PROMPT TRACKING FRAMEWORKS SKILL A free skill by Orit Mutznik: https://www.oritmutznik.com/ai-search-optimisation/prompt-tracking-frameworks HOW TO USE: Attach this file to a new chat with ChatGPT, Claude, Gemini, Copilot or Perplexity and send it with the words "Build my prompt set" (you can also paste the whole text in). Answer its questions. You get 25 to 50 prompts for tracking your brand in AI search, as a table you can paste into any spreadsheet. ===================================================================== INSTRUCTIONS FOR THE AI ASSISTANT ===================================================================== You are helping me build a prompt set to track my brand's visibility in AI search (ChatGPT, Gemini, Google AI Mode, Perplexity, Copilot and Claude). Follow the 6 steps below in order. Keep the tone plain and practical, use British spelling and talk to me as "you". Never write prompts before I have confirmed the framework and the level or stage descriptions. STEP 1: INTERVIEW ME Ask in 3 short rounds and wait for my answers each time. Skip anything I have already told you. Round 1, the brand: 1. Brand name and website. 2. What you sell and the one product or product line to start with (every framework is built per product, so a brand with several products repeats the build for each). 3. Do you make the products, sell other brands, or both? 4. Market (country) and language. 5. Who buys it: Up to 3 types of buyer, a sentence each. 6. The 3 competitors that come up most often. Round 2, the goal: 1. Which of these do you most need to know? a) Which buyers does the AI recommend you to, from first-timer to expert? b) At which stages of the buying journey does your brand appear in AI answers? c) Exactly where does your brand appear, by buyer level and stage, for your most important product? 2. Can your brand or marketing team help? They know how each type of buyer talks. Pick one: They describe the buyers with me (best), they review a draft I write, or no help this time. 3. Which AI platforms will you track? (Suggest ChatGPT, Gemini, Google AI Mode and Perplexity if I have no view.) Round 3, real queries: Real questions people already type beat anything invented and they go into the set first, so ask me to paste what I have from these sources: 1. Google Search Console: In the Performance report, filter Query by the custom regex on the next line, which shows queries of 10 words or more, the ones that read like prompts. Show the regex exactly as written: ^(?:\S+\s+){9,}\S+$ 2. Bing Webmaster Tools: Export the queries from the Search Performance report and sort them by word count in a spreadsheet, since it has no regex filter. If I can see its AI Performance report, ask for the grounding queries too: The phrases Copilot used to find the pages it cited. 3. The prompt library of an AI visibility tool such as Ahrefs Brand Radar, site search terms, support tickets, sales call notes or reviews. If I have none, say the set will be drafted from scratch and carry on. STEP 2: RECOMMEND A FRAMEWORK Recommend one of these 3 frameworks from my round 2 answer, say why in up to 3 bullets tied to what I told you, say how many prompts it produces and ask me to confirm or switch. Framework 1, prompt tracking by buyer expertise. Groups prompts by how well the buyer knows the category, in 5 levels. Logic: Buyers ask about the same product in different words depending on what they know. The AI answers each vocabulary differently, so a set written in one vocabulary only shows one type of buyer. Goal: Capture which buyer levels the AI recommends the brand to, so the content can speak each level's language. Recommend it for answer (a), or when the product serves first-timers and specialists alike (investing, software, luxury, education). Size: 5 prompts per level, 25 in total. The 5 levels: - Newcomer: First time buying this. Plain words, no category terms, often no product named. - Learner: Picking up the terms. Names the category and asks what things mean. - Informed: Knows the features. Asks for features by name. - Seasoned: Has bought before. Asks about the catches: Fees, returns, what goes wrong. - Expert: Knows the small print. Asks about details most people never reach. Newcomers and learners rarely name a brand, so brand prompts at those levels are trust checks, value checks or simple fact checks ("is [brand] a safe place to buy my first...", "can I send them back to [brand] if they don't fit"). Framework 2, prompt tracking by funnel stage. Groups prompts by where the buyer is in the journey, in 5 stages. Logic: The stage decides what the person asks for: Explainers early on, comparisons in the middle, fact checks at purchase, stay-or-leave questions at retention. Goal: Capture which stages of the journey the brand appears in, from the first explainer to the decision to stay. Recommend it for answer (b), or when the user already reports by funnel stage, sells to one main audience or has little time from the brand team. Size: 5 prompts per stage, 25 in total. The 5 stages and what each usually chases: - Awareness: Spotting the need, no product in mind. Chases a citation. - Interest: Learning how to solve it. Chases a citation. - Consideration: Comparing options. Chases a mention. - Purchase: Ready to buy, checking the facts. A brand prompt. - Retention: Using it or thinking of leaving. A brand prompt where a competitor can get named. Non-brand questions about using or replacing the product (how long it lasts, whether to repair it) go under interest. Retention prompts must be the ones that could lose a customer ("is it still worth it", "a cheaper alternative to", "better offers for existing customers than"), never support questions like "how do I cancel". Framework 3, the buyer journey matrix. Both at once for one product: 5 levels down the side, 5 stages across the top, 25 cells. Logic: The stage shows which intents the brand appears for and the level shows which words, so crossing them pinpoints the cell. Goal: Pinpoint the exact cells where the brand appears and where it doesn't (missing for seasoned buyers at consideration, say). Recommend it for answer (c), when one product matters most and the brand team can review the rows. Size: One or two prompts per cell, 25 to 50 in total. One product per matrix, with framework 2 for the rest of the range. If I am unsure, ask which result would change what I do next: A list of buyer levels to write for (1), a list of journey stages to fix (2) or one precise cell for the top product (3). STEP 3: DESCRIBE THE LEVELS OR STAGES Write the 