How to Build an AI Search Prompt Tracking List

SparkToro's research shows that people word the same need very differently, and that AI answers change from one run to the next. A prompt tracking list has to allow for both. This playbook builds one from real buyer questions to a run protocol whose numbers a team can defend.

This applies to B2B teams setting up or rebuilding AI visibility tracking in any tool, or tracking by hand. A prompt list samples a space of phrasings that never repeats exactly. Its numbers show whether a brand sits in the AI's consideration set, not a fixed ranking, and they need several runs before they mean anything.

What Is An AI Search Prompt Tracking List?

An AI search prompt tracking list is the fixed set of buyer questions a team runs through ChatGPT, Google AI Mode, Perplexity and similar engines to see whether its brand is mentioned or cited, and it matters because 51% of B2B software buyers in a G2 survey of 1,076 start research with an AI chatbot more often than Google.

A prompt list differs from a keyword list in three ways. Prompts are conversational and longer. They carry the buyer's constraints, such as team size, current stack, budget or region. And the answer to the same prompt changes from one run to the next, so a single check proves very little.

The same G2 research found that 71% of buyers rely on AI chatbots at some point in their research, and that buyers ask AI to synthesise options and run head-to-head vendor comparisons. A tracking list has to cover both kinds of question, so it needs unbranded category prompts and comparison prompts.

The list is the measurement layer of AI search visibility work, not the whole discipline. For the wider picture of what AI search visibility means and how to measure it, start with the definition guide. The rest of this playbook stays on the list itself.

How To Build An AI Search Prompt Tracking List Step By Step

To build an AI search prompt tracking list, start from the topics your team would actually create content for, collect the words buyers use in sales calls, support tickets and search data, rewrite them as conversational questions with real constraints, tag each by intent, buyer stage and persona, and run every cluster several times on two or three engines before reading the results.

The eight steps below run in order. Each names the decision and the section that goes deeper.

• **Start from business questions and topics.** List the topics your team would create content for, kept at one level of detail. Conductor says most organisations track 10 to 30 topics, and Aleyda Solis builds a matrix of product, audience, market and journey stage before writing a single prompt.
• **Collect real buyer language.** Pull wording from sales calls, support logs, Google Search Console and Bing queries, People Also Ask and community threads. The prompt research section ranks these sources.
• **Turn seeds into conversational prompts.** Add the constraints a real buyer states, such as role, team size, current stack, budget and region.
• **Write several phrasings per intent.** Group the variants into a cluster, so results are read across the cluster and never from one prompt.
• **Tag every prompt.** Give each one a topic, intent, buyer stage, persona, market, prompt type (branded, unbranded or competitor) and tier (core, experimental or monitoring).
• **Set the mix and the size.** Weight the list toward unbranded prompts and size it to cover every core topic within the tool's cap. The mix and list-size sections give the numbers.
• **Choose engines, market and a run protocol.** Pick two or three engines, set the market, and run each prompt several times inside a fixed window.
• **Validate, check usage, then launch.** Read every prompt by hand and check the tool's credit or run usage before activating, since Conductor notes that engines carry different cost rates. Then leave the list alone for a few cycles.

Aleyda Solis puts the principle behind these steps plainly. A representative library does not have to be exhaustive. It has to cover the business questions, with enough variants per question that one odd answer cannot swing the result.

The template below holds every tag column. Copy it into a spreadsheet, one row per prompt, and keep the source column honest. It shows later which wording came from buyers and which came from the team.

Prompt | Topic | Intent | Buyer stage | Persona | Market | Prompt type | Tier | Source of the wording | Date added
Which expense management software works for a 200-person company with UK and US entities? | Expense management | Commercial | Evaluation | Finance director | UK | Unbranded | Core | Sales discovery call | YYYY-MM-DD
How do I stop duplicate supplier invoices without adding approval steps? | Accounts payable controls | Informational | Awareness | AP manager | US | Unbranded | Experimental | Support ticket | YYYY-MM-DD

The method stays the same whichever tool runs the prompts. Where each research and tracking tool fits comes later, in its own section.

How To Do Prompt Research For AI Search

Prompt research for AI search starts with first-party language from sales calls, support logs and search queries, because SparkToro found that prompts written by different people for the same need had an average semantic similarity of just 0.081, so one marketer's phrasing rarely matches what buyers type.

Rank the sources by how close they sit to real buyer language:

1. Sales and support. Discovery calls, support logs and a "how did you hear about us" form field, which Profound suggests adding.
2. Search data. Google Search Console queries and the grounding queries in the Bing Webmaster Tools AI Performance report, which Microsoft groups by intent and topic.
3. Question-research tools. AlsoAsked shows live Google People Also Ask trees, and AnswerThePublic shows autocomplete-style questions, prepositions and comparisons.
4. Suggestions inside trackers. Semrush Prompt Research shows an estimated AI Topic Volume and the intent behind prompts. Peec AI and Promptwatch both suggest prompts, and Promptwatch builds its suggestions from Search Console queries, site content and competitor gaps.
5. The team's own guesses. These are the weakest source and belong at the bottom.

