Guides
How to build a prompt set for GEO: choosing what to track
Size a prompt set from the answers each report needs: about 300 answers per period per topic, which is 9 prompts per topic on four engines run twice a week.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
A prompt set is the fixed list of questions you send to AI engines on a schedule, and its size follows from the answers each reported figure needs rather than from a round number of prompts. cloro's published study on sample size puts about 300 answers per period at roughly ±5 percentage points, so a set that reports four topics separately every month needs about 300 answers per topic per month: nine prompts per topic, on four engines, run twice a week. That is 36 prompts and 1,248 answers a month. Report those same four topics weekly instead and the same precision asks for about 152 prompts, because the period shrank and the answer target did not.
This guide is for a brand team writing its first prompt set, or re-cutting one that produces charts nobody trusts. API for GEO is our editorial pick after hands-on testing by Mathias C., and it serves as the worked example here because it publishes its prompt fields and its unit of billing on the same page. Below: how each vendor counts a prompt, six steps from intent to a frozen list, where real question wording comes from, the arithmetic that sizes the set, what a set costs a month, and the mistakes that break a trend line. MentionBox, which publishes this site, has a commercial link with API for GEO.
What counts as one prompt depends on the vendor
The word "prompt" is a billing unit before it is an editorial decision, and the unit is not the same in two places. Read the row that matches your vendor before you write a single question, because it decides whether a second country, a second engine or a second run is free or doubles your bill.
| API | What one prompt costs | What the vendor's unit includes |
|---|---|---|
| API for GEO | 1 credit per prompt, per model, per run: $0.013, or $13 per 1,000 answers | "One prompt, executed once, against one model, in one locale", enrichment included; creating prompts and reading their stats is not billed |
| OtterlyAI | 1 prompt slot per country, whatever the run count: 15 on Lite, 100 on Standard, 400 on Premium, extra packs of 100 at $99 a month | "Each prompt is tied to a specific country", and "one prompt = one slot, regardless of how many reports it feeds" |
| Promptwatch | 1 response per prompt, per model, per run, against a monthly response quota | its own example: 50 prompts on 4 models daily is 6,000 responses a month, and monitoring pauses when the limit is reached |
| Peec AI | monthly responses counted as prompts × active models × frequency | one topic per prompt, several tags per prompt, and an Active, Archived or Deleted status |
| Gumshoe | 1 conversation per prompt, per model, per persona: 480 included on the $99 Starter plan, $0.10 beyond | a conversation is "one prompt sent to one model, asked as one persona, plus the answer that comes back" |
| Rankscale | 0.25 credits per engine per prompt on Google AI Mode, ChatGPT, Perplexity and Gemini, 1 credit on DeepSeek, 2 on Claude | the engine list changes the price of the same prompt set |
| Conductor | prompt slots per topic: 5 to 500 written by hand, up to 100 generated | cadence is set per topic, daily, weekly or monthly |
| Profound | prompt rows inside a category, on an Enterprise quote | one prompt row carries a language, a topic, tags, regions, platforms and personas |
| SE Visible | 200 prompts and about 30,000 AI responses analysed a month on the $99 plan | about 150 answers per prompt per month at that ratio |
Units and allowances from each vendor's documentation and pricing pages, checked on October 23, 2026, with the API for GEO prompt fields, the OtterlyAI prompt limits page, the Promptwatch response limits page and the Peec AI setup guide read again on November 9, 2026. For API for GEO, the Gemini, Google AI Overviews, DeepSeek and Meta AI coverage, the webhooks, the OpenAPI spec and the official SDKs were confirmed in our hands-on test on October 26, 2026 and are not yet documented on apiforgeo.com.
Two consequences follow for anyone about to write a list.
- A prompt count is not a volume. "We track 100 prompts" says nothing until the engines, the countries and the runs per cycle are attached to it. The same hundred prompts is 1,733 answers a month on four engines run weekly and 12,000 run daily.
- The billing model decides which growth hurts. A slot model like OtterlyAI's charges for a second country and gives away extra runs; a credit model like API for GEO's gives away the slot and charges for every run. Our guide to the cost of tracking 100 prompts on five engines prices that fork across seven vendors.
