How to choose prompts for AI brand monitoring

Why a prompt is not a keyword

A keyword is an abbreviation. Someone who types "help desk online store" into a search engine knows a list is coming and that the filtering is their job. A prompt is the question or instruction sent to an AI, and someone who asks an AI hands over the filter with the question: they describe the situation, say what cannot be missing, sometimes ask for the format of the answer.

Google measures that difference in its own data. In April 2025 the company reported that AI Mode queries were, on average, twice as long as traditional Search queries; in May 2026, its review of the product's first year in the United States put the average AI Mode query at triple the length of a traditional search. Both numbers cover a single product, and the 2026 comparison covers the U.S. market. OpenAI, Anthropic and Perplexity have published no equivalent figure; on other platforms the direction likely repeats, and the direction is what matters here.

The content of those questions is different too. An NBER working paper from September 2025, written by researchers from OpenAI and universities, estimated that about 49% of messages sent to ChatGPT are "Asking": the person is seeking information or clarification to inform a decision. The sample is a random selection of messages from the consumer plans (Free, Plus and Pro) between May 2024 and June 2025, classified automatically by an AI model. "Asking" covers any information-seeking, not purchase decisions in particular, but it is enough to show that a large share of what reaches an AI is already a decision in progress, with context and constraints built in.

That changes what you can measure. Three consequences weigh on prompt selection.

The first is that the answer changes with the phrasing. "Which help desk should I use?" and "Which help desk suits an online store with five support agents that handles orders through email and live chat?" produce different lists, with different brands, because the second question carries criteria the AI uses to filter. Google's experiences add another layer: 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. Your prompt is therefore a bundle of sub-questions you never see, and every constraining word can open or close a retrieval path.

The second is that the same question, asked twice, can receive different answers. Thinking Machines showed that a thousand runs of the same prompt on the Qwen3-235B-A22B model, at temperature zero, produced 80 distinct completions, an effect of inference batch size shifting with server load. In OpenAI's API, outputs are non-deterministic by default, and determinism can be affected even with its reproducibility parameter. For prompt selection the consequence is direct: near-identical variations of the same question do not measure your presence, they measure the model's variance. Collecting the same prompt again on different days does that job better, without inflating the portfolio.

The third is what is missing. There is no ranking position, no exact match and no search volume for a prompt: no platform publishes how often users ask a specific question to its AI. At most there is aggregate data about query behavior, such as the average length and the growth by task type that Google released in 2026, or, in Microsoft's case, a sample of the phrases the AI used to retrieve content it cited, which are not the questions users typed.

Anyone coming from SEO has to trade the logic of "terms people search for" for the logic of a sample. You choose prompts the way a researcher chooses survey questions: to represent the decisions that matter to the business, not to reproduce the most popular strings. Keyword volume can indicate that a topic matters; it does not indicate how the question is put to an AI.

Start from a buying situation, not from a list

The example is hypothetical and runs through the end of the guide. Brand A sells customer support software, a help desk, to small online stores in the United States. The typical buyer runs operations at a store with a support team of three to five people, receives orders and complaints through email, live chat, social media messages and marketplace inboxes, and wants to stop losing conversations. The tracked competitors are Brand B and Brand C.

What makes the situation yield prompts is the sequence of decisions the buyer moves through. Each moment has a different question, and each question measures something different about your presence.

Journey moment	What the buyer is trying to resolve	Example prompt
Category discovery	Doesn't know the category exists yet; describes the problem	"My online store gets support requests through email, Instagram DMs and marketplace messages, and my team keeps missing them. What kind of tool fixes this?"
Comparison	Knows the category; wants options	"Which help desk platforms work for a small US ecommerce store that runs on Shopify?"
Fit	Tests whether the category suits their size and context	"Is a help desk worth it for a five-person support team, or is it overkill at that size?"
Objection	Fears losing something in the switch	"Is it worth replacing a shared Gmail inbox with a help desk? What do you lose in that move?"
Price	Wants the order of magnitude before talking to sales	"How much does a help desk cost per agent per month for a small team? Are there options with a free plan?"
Implementation	Wants to know the effort and risk of starting	"How long does it take to set up a help desk for a small store and migrate the existing support history?"