The SparkToro finding is why the last source ranks last. SparkToro's author argues that tracking a small set of "ideal" prompts tells a team very little, and recommends starting with research into how the audience describes the problem and its constraints. In practice that means several phrasings per intent, each drawn from a different source where possible.

SE Ranking adds sources worth a pass once the first-party material runs out. These include Reddit and Quora threads, paid search data, your own website and the answers an LLM gives when asked what buyers want to know. It then filters ideas on competitive relevance, influenceability, business intent alignment and scope.

Convert a keyword in three moves: Add the role, add the constraint and add the decision the buyer is trying to make.

Treat any volume figure as a guide to relative interest, not search volume. Semrush labels its AI Topic Volume an estimate, and Peec AI's Volume score runs from 1 to 5 relative to the industry. Both help order a list. Neither tells you how many people typed a prompt.

How To Segment Prompts By Intent And Buyer Stage

Segment a prompt tracking list by buyer stage first and intent second, so awareness questions are informational, evaluation questions are commercial comparisons and alternatives, purchase questions are transactional checks on price and setup, and post-purchase questions are how-to and support prompts, then add persona and market as tags.

The stage-to-intent mapping follows Profound's guidance, which pairs awareness with informational prompts, evaluation with commercial prompts and some branded options, conversion with transactional prompts and post-purchase with how-to and support. SE Ranking sizes the stages for a starter list at 10 to 20 awareness prompts, 20 to 30 consideration prompts and 5 to 10 brand-evaluation prompts.

Persona and market are the second and third layers. Conductor segments by persona and intent. Aleyda Solis goes further, adding product line, market and buyer constraint, so a prompt for a UK finance director and one for a US controller sit in the same topic but report separately.

Each prompt belongs to exactly one topic and carries several tags. That is how Peec AI structures prompts, and it is the right model in a spreadsheet too. Results can then be filtered by stage, persona or market without counting a prompt twice.

For B2B software, the guide to AI visibility tools for B2B SaaS teams shows how trackers handle buyer-intent prompt sets once the stages are set.

Vendor defaults are starting points, not evidence. Peec AI's prompt discovery defaults to 20% branded prompts and an intent split of 25% informational, 50% commercial and 25% transactional, and both splits can be adjusted. That split says nothing about what your buyers ask, so set it from your own sales and search data.

What Mix Of Branded And Unbranded Prompts Should You Track?

Track mostly unbranded prompts, because Conductor recommends roughly 75% unbranded to 25% branded prompts and warns that branded prompts tend to overstate how visible a brand really is in AI answers. A branded prompt already points the engine at the brand, so it measures recall rather than discovery.

The other published defaults agree. Profound tells teams to prioritise unbranded category prompts, and Peec AI's discovery default sets branded prompts at 20% of the list.

In a 100-prompt list, Conductor's ratio means about 75 unbranded prompts and up to 25 branded prompts. Profound's stage mapping shows where branded prompts can fit. It includes some branded options in the evaluation stage and places how-to and support questions in post-purchase.

Competitor prompts are a separate, deliberate share. Questions such as "alternatives to" a named rival, or one product "vs" another, belong to the evaluation stage, and Profound keeps them a small share of the list. Tag them as competitor prompts. SE Ranking tracks brand-evaluation prompts separately for the same reason, so they do not inflate category metrics.

How Many Prompts Should You Track?

Most B2B software teams should start with 30 to 100 prompts and grow from there, since Aleyda Solis suggests 30 to 50 prompts as a minimum viable library and 100 to 250 for SaaS and services companies, while tracking vendors such as Conductor recommend far larger sets. The right number covers every core topic with several phrasings.

The spread follows who is speaking. Profound says its users typically track 100 to 1,000 prompts, and Conductor puts the upper limit at about 300 to 500 per topic. Those are the sets their products are built to run. Aleyda Solis, who sells no tracker, sizes the library by business complexity instead, with 30 to 60 prompts for a single product and market.

Decision rule: Start with the smallest list that covers every core topic with several phrasings per intent, sized to the tool's cap, then grow once the first cycles have run.

The cap is often set by the tool, not the team. Semrush's base AI Visibility Toolkit tracks 25 prompts, with 50 more available as a paid add-on. Peec AI counts accepted suggestions toward the plan's active prompt limit, so saying yes to every suggestion fills the list with prompts nobody chose.

Which AI Engines Should You Run Your Prompts On?

