Should tracked prompts be written as questions or as keywords?

Write them as full questions. People send chat assistants complete requests with context and constraints, so a keyword string run in a chat engine measures a query almost nobody sends. A full question also carries the subtopics, such as region, budget or team size, that can shape which sources an AI engine retrieves and which brands it cites. Keep each prompt to a single question with a single intent.

Should I include my brand name in the prompts I track?

Only in branded evaluation prompts, and report those on their own line. Naming your brand guarantees that the answer mentions it, so mixing branded prompts with category prompts inflates your visibility. Category and comparison prompts should describe the buyer's situation and need without naming you or a competitor, so you can see whether engines recommend you unprompted.

How many prompts should I track for AI visibility?

There is no reliable fixed figure. Published recommendations conflict and rarely show their evidence. Size the set from your own situation: list the topics tied to business goals, then multiply by the runs each prompt needs before its share can be read. Keep the set focused enough that every prompt can be run, reviewed and pruned on schedule, rather than aiming for an exhaustive list.

Do prompting frameworks such as the 'P's of effective prompting' apply to tracked prompts?

Only in part. Prompting frameworks list elements that make a prompt produce a better answer, such as persona, context, examples, tone and output format. A tracked prompt has a different job: it reproduces what a buyer would type. Keep persona and context, because buyers describe themselves and their situation. Drop examples, assigned roles and format instructions, because they steer the answer away from what buyers actually see.

Can I edit a tracked prompt without losing my trend data?

You keep the old data, but the edited prompt starts a new series. Give each prompt a stable ID, a version suffix and tags for topic, intent, persona and language, and record every edit with its date in a change log. Compare the new version only with its own readings, and keep the old series as history rather than splicing the two together.

How to Write Tracked Prompts That Mirror Real Buyer Questions

A tracked prompt should be a full, neutral question that a specific buyer would type. Build it from persona, situation, constraint and ask. Leave the brand out of category prompts, run each draft several times in a clean session before it enters the set, and version every edit so wording changes never break the trend.

What a tracked prompt is for, and why it is not prompt engineering

Keyword tracking follows a head term with a search volume. Prompt tracking follows a question tied to a topic and an intent. The unit changes, so the method changes: you are not checking a position for a string, you are checking whether an engine names your brand when a buyer asks in their own words.

Group the set by topic or category, and tie each topic to a business goal: a product line you want recommended, a market you are entering, a use case you want to be known for. A prompt with no goal behind it adds noise to the report and costs runs.

Prompt engineering has a different aim. It adds instructions to get a better answer: a role, context, examples, an output format. OpenAI's prompt engineering guide, consulted on 15 October 2026, describes one such technique: "Few-shot learning lets you steer a large language model toward a new task by including a handful of input/output examples in the prompt, rather than fine-tuning the model." Steering is the point when you want a good answer. A tracked prompt needs a representative answer, the one a buyer would get. Our inference, not a statement from OpenAI: a role, examples or a required format inside a tracking prompt shape the answer, so the result stops reflecting what your buyers see.

The same logic applies to prompting frameworks, the lists of letters or numbered elements that describe a good prompt. Keep the elements a buyer would supply, persona and context, because buyers do describe themselves. Drop the elements that steer the output: assigned roles, examples, tone and format instructions.

Where buyer wording comes from, and why full questions beat keywords

Real buyer wording already sits in material you can read: sales call notes, support tickets, community threads, long queries in Search Console and People Also Ask boxes. For sourcing in depth, follow a prompt research workflow built on free data sources.

Favour quality over volume: a short list of realistic questions tells you more than an exhaustive grid of keyword permutations.

Write full questions. People address chat assistants with a request, usually carrying some context and a constraint, rather than with a keyword string. Run "crm small agency" in a chat engine and you measure a query almost nobody sends there. "Which CRM fits a small design agency that bills by project?" measures one buyers plausibly do send. Treat this as a reasoned expectation about chat behaviour, not a measured share.

The details in a question also shape retrieval. Google's "AI Features and Your Website" documentation, consulted on 15 October 2026, states: "Both AI Overviews and AI Mode may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response." Our inference: when a buyer mentions project billing, a region or a team size, those details can become subtopics of the background searches. They change which sources are retrieved, and so which brands are cited. Remove them and you test a different retrieval path.

The anatomy of a tracked prompt: a reusable formula and neutral rewrites

The formula: persona, situation, constraint, ask

Use a fixed template: [persona] + [situation or job to be done] + [constraint] + [ask]. The constraint is whatever narrows the choice for a real buyer: budget, team size, region or industry. The fields encode the grouping. Situation and ask tie the prompt to one topic, and the ask sets one intent category, such as learning, comparing or evaluating. Tag both, so every prompt sits in exactly one topic and one intent.

Keep the formula tight: one question per prompt. A draft that asks for options, pricing and setup advice at once returns an answer you cannot score against a single intent. Split it or cut it.

Filled-in examples from different sectors, both illustrative:

Commercial: "I run operations at a small logistics company in Spain. We need to track delivery exceptions without hiring a developer. Which tools suit a team with no in-house engineers?"
Publisher: "I edit a regional news site with a small newsroom. We want more newsletter sign-ups without adding a paywall. What approaches work for local publishers with limited staff?"

Leading versus neutral wording, with before and after rewrites

These wording traps bias results:

Superlatives. "The best" or "top" tends to push the engine toward a ranked list of the most established names.
A brand named in a category prompt. Yours or a competitor's: the named brand appears because you put it there.
Presupposed answers. "Why is Acme the right choice for agencies?" assumes the conclusion.
Double questions. Stacked asks blend intents, as described above.
Output instructions. "Answer in a table" or "rank them" fixes the shape of the answer, something buyers rarely ask for.