Step 1: decide what each prompt has to prove
A prompt set is not a keyword list, and the vendors who publish a taxonomy agree on the split. API for GEO's prompt tracking page gives its prompt type field five values, "informational · commercial · comparison · branded · navigational", and lets tags carry "the topics, markets and funnel stages of the project". Peec AI's setup guide uses three intents, informational for someone who "wants to learn or get instructions, not compare or buy", commercial for someone "comparing or researching options before a decision", and transactional for someone "ready to act, buy, or take a specific step". Profound attaches an analysis type to each prompt instead, among visibility, sentiment and accuracy, so the same question can be measured for presence or for how it is described.
Each slice answers a different question, and mixing them without tags produces an average that means nothing:
- Category prompts ("best X for Y") ask whether the engine puts you in the category at all. They name the most brands, so they carry share of voice.
- Use case and constraint prompts ("X for a team of five with no dedicated admin") are where a smaller brand can win, because the answer has to satisfy a condition rather than list the market leaders.
- Comparison and alternatives prompts ("X vs Y", "alternatives to X") show how the engine frames your differences, and they are the ones a competitor's page can swing.
- Problem-stage prompts ("why does my X keep Y") name few brands and return a source list. The value sits in which domains the engine read.
- Branded prompts ("what is X", "is X any good") measure accuracy and sentiment, not competition.
In our view a first set of 36 prompts splits well as nine category, eleven use case, five comparison, five persona or segment, four problem stage and two branded. Peec AI's documented default is more generous to brand terms, at "20% branded and a 25/50/25 intent split"; we keep branded under 10%, because those answers move with your own site rather than with the category, and the budget is better spent on questions where an engine is choosing between vendors. Either ratio works as long as it is written down and tagged, since a share of voice computed across all six slices at once is an average of six different populations.
Step 2: take the wording from places that record real questions
Invented prompts read like keywords. Peec AI's guide puts the difference plainly: prompts "are longer and conversational" and should carry "context and constraints, not just topics". Four free sources give you that wording before any tool does: recorded sales calls, support tickets, your site's internal search log, and the question queries in Search Console.
Two platform reports sit above those, and only one of them gives you question text.
- Bing Webmaster Tools. Microsoft's AI Performance report, in public preview since February 10, 2026, reports total citations, average cited pages, page-level citation activity and grounding queries, defined as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers", across Microsoft Copilot, AI summaries in Bing and partner integrations. Microsoft describes the grounding query data as a sample of overall citation activity, so read it as phrasing evidence rather than as volume.
- Google Search Console. The Generative AI performance report groups data by pages, countries, devices and dates, and by text or multimodal search. There is no query dimension, so Google's own AI report cannot tell you which questions produced your impressions.
After that come the prompt databases, and they are built in three different ways. Observed data comes from conversations the vendor licensed or captured. Derived data comes from search keywords, with a model in between.
| Source | How the questions are obtained | Published price | What to watch |
|---|---|---|---|
| Profound Prompt Volumes | "licensing AI answer engine conversations, cleaning the data, applying probabilistic modeling" for ChatGPT, Gemini and Perplexity | API on an Enterprise quote | US data from January 2025, other regions from January 2026, and coverage varies by region |
| DataForSEO LLM Mentions | aggregated LLM chat data, on a counter showing 420,701,746 prompts when we checked it | $0.10 per task plus $0.001 per row | its ChatGPT mention data is limited to the United States and English |
| DataForSEO AI Keyword Search Volume | volume estimates for keywords, as a task endpoint | $0.01 per task plus $0.0001 per row | a keyword volume, not a prompt volume |
| Ahrefs Brand Radar | Google's People Also Ask corpus plus a semantic fan-out that expands it for "logical completeness rather than behavioral popularity" | subscription plus API units; adding custom prompts consumes none | its impressions weight mentions by Google search volume, which Ahrefs calls "a modeling choice, not a measured relationship" |
| Semrush Enterprise AIO | a prompt base whose responses are "captured from real requests and not via any APIs of LLMs" | Enterprise quote | the AI Visibility Toolkit data is not exposed in the public API documentation |
| LLM Pulse prompt research | prompts generated from your domain and related keywords, localised per country and language | included in the plan, up to 100 suggestions per run | generated suggestions, not observed questions |
Methods, prices and caveats from each vendor's documentation, pricing pages and methodology posts, checked on October 23, 2026 and read again on November 9, 2026.
The cheapest source of engine-authentic phrasing is the engines themselves. The sub-queries a model issues behind your existing prompts come back as a field on most answers, and they are written the way the engine frames your category rather than the way a marketer would. Our explainer on query fan-out lists which APIs return them and at what price, and a fan-out list collected over a month is a prompt-set candidate list for free.
