on different sources and weighs them differently. Tracking the same prompts across engines shows you gaps that a single-engine view (or a Google-only view) would never reveal.

Buyers are asking AI at decision moments. Research from Ahrefshas shown that AI-generated answers substantially reduce clicks to traditional results, which means the answer itself is increasingly where the shortlist gets made. If you are not in the answer when someone asks “best [your category] for [their situation]”, you were never in the running, and you will not see it in your analytics either.

Not sure where you stand today? Our free AI visibility checker will show you in about a minute, no email required.

Prompts are not keywords

If you come from SEO, the biggest mistake you can make is treating prompts as long keywords. They behave differently in four important ways.

People phrase them like speech, not search.Nobody asks ChatGPT “accountancy software UK small business”. They ask “what accounting software should a small UK business use if they mostly invoice in euros?” Prompts carry context, constraints and intent that keywords never did. Your tracked prompts should read the same way.

Engines expand your question behind the scenes. When you ask a modern answer engine something, it often rewrites your prompt into several background searches (sometimes called query fan-out) and synthesises the results. This means a single well-chosen prompt effectively tests your visibility across a cluster of related searches, not one string.

There is no traditional search volume. You cannot look up how many people asked a specific prompt last month the way you could with keywords. Platforms such as Profound now offer prompt volume data that estimates how often themes are asked, which helps with prioritisation, but the honest position is that prompt selection is guided by judgement about your buyers, not a volume column sorted high to low.

Journey stage matters more than string matching.Two prompts with nearly identical wording can sit at completely different points in a buying journey. “What is answer engine optimisation?” and “answer engine optimisation agency for a UK law firm” deserve different priority, different competitor sets, and different expectations. If you want the academic grounding for how generative engines select and cite sources, the original GEO paper from Princeton researchers is worth your time.

The questions to ask before creating a single prompt

This is where prompt sets are won or lost. Before writing anything, work through these questions honestly. We use a version of this list with every client, and the answers do most of the work.

Who is actually asking? Define two or three real buyer personas, then write in their voice, not yours. A finance director, a marketing manager and a founder ask about the same product in completely different ways. If your prompts all sound like your own team wrote them, they probably did, and that is the problem.

What stage of the journey are they at?Awareness prompts (“why is my brand not showing up in ChatGPT?”), consideration prompts (“how do I improve AI search visibility?”), comparison prompts (“X vs Y for mid-size firms”) and purchase prompts (“best AEO agency UK”) each tell you something different. A good set covers the journey, weighted towards the stages where deals are decided.

Would a real person ever say this?Read each prompt aloud. If it sounds like a keyword with filler words bolted on, bin it. “AEO agency UK best 2026” is not a prompt. “Who are the best AEO agencies in the UK right now?” is.

What is commercially at stake if we are absent from this answer?Some prompts are interesting; some are revenue. “What does AEO stand for?” builds category awareness. “Which agency should we hire to fix our AI visibility?” is a deal happening in real time. Track both, but know which is which, and weight accordingly.

Which competitors should we be benchmarked against here? Competitors vary by prompt. On a broad category prompt you might compete with global players; on a local or niche prompt, with two specialists. Assign competitor sets per topic cluster rather than one blanket list.

Which engines does our audience actually use? A developer-heavy audience skews towards Claude and Perplexity. A consumer audience leans on ChatGPT, Gemini and Google AI Overviews. Track everywhere if you can, but read the results through the lens of where your buyers actually are.

What is our branded vs unbranded balance?Branded prompts (“is [your brand] any good?”) measure reputation and citation accuracy. Unbranded prompts measure whether you get discovered at all. Most brands over-track branded prompts because the results feel good. Resist that.

Does geography or language change the question?“Best payroll software” and “best payroll software for UK businesses” produce meaningfully different answers with different brands in them. If you sell regionally, your prompts should say so.

How many prompts, and how often? More prompts means more signal and less noise. Daily tracking catches the volatility that weekly tracking smooths over. We cover sensible numbers below.

A tracked prompt taxonomy: the six categories every set needs

Here is the taxonomy we build every client prompt set around. The examples use a fictional UK project management tool called “Planbird” so you can see the pattern.