Each moment appears here with one prompt; in the portfolio worked out below, it gets two to six, with different constraints, spread across the topics. What does not change is the origin: every prompt answers a decision someone would make before buying from Brand A.

Each row measures something different. The discovery prompt may get an answer with no brands at all, only tool types; it measures whether the category is framed in a way that leads to your product, and only starts measuring presence once the AI names options. The comparison prompt is where brands appear most, and it is the one that most resembles the SEO query "best help desk".

The fit and objection prompts show how the category is described: if the AI says a help desk "is for large companies", Brand A has a positioning problem no mention ranking would reveal. The price prompt tests whether the information circulating about the market is right. The implementation prompt shows whether the documentation you publish is being used.

The real questions come from the same sources you already use to understand the buyer: sales calls, support tickets, objections logged in the CRM, questions that bring people to your site through search, forum threads and reviews. Record the origin of each prompt with a simple label, "observed" when the question actually appeared in one of those channels and "hypothesis" when the team wrote it. Months later that label lets you see whether the prompts you invented produce answers with brands at the same rate as the observed ones. A prompt the team wrote is a bet, not measured demand, and the label exists so the bet can be called in later.

How to build the portfolio: fixed core, exploratory group and topic coverage

A prompt portfolio has two jobs that pull in opposite directions. It has to be stable, so that a comparison between September and October says something about your presence and not about a change of questions; and it has to evolve, because the product changes, the market changes and you discover better questions. The solution is to separate the two jobs into distinct groups with distinct rules.

The fixed core is the set you compare over time. It has one rule: it does not change inside a comparison period. If you compare months, October's core is identical to September's; any change lands on the first day of a new period, with a record. The core should be the larger part of the portfolio and should cover the journey moments and the topics that carry revenue.

The exploratory group rotates. That is where new hypotheses go: a question that showed up in three tickets the same week, a constraint you suspect changes the answer, a topic the product has started to serve. Each exploratory prompt stays for a full period, and at the end it is either promoted to the core, from the following period on, or dropped. The exploratory group's metrics are reported separately, never added to the core.

Topic coverage is the third axis. Assign a topic to every prompt and guarantee a minimum per topic, because the portfolio's aggregate is a weighted average of your own choices: if "implementation" has twelve prompts and "price" has two, total visibility is mostly implementation visibility, and a drop in price barely shows. The topic is also the unit of reading: you will look at answers by topic before you look at the total.

For Brand A, a first portfolio could look like this. Three topics: "choosing a help desk"; "live chat and shared inbox", because for a small store the choice between a shared inbox and a ticketing tool is where the buying decision actually happens; and "implementation and migration". The six journey moments crossed with the three topics form an 18-cell matrix, and the portfolio does not fill every cell. A cell goes in when the buyer would ask that moment's question inside that topic and the answer would lead to a decision no other cell covers; it stays out when the question would be the same as in another topic (the price asked inside "implementation" is the same price as in "choosing") or when the buyer would not ask it there (someone asking about migration has already discovered the category, so "implementation and migration" has no discovery cell). Brand A's core occupies ten cells with 24 prompts:

Moment	Choosing a help desk	Live chat and shared inbox	Implementation and migration	Total
Discovery	2	1	—	3
Comparison	3	3	—	6
Fit	1	2	—	3
Objection	—	2	3	5
Price	2	—	—	2
Implementation	—	—	5	5
Total	8	8	8	24

Prompts in the same cell differ by constraint, not by wording: the three comparison prompts in "choosing a help desk" could be "for a small US store that runs on Shopify", "for a five-person support team" and "with a free plan". Six exploratory prompts, two per topic, bring the total to 30. Each prompt carries four labels: topic, journey moment, origin (observed or hypothesis) and its relation to the brand, which is the subject of the next section.

