DeepSmith

Nov 26 · AEO & AI Visibility

19 min read

How to Estimate Prompt and Engine Needs Before You Buy a Plan

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Dark abstract cover with a grid of squares linked by thin lines to a stack of three layered cards, with the centered words Size Your Plan First.

Most teams pick an AI visibility plan by looking at the price and the biggest prompt number they can afford, and then try to fit their questions into it. That works backwards. This guide walks you through AI visibility plan sizing the other way around: you work out how many prompts and which engines your team really needs, and then you check that number against what a plan allows. It is written for marketing leads who are about to compare plans on the AI visibility tools market and want a number they can defend. By the end you will have a simple inventory of your prompts, your markets, your competitors and your engines, plus a short checklist to take into a trial or a vendor call.

You will need a spreadsheet, a few notes from your sales and support teams, and about two hours. This guide only sizes your own requirement. It does not explain how vendors build their pricing models, and it treats the list you end up with as a deliberate sample of the questions your buyers ask, not a full count of everything people type into AI tools.

Step 1: Set the business and measurement boundary

Start by deciding what the tracking is supposed to tell you. Pick one product line, brand, campaign or market to begin with. Write down who the buyers are, which markets matter, and how long you will watch before you judge the results. A quarter is a common choice, but use whatever fits your planning cycle.

Then gather the real language your customers use. Sales call notes, demo recordings, support tickets, customer interviews, reviews and your own site search are all good sources. Add the company priorities that sit on top of those, such as a product launch, an objection that keeps coming up, a competitor comparison you keep losing, or a reputation risk. If you want more help with how to find the prompts worth tracking, that guide goes deeper on discovery.

Analytics can support this, but it should not lead it. In GA4, the Traffic acquisition report sits under Reports, then Acquisition, then Traffic acquisition. You can search it for a known source and look at sessions, engagement rate and key events. GA4 also has an AI Assistant default channel group that groups sessions referred from sources such as ChatGPT, Gemini, DeepSeek, Copilot or Grok. It leaves out Google AI Overviews and AI Mode, so the picture is partial. Google Search Console has its own generative AI performance reports for Search and Discover, with impressions, pages, countries and performance over time. Treat all of this as a hint about where to look, because referral traffic only shows visits that reached your site. It does not tell you how often AI names or cites your brand, and it can't tell you about the engines that send no traffic at all.

You are done with this step when you can write down the product or brand in scope, the buyer audiences, the priority markets, the business questions you want answered, and where your evidence came from.

Common mistake: starting with the largest prompt allowance on a pricing page, or with a keyword export, and then trying to invent a business question to match it. Start with the question and let the number come out of it.

Step 2: Build a topic-by-buyer-stage map

Make a list of the topics and use cases you want to be known for. For each topic, think about the three buyer stages: Awareness, Consideration and Decision. Awareness questions ask what a problem is or how to solve it. Consideration questions ask which approaches or product types fit a use case. Decision questions compare named options, ask whether a product suits a situation, or push on a buying objection. Include plain-language versions of the problem as well as the category words your company uses, because buyers often describe the problem long before they know your category name.

Put this in a grid with topics down the side and stages across the top. Not every topic needs a prompt at every stage, and that is fine. If you have no real Decision question for a topic, leave the cell empty and write down why, rather than inventing a question to fill the gap. A neat full grid looks tidy, but it quietly adds prompts that you will pay for and never act on.

Keep the grid small enough to read. One prompt should stand for one intent and sit in one topic and stage cell. A prompt deserves its own slot when something real about the audience, the use case, a constraint or the outcome could change the answer, the recommendation, the sources or the reasoning behind it.

You are done with this step when every candidate prompt has a topic, a stage, a note about the audience or use case, and one sentence on why you would watch it.

Common mistake: treating a complete grid as the goal. Empty cells are valid when that decision does not matter for the product or is not a priority right now.

Step 3: Choose representative prompts and remove the redundant ones

Now write one candidate prompt for each important buyer decision in the grid. A healthy mix usually includes:

  • Category questions from people who do not know your brand yet.
  • Evaluation questions from buyers who already know it.
  • Competitor and comparison questions, where the alternative is relevant.
  • Use-case, objection and reputation questions that come from real customer concerns.

If you are not sure how those groups differ, the guide on branded, unbranded and competitor prompts walks through each one.

Next, cluster the prompts that ask the same underlying question. Different wording alone does not always justify a separate slot. Keep a variant when the audience, the market, a constraint or the buying situation could change what the AI recommends or cites. A general question about the best option for a small team is not always the same question as one that adds a regulated-industry requirement, but it is only different if the added requirement changes the answer. Treat that as a rule of thumb, not a promise that every variant will return something new.

A good order for deciding what stays is business relevance first, then buyer-stage relevance, then whether you have a meaningful offer for the topic, then whether competitors show up where you do not, and last whether the prompt adds coverage you do not already have. A competitor showing up is a reason to look at a relevant prompt. It is not a reason to track every competitor against every topic.