Step 3: test every candidate before it takes a slot
Four checks, in this order, remove most of the dead weight from a first list.
- Does the engine search the web for it? A prompt answered from the model's own knowledge returns no source list and nothing you can influence with a page. Anthropic's web search tool guide states that simple factual queries typically use one to three searches while comparative research "can use 10 or more", and that search is steerable from the system prompt; OpenAI's web search guide states that the model "can choose to search the web or not based on the content of the input prompt". The collection APIs expose the same switch: Bright Data returns a
web_search_triggeredfield, cloro forces search on and offersdisableWebSearch, and DataForSEO's LLM Responses endpoint ships withweb_searchat "default value: false". - Does any brand get named? A prompt that returns a generic explainer cannot produce a share of voice, however central it looks on a whiteboard. In an analysis we ran in October 2026, 7.1% of the answers we collected named no brand at all.
- Could your pages be reachable for it? The documented fan-out examples break a question into subtopics, so a page answering a subtopic is reachable and a brand page is not. If no page of yours could plausibly be read for a prompt, the figure it produces is a competitor report rather than a measurement of your own work.
- Does it belong to a segment you will report? Every prompt must sit in a topic you intend to show on a chart. Prompts that belong to no reported segment pay for answers nobody reads.
Step 4: size the set from the answers each report needs
Two formulas carry the whole decision, and the second is the one teams skip:
Answers per segment per period = prompts in the segment × engines × runs per period
Prompts per segment = answers you need ÷ (engines × runs per period)
The answer target comes from the precision you are claiming. cloro's study puts about 300 answers per period at roughly ±5 percentage points, and about 294 answers on each side to support a claimed move from 20% to 30% visibility. So the reporting cadence, not the size of the company, sets the prompt count:
| Reporting period | Collection cadence | Answers per prompt per period, 4 engines | Prompts per topic for 300 answers | Four topics |
|---|---|---|---|---|
| month | twice a week | 34.7 | 9 | 36 |
| month | once a week | 17.3 | 18 | 72 |
| month | once a day | 120 | 3, raised to 8 for coverage | 32 |
| week | twice a week | 8 | 38 | 152 |
Our arithmetic on the ±5 point figure published in cloro's study of August 14, 2026, counting 4.333 weeks and 30 days in a month. The coverage floor of eight prompts per topic is our own rule: below it, a topic cannot hold the five intent types above.
Three readings of that table are worth keeping.
- Weekly reporting is expensive in prompts. The same precision costs four times the set, because a week holds a quarter of the runs a month does. Teams that want a weekly chart usually want a monthly figure with a weekly collection behind it, which is a different and much cheaper thing.
- Repetition and breadth cost the same and tell you different things. Thirty-six prompts twice a week and 72 prompts once a week both produce 1,248 answers a month at four engines. Only the repeated design tells you whether a single prompt is stable, which is what you need the day someone asks why one question lost you three positions.
- Every split you report multiplies the need. The 300-answer target applies to each segment you show on its own. Four topics crossed with four engines is sixteen segments, and sixteen segments at 300 answers each is 4,800 answers a period. Decide which cuts are reported and which are diagnostic before sizing, because only the reported ones have to be powered.
Step 5: price the set before you freeze it
Thirty-six prompts on four engines in one country is 144 answers per cycle. Three cadences, on the two billing models a brand team usually chooses between:
| Cadence | Answers a month | API for GEO, $0.013 per answer, metrics included | DataForSEO standard queue, $0.0012 per answer, counting excluded |
|---|---|---|---|
| once a week | 624 | $8.11 | $0.75 |
| twice a week | 1,248 | $16.22 | $1.50 |
| once a day | 4,320 | $56.16 | $5.18 |
| twice a day, two countries | 17,280 | $224.64 | $20.74 |
Our arithmetic on the rates each vendor publishes, checked on October 23, 2026. DataForSEO's rate applies to its LLM Scraper standard queue and needs a $50 minimum payment, and its response carries no mentions, position, share of voice or sentiment. A slot plan prices the same set differently: OtterlyAI's $189 Standard plan covers 100 prompts in one country each, and Promptwatch's $245 Professional plan covers 150 prompts and 18,000 responses.