With or without the brand name: what each group measures

A prompt that names no brand measures consideration: when someone describes the problem or asks for options, does your brand enter the answer? That is the question you cannot answer any other way, which is why the core should be mostly unbranded.

A prompt that names your brand measures something else. "Is Brand A a good fit for a small store?" or "What are the most common complaints about Brand A?" almost always produce an answer containing the brand name, because the name was already in the question, and tools that measure presence by textual detection of the name, Lastfind among them, find that name in the answer. The presence metric is saturated by construction. What these prompts measure is the narrative: whether what the AI says about you is accurate, which objections it repeats, which sources it uses to describe your product. They are valuable prompts, and few, because the answer changes little from one day to the next and each collection costs the same as an unbranded prompt.

Two hybrid types deserve their own group. The head-to-head prompt, "Brand A or Brand B for a store with five support agents?", measures how the AI frames the direct contest and favors whoever is already on the buyer's shortlist. The competitor-branded prompt, "What are the alternatives to Brand B?", is unbranded from your point of view and branded from theirs; it measures whether you appear as an alternative when the buyer already knows the competitor. Classify it as "competitor", not as "unbranded", so it does not inflate the consideration group.

Mixing these groups into one number destroys the reading. A portfolio of only unbranded prompts shows whether you are remembered but not what is said about you; a portfolio of only branded prompts shows the narrative but never consideration. Mixing both into the same visibility figure produces a percentage that rises when you add branded prompts and falls when you remove them, with nothing having changed in the answers. Report each group on its own.

How many prompts: the arithmetic that decides

The number of prompts is not a standalone choice; it is the result of a multiplication. What you will produce per period is prompts × platforms × collections, and what limits that product is the cost of each answer and someone's capacity to read what the tool detected.

The cost is concrete. Every collected answer is a paid call to a provider, and tools pass that on as prompt limits, platform quotas or monthly credits. On Lastfind, the free Cloud plan accepts 10 prompts in 1 topic, covers only ChatGPT and collects each prompt once a week, on Mondays; a portfolio like Brand A's 30 prompts does not fit it. The independent installation (AGPL-3.0-only) collects daily on your own provider account, with capacity set by you and by your provider and infrastructure costs. Other tools have other limits, but in any of them daily collection of a large portfolio across many platforms has a price, and it belongs in the decision before the enthusiasm.

Reading capacity is the most ignored limit. Automatic mention detection reads every answer; a person does not. In the hypothetical situation, Brand A's 30 prompts, on two platforms, collected daily through a 30-day month, generate 30 × 2 × 30 = 1,800 answers. Nobody reads 1,800 answers; somebody reads a sample.

If the rule is to read one answer per prompt and per platform each week, that is 30 × 2 × 4 = 240 answers a month; at two minutes per answer, 480 minutes, or 8 hours of reading per month. Extending collection to six platforms triples everything: 5,400 answers collected, 720 read under the same rule, 24 hours of reading. The arithmetic is hypothetical and the time per answer is a guess; redo it with your own numbers. What it shows is that platforms multiply cost as much as prompts do, and that a portfolio is only truly large if someone reads what it produces.

There is a trade-off that favors smaller portfolios at the start. Ten prompts collected every day produce, per platform, 300 observations in a month; a hundred prompts collected once produce 100. Because the same question gets different answers across runs, what gives a period-over-period comparison its confidence is the number of observations per prompt, not the number of prompts. Start with a portfolio you can read, collect often, and expand once you know which prompts produce answers with brands and which only produce silence. The final count is a consequence of how many topics carry revenue, which platforms your buyer uses and how much reading time you set aside, and no industry range replaces that arithmetic.

How to phrase each prompt

Phrasing decides what the answer will contain. Six criteria solve most of the problems.