Split what survives into two lists. The Core list is the stable set you will compare over time. The Explore list holds new questions and experimental wording. When an Explore question keeps surfacing something your business cares about, promote it into Core. When you retire or replace a Core prompt, label the change and keep the old version, so a different list is never mistaken for better visibility. If you want a longer walk-through of setting this up, see how to set up a repeatable prompt set.

On how big the list should be, the honest answer is that no source gives one right number. One Semrush guide suggests a starting set of around 25 prompts, and a DeepSmith guide on how many AI prompts to track suggests beginning with 15 prompts across two engines and says a working baseline may land between 30 and 80. Those are editorial suggestions, not averages and not minimums, so use them as a sanity check on your own count and nothing more.

You are done with this step when near-duplicates are grouped, each prompt you keep tests a distinct decision, and the Core and Explore lists are separate.

Pro tip: a list that covers a handful of distinct buyer problems is more useful than a much longer list of paraphrases of one category question.

Step 4: Add the markets, languages and competitor comparisons that matter

Write down the exact markets and languages you need to compare. If the wording, the language or the market changes, record it as a separate version in your inventory and keep the earlier history with it. Mixing different market and language groups into one trendline makes the trend hard to trust.

For sizing, the safe move is to count each required market and language version as its own planned item, unless the vendor tells you in writing that it counts regional variants differently. A region feature does not mean unlimited prompt capacity, and one prompt entered once is not automatically reported separately for every country and language. This matters because it can double or triple your count before you have added a single new topic.

Competitors need the same restraint. Add a competitor prompt only where it tests a real buying decision, such as a named alternative, a category comparison, or an objection your sales team hears. Setting up a competitor for competitive analysis does not mean you need a separate prompt for every topic. If a prompt changes the matchup in a way that changes the question you are asking, though, it can be a distinct test. If you are still deciding who belongs on the list, the guide on how to choose which competitors to track can help.

A note on DeepSmith here, since you will probably compare it. The published plan comparison shows one geographic region for Pro, Grow and Scale, and custom multi-region tracking on Enterprise. The product pages do not specify supported countries or languages. If you need several markets, ask DeepSmith to confirm the exact region and language setup before you decide that a self-serve plan is enough.

You are done with this step when every planned row has a market and a language, or is marked as a single-market prompt, and the competitor comparisons you need are listed on their own.

Common mistake: multiplying the whole set by every competitor, or assuming one regional setting gives you the same coverage as separate local tracking.

Step 5: Choose engines using audience evidence and required coverage

List the engines your buyers actually use, and then split them into must-haves and nice-to-haves. Lean on customer and sales evidence, your market plans and the reporting question you set in Step 1. If you want a longer method for ranking them, see which engines to track first. Keep the same prompt wording on every engine when you want a like-for-like comparison, because an engine-specific rewrite is a different test and not a copy of the first one. The guide on tracking the same prompts across several engines shows how to do that.

Do not choose engines from referral traffic alone. Some engines send little traffic and still shape what buyers believe.

DeepSmith covers ten engines in total: ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, Google AI Mode, Grok, Meta AI, Microsoft Copilot and DeepSeek. Which of them you get depends on the plan, as this table shows.

PlanTracked promptsEnginesRegionsMonthly priceMonthly price billed annually
Pro50ChatGPT1$99$80
Grow100ChatGPT, Perplexity1$199$160
Scale200ChatGPT, Perplexity, Gemini1$399$299
Enterprise / CustomCustomAll 10 enginesCustom multi-regionCustomCustom

If you plan to buy prompt tracking 200 prompts deep or more, you will want Scale or Enterprise. Use this as a capacity reference for your own sizing, and look at the published plan page for anything you need to double check. Enterprise adds one-to-one expert onboarding and a dedicated account manager, and it carries custom limits on every metric. The material available for this guide does not describe per-engine prompt quotas, so if that detail matters to you, ask about it.

You are done with this step when you have a must-have engine list, an optional engine list, and a clear yes or no on whether one geographic region is enough.

Common mistake: treating more engines as more value before you have checked which engines your buyers use and which ones the plan includes.

Step 6: Count your unique tracking items and keep some headroom

Here is the formula to use for AI visibility plan sizing:

Planned prompt count = distinct prompt questions × required market and language versions, after removing redundant variants.

Count the questions, or the separately identified variants, and not the answers each engine gives back. If you track 100 prompts on three engines, that is 100 tracked prompts and 300 answer observations in a matched round. The observations are results, not extra prompt slots. A one-round observation estimate is simply prompt count times the number of engines you pick, and it says nothing about billing or how often a vendor runs your prompts, so keep those two numbers apart in your notes.

Then add a buffer. A reserve of 10 to 20 percent gives you room for new launches, new markets and replacements for prompts that stop being useful. That range is a planning suggestion in this guide and not a DeepSmith rule or an industry benchmark. If a plan allows 50, 100 or 200 prompts, you do not have to fill every slot before your first review.

Here is an example for a team that sits near the 50-prompt allowance, in one market and one language:

  • Five priority topics, three buyer stages and two distinct decisions per cell gives 30 Core prompts.
  • Four separately justified brand or competitor-comparison questions bring it to 34.
  • Six Explore questions for emerging needs bring it to 40 active prompts.
  • Ten slots stay unallocated as headroom, so 40 active prompts plus 10 spare slots fit inside 50.