Run a prompt tracking list on the two or three engines your buyers actually use, and add Microsoft's free AI Performance report in Bing Webmaster Tools, which shows citations across Copilot, Bing and select partner AI experiences, as a first-party check on what any paid tracker reports.

SE Ranking suggests two or three priority models. Conductor chooses engines by purpose:

• ChatGPT for a baseline of what the model says.
• ChatGPT Search for which sources it pulls in.
• Google AI Overviews for summary-level visibility.
• Perplexity for citations.
• Gemini and newer engines only when they matter to the audience.

Tool coverage narrows the choice. Semrush Prompt Tracking runs on ChatGPT Search, Google AI Mode and Gemini. Promptwatch lists ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews among its engines, compared side by side.

Bing's report adds features in preview that group grounding queries into intents and topics, which makes it a free source of the language people use with Copilot. Keep each engine's results separate rather than blended into one score, as Aleyda Solis advises, because each engine answers the same prompt differently.

How Often Should You Run And Refresh Your Prompt List?

Run tracked prompts at least weekly and several times inside a short window, read trends only after about 30 days, and refresh the list quarterly, keeping core prompts fixed so trend lines stay comparable while experimental prompts rotate. Running and refreshing are two different clocks, and mixing them up breaks the trend line.

Run cadence: Peec AI and Semrush both run tracked prompts daily. Peec AI runs active prompts on a 24-hour cycle, and Semrush Prompt Tracking gives daily updates. Conductor describes weekly tracking as the industry standard. For manual or spot checks, Aleyda Solis runs each prompt 3 to 5 times inside the same 24 to 72 hour window.

Reading window: SE Ranking says to track for at least 30 days before drawing conclusions. Conductor expects early signals within a few cycles and meaningful trends after several weeks.

Refresh cadence: SE Ranking reviews the list every 30 to 60 days. Conductor updates prompts or topics quarterly or twice a year. Aleyda Solis sets the refresh by tier, keeping core prompts stable, rotating experimental prompts and watching monitoring prompts for change.

Which Tools Help Build A Prompt Tracking List?

Question-research tools such as AlsoAsked and AnswerThePublic supply the seed wording for a prompt tracking list, while trackers such as Semrush's AI Visibility Toolkit, Peec AI and Promptwatch run those prompts on AI engines, so most B2B teams need one of each rather than one tool for everything.

Each tool below is mapped to the step it serves. None is ranked.

AlsoAsked

Role in the list: AlsoAsked supplies seed questions from live Google People Also Ask results, grouped by Google's own intent clustering.

What it adds: Searches can be set by country, city and language, and a Deep Search returns about 150 questions per query on average. Results export as PNG or CSV, and cached searches can be recalled for free.

Setup and permissions: Bulk search takes a CSV of up to 1,000 queries at once. Processing takes minutes to hours and returns a zip of CSVs with the related questions.

Cost driver: Search credits. A bulk run needs at least as many credits as terms, the full amount is reserved up front, and credits for queries with no results are refunded.

Limitation: These are Google search questions, shorter and less specific than a buyer's AI prompt, so every seed needs rewriting with a role, a constraint and a decision before it goes on the list.

AnswerThePublic

Role in the list: AnswerThePublic shows the words an audience uses across Google, YouTube, Amazon, ChatGPT and more, which widens the pool of seed phrasings.

What it adds: Its Ideas reports group suggestions into questions, prepositions, comparisons, alphabetical expansions and related terms.

Setup and permissions: Plans set the number of users, from a single user on the entry plan to up to 10 on the largest.

Cost driver: Each plan caps monthly searches, prompt suggestions and prompt responses, from 100 searches and 1,000 prompt suggestions a month on the entry plan to 300 searches and 3,000 prompt suggestions on the largest.

Limitation: An independent review on diyai.io notes that raw results need de-duplication and clustering because the same intent appears several times, that volume and broad intent labels are not enough for final prioritisation, and that the AI Models report's figures should not be read as audited market-share evidence.

Semrush

Role in the list: Semrush covers both research and tracking in its AI Visibility Toolkit. Prompt Research, which Semrush calls "keyword research for the AI era", takes a broad topic or product category and shows an estimated AI Topic Volume for related topics and the intent behind prompts.

What it adds: Prompt Tracking sits inside the Position Tracking tool. Choose ChatGPT Search, Google AI Mode or Gemini as the engine, add target prompts, and receive daily updates on Visibility and Average Position. Adobe completed its acquisition of Semrush in April 2026, and the Semrush name continues.

Setup and permissions: In corporate accounts each user needs their own toolkit licence, and the toolkit cannot be shared through shared limits. Brand Performance campaigns are shared with eligible teammates, and other data is shared by export or through My Reports.

Limitation: The base toolkit tracks 25 prompts in one project, with 10 CSV exports a day of up to 1,000 rows. Treat Average Position as directional only, for the reasons in the measurement section.