Neutrality. The prompt cannot carry the answer. "Why is Brand A the best help desk for small stores?" pushes the AI to justify a premise; "Which help desks suit a small online store?" leaves the list open. Neutrality is not avoiding the word "best": "what is the best help desk for a small store" is a question buyers ask, and it should be monitored as such. The problem is presupposing who wins.

Specificity, one constraint at a time. "For a five-person team", "that integrates with Shopify", "under US$50 per agent per month", "with a free plan" are constraints buyers state and that change the list of brands. Use one or two per prompt, in the buyer's language. Stacking five constraints into one sentence creates a question nobody asks and an answer that represents no real situation. If two constraints matter, that is two prompts.

Language and market. Ask in the language your buyer asks in, because the set of brands in the answer changes with the language and with the market implied by the question. "In the US" inside the prompt is part of the question and should be a deliberate, consistent choice across the portfolio, not a decoration on a few prompts. Also check the tool's language limit per platform before building the portfolio: on Lastfind, the Gemini integration requires the project to be in English, and a portfolio in another language does not run on that platform. A platform accepting your language in its interface does not guarantee that the integration the tool uses accepts it.

Time and format. Avoid "in 2026" and "updated" unless your buyer says it; the date ages the prompt and forces a swap that breaks the series. Do not ask for a format, "answer in a table", "list five options", unless the buyer would; the requested format changes how many brands fit in the answer. One question per prompt: "how much does it cost and how long does it take to set up" are two decisions, and the answer will treat one of them in passing.

When the prompt already contains the brand. Write it the way the buyer would, not the way your marketing team would like. "Does Brand A work well for stores that sell on marketplaces?" and "What are the most cited problems with Brand A?" are legitimate branded prompts; "What are Brand A's innovative differentiators?" is brochure language, and the answer will hand the brochure back. Branded prompts go in the narrative group, with their own label.

Length. A real prompt rarely runs past three or four lines. Tools impose technical limits, on Lastfind the CSV import accepts 5 to 1,000 characters per prompt, but the useful limit is the buyer's: if you needed a paragraph to describe the situation, you probably joined two questions.

Google's AI answers call for a specific caution. AI Overviews are only shown when Google's systems determine they add to classic Search and, as such, often don't trigger. A prompt that does not trigger an AI Overview is neither a wrong prompt nor an answer without your brand: it is an observation with no AI answer, and a serious tool records it that way. Do not rewrite prompts to force the block to appear; you would be monitoring the question that triggers the feature, not the one the buyer asks. And because AI Mode and AI Overviews may use different models and techniques, so the responses and links they show will vary, treat the two as separate platforms in the portfolio.

What not to monitor

These prompts take up a slot without producing a decision.

• Questions whose answer never names brands. "What does SLA mean in customer support?" gets a definition, and the presence metric stays at zero forever without saying anything about you. If you want to know how the category is explained, keep one prompt of this kind per topic, labeled "category", outside the presence core.
• Redundant variations of the same question. "Which help desk for a small store", "help desk for a small store, which one to pick" and "recommend a help desk for a small store" measure the model's variance, not your presence, and triple the cost of a single data point. If variation between phrasings interests you, pick one and collect it again.
• Questions with no decision behind them. "What is the future of customer support?" may yield an interesting answer and no action: there is no page to fix, no attribute to document, no source to chase.
• Questions whose answer you cannot change. If the AI recommends tools with a feature your product does not have and will not have, the answer is actionable for the product, not for content; record the gap and remove the prompt from the monitoring portfolio.
• Prompts in your marketing language. Nobody asks an AI for "omnichannel customer experience solutions" when they want to stop missing messages in a shared inbox. The prompt has to reproduce the buyer's vocabulary, even when it is imprecise.
• Prompts with several questions. An answer that touches price, implementation and integration in passing is no use for evaluating any of the three.

Maintenance: how to swap prompts without breaking the series

The portfolio will change, and the risk is not change; it is uncontrolled change later read as a result. Four situations justify a swap: a core prompt went a full period without producing a mention of any brand, on every platform, and was not deliberately included as a category prompt; the topic stopped mattering to revenue; the product changed and the question lost its meaning; or a real question, from sales or support, proved better than a hypothesis sitting in the core.