Every example in this table is illustrative, and Acme is a placeholder brand.

Wording type	Illustrative example	What it actually measures	Keep or rewrite
Keyword string	"crm small agency"	A query almost nobody sends to a chat assistant	Rewrite
Leading question	"What is the best CRM for agencies?"	How the engine ranks familiar names under a superlative	Rewrite
Branded question	"Is Acme CRM a good fit for agencies?"	What the engine says about a brand you named	Keep, report separately
Neutral buyer question	"Which CRM fits a small design agency that bills by project?"	Whether the engine recommends you unprompted	Keep

Before and after rewrites, both illustrative:

Before: "Top project management tools for agencies, ranked in a table." After: "I manage client work at a small agency. Which project tools handle client approvals without paying for extra seats?"
Before: "Why is Acme the most reliable payroll option for startups?" After: "Which payroll options suit an early-stage startup hiring across several European countries?"

Branded, category and comparison prompts: balancing the set

Each prompt type answers a different question about your visibility:

Informational prompts ask how to solve a problem or understand a topic. They show whether engines treat you as a reference before a buyer has a shortlist.
Comparison prompts ask for options or alternatives within a category. They show whether you make the shortlist and who sits beside you.
Branded evaluation prompts name your brand and ask about fit, pricing or reputation. They show what engines say about you, and whether it is accurate.

Keep branded prompts apart from non-branded category prompts and report them on separate lines. Naming the brand guarantees a mention, so mixing them inflates visibility and hides whether engines recommend you unprompted.

Aim for a focused set tied to business goals rather than an exhaustive one. Published recommendations on prompt counts and type mixes conflict, and they rarely show the evidence behind them. Size the set from your own constraints instead: the number of topics you need to report on, multiplied by the runs each prompt needs before its share can be read.

Persona and scope modifiers create legitimate segments, not duplicates. "Which CRM fits a small design agency?" and "Which CRM fits an in-house marketing team at a bank?" describe different buyers, and answers may differ with the person described and with what the engine holds about the user. Write persona wording explicitly and keep it fixed, so a change in the answer reflects the engine rather than your phrasing.

Variants, follow-ups and languages without breaking comparability

How many phrasing variants per prompt, and how to report them

Buyers phrase the same need in different ways, so a single wording can overstate or understate your presence. Add a small, fixed number of phrasing variants to each parent question and keep that number constant over time. Tie every variant to its parent prompt and topic so grouping by topic stays intact. Report the parent as one aggregated line, with the spread across variants shown beside it, rather than as separate prompts that dilute the set.

Multi-turn prompts and follow-up questions

Chat engines generally carry earlier turns of a conversation into later answers, so a follow-up asked in an old thread is a different test from the same follow-up asked fresh. When the buyer journey you want to track is a sequence, such as a category question followed by "which of these works for a team without engineers?", script the whole sequence and run it in a new conversation each time. Track the sequence as one prompt with its own ID.

Translating and localising tracked prompts

A literal translation measures how a translator writes, not how a local buyer asks. Write the market-language prompt from native buyer wording, collected the same way as in English. Keep a shared parent ID across languages, and compare markets on the parent, not on the wording, because the phrasing will legitimately differ.

Naming, tagging and change logs

Give every prompt a stable ID, for example crm-agency-fit, and a version suffix that increments with each rewording. Tag it with topic, intent, persona and language. Keep a change log that records the date of each edit, what changed and why. A reworded prompt starts a new series. Never splice its readings onto the old line: a shift at the edit date would have no clean explanation.

Test a draft prompt before it enters the tracked set

AI answers are non-deterministic. SparkToro's research report, published in January 2026 and consulted on 15 October 2026, reports a less than 1 in 100 chance that ChatGPT or Google's AI gives the same list of brands in any two responses to the same prompt. A single run proves nothing about a draft, in either direction.

Test under conditions you can repeat: a logged-out or temporary session, memory and personalisation turned off, the same engine and model settings each time, and a fresh conversation for every run. Saved context about you can colour an answer; a clean session keeps it out of the reading.

Keep a draft only if it passes every check:

The answer addresses the intent the prompt was written for.
The category's real competitors appear, a sign the engine read the prompt as intended.
The answer is not a refusal or a request for clarification.
Results are stable enough across several runs to read a share.

If a check fails, rewrite and test again. If the rewrite fails too, drop the prompt.

Read results as a share over many runs, never as a yes or no from one. Treat a small move between consecutive readings as noise unless it persists. For run counts, see how many runs it takes before a competitor gap is real.

Once a prompt enters the set, give it time before drawing conclusions, and set a regular review date. At each review, prune prompts that keep failing the checklist or no longer match how buyers talk.

Where to start this week

Work through these steps in order:

Collect. Pull a short list of real buyer questions from sales notes, support tickets or Search Console.
Rewrite. Put each one into the formula: persona, situation, constraint, ask.
Neutralise. Strip superlatives, presuppositions and brand names from category prompts.
Vary. Add a small number of phrasing variants per parent question.
Test. Run each draft several times in a clean session and apply the acceptance checklist.
Lock. Assign IDs, versions and tags, and open a dated change log.
Schedule. Set a review date, then read about automating AI visibility checks on a schedule to run the locked set.

Keep branded and category prompts on separate report lines from the first reading.

Before you lock your first prompt set, run the free Namedrop scan to see how ChatGPT, Perplexity, Gemini and Claude currently answer about your brand.