API for GEO is the route we use for this kind of set, because the prompt layer and the metric layer sit in the same object: POST /v1/tags creates the topics, markets and funnel stages, POST /v1/queries upserts each prompt with its type, language, country and tags, POST /v1/jobs/execute runs them with a models list and an executionsPerQuery count, and POST /v1/queries/stats reads back runs, visibility, average position, entities and sources per prompt. It scores 9.0 of 10 in our ranking of 55 APIs, the highest of the set, it covers the ten engines we track, with ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok listed on apiforgeo.com and Gemini, Google AI Overviews, DeepSeek and Meta AI confirmed in our hands-on test on October 26, 2026, and its terms allow resale, which decides whether an agency can put the numbers in a client report. Said plainly, the trade-offs are a price per answer about eleven times DataForSEO's, a key obtained through a form rather than at sign-up, and those four engines plus the webhooks, the OpenAPI spec and the official SDKs resting on our test rather than on the vendor's documentation. To price your own list, request a key on apiforgeo.com: it comes with a sandbox project.
The competing strengths are published too. OtterlyAI's slot model is the cheapest way to run one list very often, since extra runs cost nothing. Promptwatch publishes the clearest quota arithmetic in the market and advises trimming models first, noting that "the number of models is the quickest lever on your usage". Conductor generates up to 100 prompts per topic for you. DataForSEO remains the floor at $0.0012 an answer for any team with a developer to write the counting, and our ranking for enterprise brand teams scores the whole set on coverage and presentation-ready metrics.
Step 6: freeze the wording, then version the changes
A prompt set earns its value in month three, and only if nothing moved. The APIs are built on that assumption: Peec AI gives a prompt an Active, Archived or Deleted status, and Profound exposes an endpoint that enables or disables prompts. None of them offers an edit that preserves the series, because there is no such thing.
Four rules keep a set comparable:
- Never rewrite a prompt in place. Add the new wording as a new prompt, let the old one run one more period alongside it, then archive it. The overlap period is what lets you show the two series meet.
- Tag before you collect, not after. One topic per prompt and as many tags as you need, the way Peec AI splits "folder-like structure for navigation" from "flexible categorization for analysis". Retro-tagging works only where the vendor stores the raw answers.
- Log every change with its date, next to the chart. A set that gained four prompts in week six explains a step in the curve that otherwise looks like an engine update.
- Review quarterly, not monthly. A quarter is the shortest window in which a prompt's own numbers can justify keeping or dropping it.
Peec AI's documentation makes one claim worth acting on here: "Exact phrasing doesn't matter much. Semantically similar prompts will lead to very similar results over time." cloro's run-to-run measurement points the same way, since two runs of the identical prompt on the identical model repeated only about 40% of their cited domains. Taken together, paraphrases are the worst use of a slot: three wordings of one question cost three times as much as one wording run three times, and the second design is the one that gives you an error bar.
Locales and personas, the two multipliers
Both of these double a set quietly, and the vendors handle them in incompatible ways.
- Locales. OtterlyAI ties each prompt to a country, so the same question in the US and the UK takes two of your slots. API for GEO puts the locale inside the credit, so a second country doubles the answer count rather than the prompt list, and Profound attaches regions to the prompt row. Translating a prompt is not the same as localising it: the question a French buyer types is rarely the French translation of the English one.
- Personas. Gumshoe builds the persona into its billing unit, where one conversation is one prompt, one model and one persona, and its $99 Starter plan covers 480 conversations as 8 personas × 10 prompts × 6 models. Profound carries personas as a field on the prompt. Ahrefs takes the opposite route and states that it "enters those questions into the web version of each AI chatbot using the default model", with no personalisation and no pre-prompt.
In our view a persona belongs in the prompt text rather than in an account setting whenever you might change vendors. A prompt that reads "I run a 40-person ecommerce brand and need X under $500 a month" is reproducible anywhere; a persona configured in a dashboard is a setting you cannot export, and it quietly changes what the engine was asked.
Five mistakes that waste a prompt set
- Sizing on a round number. Twenty-five prompts across four engines run once is 100 answers, which carries a margin of error wide enough to hide any move you would act on. The number comes from the segments you report, not from a template.
- Filling the list with paraphrases. Five wordings of "best X" measure one question five times and look like coverage on a slide. Spend those slots on five different buyer situations, or on five extra runs.
- Tracking prompts the engines answer from memory. No search means no sources and nothing your pages can reach. Screen for it once, cheaply, rather than discovering it in month two.