How you swap matters more than why. Never edit a prompt in place: an edited prompt carries the old text's history under the new text's name, and the series starts comparing two questions as if they were one. Pause or archive the old prompt and create a new one. On Lastfind, pausing keeps the history and the slot, archiving frees the slot, and each prompt keeps its own history; on other tools, confirm that deletion does not erase past answers before touching anything.

Change the core only at a period boundary, never in the middle. Log every change in a simple journal, which can be a spreadsheet: date, prompt, action (created, paused, archived, promoted from exploratory), reason and owner. Version the core with a tag, "core-v1", "core-v2", so any report says which portfolio it refers to. When comparing two periods that straddle a change, compare only the prompts present in both, or recompute the earlier period with the new portfolio; presenting the old total against the new total compares different questions.

The exploratory group is the entry door, and that is its main use: no new question enters the core without having spent a period in exploratory, with its own metrics kept apart. That way the core changes little, in a planned way, and the historical series survives the portfolio's evolution. On Lastfind, new prompts join collection at the next scheduled window (Monday on the free Cloud plan, the next daily window on an independent installation); check the equivalent rule in your tool. The first day of a new period is the moment to import the new version at once, by CSV, rather than adding prompts through the month.

What to demand from a monitoring tool

The method above only works if the tool preserves the distinctions it creates. Before choosing, check whether it lets you organize prompts by topic and by free-form tags, because without that the core, exploratory, branded and unbranded groups collapse into a single list; whether it imports and exports the portfolio, so the core's version exists outside the interface; whether it keeps per-prompt history and lets you pause or archive without erasing past answers; whether it preserves the original answer from every collection, without which a change cannot be explained; whether it states language limits per platform; and whether it treats the absence of an AI Overview as an observation with no answer rather than as an answer without your brand. These criteria apply to any tool, and they are what you should ask for in any demo.

Lastfind organizes prompts by topics and tags, imports by CSV, keeps per-prompt history with pause and archive, stores the original answer and does not count a missing AI Overview as an answer. Its CSV export covers collected answers, not the portfolio with its labels, so keep your own portfolio file. The guide to GEO monitoring develops the process's other requirements.

A protocol for building the portfolio today

Follow the steps in order. Each takes under an hour, except the first.

1. Write the buying situation in one paragraph: who buys, with what team size, with what problem, in which language and market. Everything that follows derives from it.
2. List two to four topics that carry revenue in that situation. If you would not invest content in a topic, do not monitor it.
3. Cross the topics with the six journey moments (discovery, comparison, fit, objection, price and implementation) and mark only the cells where the buyer would ask the question inside that topic and the answer would lead to a decision no other cell covers. Write one to five prompts per marked cell, differing by constraint, and mark each one's origin as "observed" or "hypothesis".
4. Classify each prompt as unbranded, your brand, head-to-head or competitor. The consideration core stays mostly unbranded.
5. Cut by the rules in "What not to monitor": no redundant variations, no questions without a decision, no marketing language, one question per prompt.
6. Choose the platforms where your buyer is and check each one's language limit in the tool.
7. Do the arithmetic: prompts × platforms × collections per month, and the sample somebody will read. Reduce until it fits the plan and the reading time.
8. Separate the fixed core from the exploratory group, version the core and import the portfolio in one go.
9. Open the change journal with today's date and the version 1 prompt list.
10. Mark the first period boundary on the calendar; until then, the core does not change.

At that boundary, promote or drop each exploratory prompt and log the decision in the journal before opening the next period. You will then have something a keyword list never gives you: a set of buying decisions, observed on the platforms where they happen, with answers stored and a record of what was asked. Whatever changes afterward, in the portfolio or in the answers, can be explained rather than merely counted.

Platform documentation and Lastfind limits checked in October 2026.