That scenario only needs ChatGPT. If Perplexity is a must-have, the engine requirement moves you to Grow even though the prompt count would have fit on Pro by volume.

And here is a larger example for a team that is closer to 200, still in one market and one language:

  • Twelve topics, three stages and four distinct decisions per cell gives 144 Core prompts.
  • Twelve separately justified comparison, brand-evaluation or objection prompts bring it to 156.
  • Four Explore prompts bring it to 160 active prompts.
  • Forty slots stay free, so 160 active prompts plus 40 spare slots fit inside 200.

Both examples are made up to show the arithmetic. They are not measured benchmarks and not a recommended number of topics, and they assume every prompt really is distinct. If you were weighing whether to buy prompt tracking 200 prompts deep, this is the kind of count that tells you whether you would use it or only pay for it. If the same set has to run in several markets or languages, redo the count with each version listed separately and confirm the regional setup, because a one-region plan should not be assumed to cover several distinct geographic collections.

People also ask how much does it cost to track 50 prompts. For DeepSmith, the 50-prompt allowance sits on Pro, at $99 a month or $80 a month billed annually, and Pro includes ChatGPT. If you need Perplexity as well, Grow is the first plan with it, and that is true even if you only need 30 prompts. If the cost side is your main worry, check the annual and monthly rates on the published plan page before you settle on a tier.

You are done with this step when your inventory shows active prompts, reserve capacity, market and language versions, and engine count, and each total has been checked on its own line.

Common mistake: multiplying prompt capacity by engine count and calling the result a number of prompts. Keep prompt slots and answer observations on separate lines.

Step 7: Check the plan fit, validate, and revisit on purpose

Now compare your inventory with a plan, and check all of these and not only the headline prompt cap:

  1. Do your distinct prompts plus the headroom you want fit inside the allowance?
  2. Does the plan include every must-have engine?
  3. Is the region setup you need supported?
  4. Does the set cover your priority topics, buyer stages, audiences and the competitor decisions that matter?
  5. Can the team keep the Core set steady enough to compare results over time?
  6. Have you asked the vendor about every quota or regional detail that is not written down?

If you use DeepSmith, the AI Visibility module lets you write your own buyer prompts or use Discover Prompts, which suggests a starter set from your product, persona and buyer-stage context. It runs your tracked prompts on a schedule, saves the full answers, and reports mention and citation data for each prompt, along with which of your pages AI cites. A mention means an answer names your brand, and a citation means it links to one of your pages as a source. Those are different outcomes, so do not add them together as one success number.

Use the 7-day free trial, or onboarding, to test your scope before you pay. Check that your brand and competitor names are right, see whether the suggested prompts are relevant, confirm the engines and regions you want are available, read the captured answers, and see whether your list fits the allowance you have in mind. Do not assume a trial includes every plan's prompt capacity, engines or region settings, and ask DeepSmith if that matters to your decision. You can run this check with your own list.

When you report results, keep the Core set the same between reporting periods. Put new questions in Explore first, and change Core only when your product, market, buyer language or decision really changes. AI answers can vary from one run to the next, so check any surprising or high-stakes result across repeated observations where your tool supports it, and do not assume a plan includes repeat runs.

You are done with this step when you have a written prompt inventory, a plan that meets your prompt, engine and region needs, and a checklist of things to confirm during the trial or with the vendor.

Pro tip: when you share a score, say what it covers: the prompt families, audiences, buyer stages, markets, languages, engines and the period. A number without its scope is easy to misread.

What to do next

Take your spreadsheet and put the totals at the top: active prompts, headroom, market and language versions, and the engines you must have. Then set that beside two or three plans and see which one holds all of it, not only the prompt count. If nothing fits, trim the Explore list first, and look at the market count before you cut Core prompts. Once you have a plan, treat your first month as a test of your list, and expect to swap a few prompts as you see what the answers look like.

If you want to try your inventory against real data, start a DeepSmith free trial and load your Core list.

Frequently asked questions

How many prompts should I start with?

Start with the distinct buyer decisions you need to watch, not a target number. Some guidance suggests about 25 as a starting set, other guidance suggests 15 for an initial test or 30 to 80 for a working baseline. These are suggestions and not standards. A smaller set is fine if it is all your team can keep up with and read properly.

How much does it cost to track 50 prompts?

On DeepSmith, the 50-prompt allowance is on Pro, which is $99 a month or $80 a month billed annually, and Pro includes ChatGPT. If you need Perplexity, Grow is the first plan that has it, even if you track fewer than 50 prompts.

Do prompts count separately for every engine?

No. If you track 100 prompts on three engines, that is 100 tracked prompts and 300 answer observations in one matched round. Do not multiply your prompt list by the engine count when you estimate slots. If you need to know about per-engine limits or repeat runs, ask the vendor to confirm them.

Should I track every wording variation, competitor or market?

No. Group the wording variations that test the same buyer decision. Keep a separate question or a separate market and language version when the audience, the constraints or the local market could change the answer. Track competitors where they matter to a real comparison or objection, and skip a competitor-by-topic grid.