- Letting the set grow without a budget line. Each prompt added to a daily four-engine set is 120 answers a month. Twenty additions is 2,400 answers, and on a quota plan like Promptwatch's the monitoring simply stops when the response limit is reached.
- Mixing branded and unbranded prompts in one figure. Branded prompts are easy wins that lift an aggregate visibility score without telling you anything about the category. Tag them and report them apart, whatever ratio you chose in step 1.
Our view, and a 36-prompt starter sheet
A usable first set is small, derived rather than guessed, and frozen. The one we would write for a brand team reporting four topics monthly, on four engines, twice a week, costs $16.22 a month on API for GEO's rate card and breaks down like this:
- 9 category prompts, one per product line and per buyer type, in the wording your sales calls use.
- 11 use case and constraint prompts, each carrying a real constraint such as team size, budget ceiling, integration or regulation.
- 5 comparison prompts, built from the competitors that actually appear in your answers rather than the ones in your pitch deck.
- 5 persona or segment prompts, with the persona written into the question text.
- 4 problem-stage prompts, kept for their source lists rather than their brand counts.
- 2 branded prompts, read for accuracy and sentiment only.
Then collect for a quarter before changing anything. Our guide on tracking a brand in ChatGPT with an API covers the collection and counting steps that follow, GEO metrics explained covers the denominators you have to fix before the first report, and the cost of tracking 100 prompts on five engines prices the set once it grows. The prompt list is the only part of a GEO programme that nobody else can write for you, and it is also the only part that gets reviewed by accident instead of on purpose.
Frequently asked questions
How many prompts should I track for AI visibility?
Work backwards from the figures you want to report. cloro's published study of August 14, 2026 puts about 300 answers per period at roughly ±5 percentage points, so one topic reported monthly, on four engines run twice a week, needs about nine prompts: each prompt produces 34.7 answers a month. Four topics is about 36 prompts. Reporting weekly instead of monthly takes the same precision to about 38 prompts per topic, and no set below roughly eight prompts per topic can cover the five intent types vendors distinguish.
How many times should each prompt be run?
More than once, and the same number of times every cycle. cloro's study measured a mean overlap of 0.40 between the cited domains of two runs of the same prompt on the same model minutes apart, and puts about 294 answers on each side to support a claimed move from 20% to 30% visibility. Evertune states that it samples each prompt 100 times across every AI model. Promptwatch's own documentation gives the arithmetic as prompts × models × runs, with 50 prompts on 4 models daily producing 6,000 responses a month.
Where do I find the prompts people actually ask an AI engine?
Microsoft publishes a free sample: the AI Performance report in Bing Webmaster Tools, in public preview since February 10, 2026, shows grounding queries, "the key phrases the AI used when retrieving content that was referenced in AI-generated answers". Google's Generative AI performance report has no query dimension at all, only pages, countries, devices and dates. Paid prompt databases exist on both sides: Profound's Prompt Volumes is built by licensing AI answer engine conversations and modelling them, and DataForSEO's LLM Mentions bills $0.10 per task plus $0.001 per row.
Should branded prompts be in a GEO prompt set?
A small share of it. Peec AI's setup guide uses a default of 20% branded prompts with a 25/50/25 informational, commercial and transactional split; in our view 5% to 10% is enough, because a branded prompt measures whether an engine describes you correctly rather than whether it picks you. API for GEO's prompt tracking page carries both as values of its prompt type field, informational, commercial, comparison, branded and navigational, so the split can be reported separately whichever ratio you choose.
Can I change the wording of a prompt once tracking has started?
Not without breaking the series it feeds. The APIs reflect that: Peec AI gives prompts an Active, Archived or Deleted status, and Profound exposes a status endpoint that enables or disables a prompt rather than rewriting its history. Add the new wording as a new prompt, keep the old one running for one more period, and record the date of the change next to the chart.
How much does a 36-prompt set cost per month?
At 36 prompts on four engines, one country, a cycle is 144 answers: 624 a month run weekly, 1,248 run twice a week and 4,320 run daily. On API for GEO, our editorial pick after hands-on testing, that is $8.11, $16.22 and $56.16 at $0.013 per credit with metrics included and no monthly minimum. On DataForSEO's standard scraping queue at $0.0012 an answer it is $0.75, $1.50 and $5.18 after a $50 minimum payment, with the brand counting left to you. Prices checked on October 23, 2026.