A note on balance: as a starting point we suggest roughly 60% unbranded and problem-led prompts, 25% comparison, and 15% branded. It is a starting point, not a law. A challenger brand nobody knows should weight even harder towards unbranded discovery. An established brand fighting misinformation might justify more branded tracking.

How to build your tracked prompt set, step by step

1. Mine real language first. Before writing anything, collect how buyers actually talk: sales call recordings and notes, support tickets, the questions prospects ask in first meetings, Reddit and industry forums, People Also Ask boxes, your site search logs. Google Search Console queries are useful seeds too, as long as you translate them from keyword-ese back into natural questions.

2. Draft broadly. Write 80 to 100 candidate prompts across all six categories. At this stage, quantity beats polish. Involve someone who talks to customers daily; they will write better prompts than your SEO tool does.

3. Stress-test them live. Run each candidate manually in two or three engines. You are checking two things: does the engine interpret the prompt the way you intended, and does the answer space include brands like yours at all? A prompt that returns generic how-to advice with no brands mentioned is educational, not competitive, and belongs in a different bucket.

4. Score and prioritise. Rate each surviving prompt on relevance to your actual offer, commercial intent, and winnability (is there any realistic path to appearing in this answer within six months?). Prompt volume data helps here where available, but do not let it override judgement about intent.

5. Balance the portfolio. Apply the category split above, check you have coverage across journey stages, and make sure no single topic cluster dominates. Fifty prompts about one product line and five about everything else is a blind spot, not a strategy.

6. Assign competitors per cluster. For each topic group, list the two to four competitors you genuinely lose deals to. This is what turns appearance data into share of voice you can act on.

7. Set volume and cadence. For most brands, 50 prompts is a working minimum, 100 gives you reliable signal, and 200+ suits multi-product or multi-market businesses. Track daily if your platform allows it. Answer volatility is real, and daily runs are what separate a genuine visibility shift from ordinary noise. This is the part where honest effort matters most, and if you would rather have specialists build and run it, that is exactly what our AEO services exist for.

Common mistakes (we see these constantly)

Keyword-stuffed prompts nobody would ask. If your prompt reads like it was reverse-engineered from a keyword tool, the data it produces describes a conversation that never happens.

Tracking only branded prompts. Appearing in 95% of answers about your own brand feels great and means very little. Discovery happens on unbranded prompts, and that is where most brands are losing without knowing it.

Too few prompts. With 15 prompts, one flaky answer swings your visibility score by seven points and someone panics in a Monday meeting. Volume is what makes the data trustworthy.

Set and forget.The questions people ask AI evolve quickly, new competitors appear in answers, and engines change how they source. A prompt set that has not been reviewed in six months is measuring last year’s market.

Ignoring engine differences. Averaging your visibility across all engines can hide the fact that you are strong in one and absent in another. Read the results per engine, then prioritise the engines your buyers use.

Tracking prompts you can win but should not want. High appearance rates on commercially irrelevant prompts are a vanity metric wearing a lab coat.

Measuring and iterating

Once your set is live, these are the numbers worth watching:

• Appearance rate: the percentage of runs where your brand appears in the answer at all
• Share of voice: your appearances relative to your assigned competitors, per topic cluster
• Citations: which of your pages (and which third-party pages) engines cite when you do appear
• Sentiment and accuracy: how you are described, and whether it matches reality
• Position: whether you are the lead recommendation or a footnote

A sensible cadence: daily automated runs, a monthly review of movement and citations, and a quarterly refresh where you retire dead prompts and add new ones based on what sales and support are hearing.

Set expectations honestly. In months one to three, the win is baseline clarity: knowing precisely where you appear, where you do not, and why. Meaningful share-of-voice movement typically shows from months three to six as content and third-party coverage take effect. Anyone promising top-of-answer placement in a fortnight is selling you something other than measurement.

Start with prompts that tell you the truth

Tracked prompts are not glamorous, but they are the foundation everything else in AEO stands on. Choose them like a buyer, test them like a sceptic, and refresh them like the market is moving, because it is.

If you would rather not build this alone: every Tilio plan includes 100 tracked prompts monitored daily across the major AI engines, with competitor benchmarking and a monthly action plan built from what the data shows. Get in touch and we can have your baseline in front of you within the first call.