Peec AI

Role in the list: Peec AI organises prompts by topic, with each prompt in exactly one topic, an informational, commercial or transactional intent label and optional tags. Peec AI suggests starting with 3 to 5 broad topic areas, and its documentation says capturing the intent matters more than the exact wording.

What it adds: Prompts can be added by hand, by bulk CSV, through suggestions or through prompt discovery with adjustable branded and intent splits. A Volume score rates each prompt's estimated volume from 1 (very low) to 5 (very high).

Setup and permissions: Bulk add runs from the Prompts page. Tag and topic values in the CSV columns are created and assigned automatically, and an import preview appears before the upload is confirmed.

Cost driver: Active prompts. Each plan includes a set number, and accepted suggestions count toward it, while archived and deleted prompts do not.

Limitation: Prompts in a CSV are capped at 200 characters, and deleting a prompt erases its historical data. Teams weighing it against another tracker can read Profound vs Peec AI for engine coverage and cost.

Promptwatch

Role in the list: Promptwatch builds prompt suggestions from Google Search Console queries, an AI generator based on site content, and competitor gap analysis, and it can turn an imported keyword list into prompts.

What it adds: Personas let prompts carry a buyer's role, goals and context. Prompts can be tagged by Prompt Type (Organic, Brand Specific, Competitor Comparison) and by intent. Promptwatch states there is no fixed prompt limit, says some customers track over 6,000 prompts across 17 markets, and recommends 15 to 30 prompts per monitor or market to start.

Setup and permissions: Onboarding asks for the brand URL and 5 category prompts, and a project is the workspace for one brand. Agent Analytics, which shows which AI crawlers visit which pages, needs a CDN connection such as Cloudflare, Vercel or Fastly.

Limitation: Promptwatch is built as a full optimisation suite, with content gap analysis and crawler analytics, which is more than a team needs if all it wants is a place to collect seed prompts.

Decision rule: Use a question tool (AlsoAsked or AnswerThePublic) for seeds and a tracker (Semrush, Peec AI or Promptwatch) for runs. A team already on Semrush can start there. A team that wants tagging, persona and suggestion controls should look at Peec AI or Promptwatch.

For a side-by-side view of trackers on coverage and cost, see the AI visibility platforms comparison.

How To Measure Results From A Prompt Tracking List

Measure a prompt tracking list by visibility share across clusters of prompts rather than by rank, because SparkToro's study of 2,961 AI responses found less than a 1 in 100 chance that ChatGPT or Google's AI returns the same brand list twice.

SparkToro's author concludes that visibility percentage, measured across dozens to hundreds of prompts run several times, is a reasonable measure of presence in the AI's consideration set. In a follow-up of 142 human-written headphone prompts, leading brands appeared in 55% to 77% of 994 responses, which is the kind of spread SparkToro's conclusion rests on.

Track these measures:

• Visibility share. The percentage of runs in a cluster where the brand appears, reported per topic, buyer stage and engine.
• Branded and unbranded visibility. Two numbers, always reported apart.
• Citation share. How often your pages are cited. Bing's preview Citation Share is the percentage of citations attributed to your site out of all citations shown for a grounding query, and Microsoft states it does not expose competitor domains or represent traffic share.
• Competitor gaps. Clusters where competitors appear and you do not.

Report each share in the same cuts as the tags. A grid of topic by buyer stage by engine shows where the brand is present and where it drops out, and persona and market tags split it further once the list is large enough. Read Bing's intent and topic labels with care, since Microsoft says they come from evolving classifiers and may be broad while in preview.

Rank position is not on the list. SparkToro calls ranking position in AI answers unreliable, so read any tool's Average Position as directional at best.

The last step turns a measure into work. Map each cluster with low visibility to a page to create or improve, then watch that cluster's share over the next few cycles.

Common Mistakes When Building A Prompt Tracking List

A common mistake in a prompt tracking list is tracking one ideal phrasing per question, since SparkToro found that people word the same need very differently and AI answers change between runs, so a single prompt can show a brand as present or absent by chance.

• Keywords renamed as prompts. They lack the role and constraints buyers state. Rewrite each with a role, a constraint and a decision.
• One phrasing per intent. One prompt is one sample. Write several variants and read the cluster.
• A branded-heavy list. Branded prompts overstate visibility. Keep them at a quarter or less and report them apart.
• Blended engines. One score hides where the brand is missing. Report each engine separately.
• Changing prompts too early, or deleting them. Edits before a few cycles break the trend line. Wait, and archive rather than delete.
• Reading rank as a metric. AI rank positions do not hold between runs. Use visibility share.
• Accepting defaults and suggestions unreviewed. Vendor splits are not buyer evidence, and accepted suggestions in Peec AI use up plan limits. Review each one.

FAQ

Sources