5 levels (framework 1), the 5 stages (framework 2) or both (framework 3) in my category's own words, as a table: What each buyer knows or is doing and how they would phrase a question. Place my buyer types on the levels (one type often spans two) or spread them across the stages. Build it from what I told you, from the real queries I pasted and from my website if you can browse it. Then follow my round 2 answer: If the team describes, ask me to paste what they say and write the table from that. If the team reviews, ask me to have them check the table before you write prompts. If there is no help, say the table is less precise until someone who talks to buyers checks it. Ask for corrections before writing any prompts, because the prompts are only as good as these descriptions. STEP 4: WRITE THE PROMPTS Real queries first: Map each real query from round 3 to the level, stage or cell it fits, rewrite it into the structure below (keep the person's own words wherever the rules below allow) and only then fill the gaps with prompts you draft yourself. Mark every prompt as real or drafted: A prompt built from a real query counts as real even when you split, merge or reword it, while a prompt built only from a grounding query's context counts as drafted. If a level, stage or cell has more real queries than slots, keep the strongest and list the rest under the table as spares. Use grounding queries from Bing's AI Performance report only to choose contexts: They are the phrases Copilot searched with, so they never go into the set as prompts. Every prompt follows these rules: - One intent (what the person wants to do), one entity (the thing they ask about) and one context (one specific condition the answer depends on, which Mark Traphagen calls the aspect). Two contexts in one prompt measure two things at once, so split them. - Natural wording: The way a person types into an LLM, with their situation around the question, about 15 to 20 words. The context is the one condition the answer depends on, while the situation is what the person adds around it and doesn't change what you measure. - Brand or non-brand tag: A brand prompt names my brand (use the real name) and shows how the AI presents it. A non-brand prompt shows whether I get recommended when nobody asked for me by name. - Mention or citation tag. Non-brand: "Which X should I buy", "best X for Y" and "where's the best place to buy X" chase a mention (the AI recommends the brand or shop); "how do I choose X", "what X do I need", "what does X mean" and "is it worth paying more for X" chase a citation (the AI uses my page as a source). Brand: A verdict question (is it worth it, can I trust it, is there anything better) chases a mention; a fact check (does it offer, how long is the warranty, can I return it) chases a citation. - If I sell other brands rather than make them: "Which X should I buy" prompts name products and leave the shop out, so make every non-brand mention prompt a where-to-buy prompt ("where's the best place to buy X"). In framework 1, include at least one in every level. In frameworks 2 and 3 they sit at consideration, so the other stages keep the shape the stage table gives them. - Leave competitor names out of the prompts. They get tracked in the answers. - Avoid leading prompts ("why is [brand] the best"), keyword fragments and two prompts with the same meaning. Examples of the structure (intent / entity / context): - Newcomer, fintech, non-brand, citation: "I've got £5,000 sitting in my current account doing nothing, what should I actually do with it?" (what should I do with it / £5,000 in my current account / doing nothing) - Informed, fintech, non-brand, mention: "Which investing app lets me buy fractional shares of US companies from a UK account?" (which lets me buy / investing app / fractional shares of US companies) - Consideration, outdoor retailer, non-brand, mention: "I've got wide feet and most boots pinch, where's the best place in the UK to buy wide-fit walking boots?" (where's the best place to buy / walking boots / wide fit) - Seasoned, fintech, brand, mention: "Is [Brand] still worth it for someone who only trades a few times a month once the fees add up?" (is it still worth it / [Brand] / for someone who trades a few times a month) - Expert, fintech, brand, citation: "Does [Brand] let me transfer my US-listed ETFs from another broker without selling them first?" (does it let me / [Brand] / transfer US-listed ETFs from another broker) - Retention, ecommerce, brand, mention: "I bought my last pair from [Brand], is it still the best place to buy running shoes for my next pair?" (is it still / [Brand] / the best place to buy running shoes) STEP 5: BALANCE THE SET Count the 4 combinations (non-brand mention, non-brand citation, brand mention, brand citation) and show them in a small table. The floor is 4 prompts per combination in every 25. Framework 1 aims for about a quarter in each, so treat fewer than 6 as short. Frameworks 2 and 3 take their shape from the stages, roughly 10 non-brand citations (awareness and interest), 5 non-brand mentions (consideration) and 10 brand prompts (purchase and retention) per 25, with the brand prompts split between mentions and citations, so only the floor applies. If a combination is short, swap the weakest drafted prompt in the fullest combination within the same level or stage, so every level or stage keeps its count and the real queries stay in, then say what you swapped. STEP 6: DELIVER 1. The prompt set as a table with these columns, ready to paste into a spreadsheet: ID, level or stage or cell, prompt, brand or non-brand, mention or citation, intent, entity, context, source (real or drafted). Put any spare real queries under the table. If you can create files, also offer the same table as a CSV file. 2. How to run it, in 6 short bullets: Run every prompt on every platform in scope more than once, logged out where the platform allows it or with memory off, because answers change from run to run. Run the set separately for each market I sell in, written in that market's language. Report the mention rate and citation rate (the share of runs where the brand is mentioned or cited) by level, stage or cell, keeping brand and non-brand prompts apart because brand prompts name the brand by design. Log which competitors each answer names, because that shows who gets recommended in my place. Run a weekly sample of two prompts from every level or stage (the two most likely to lead to a sale) and the full set monthly, noting every major model release. Keep the set fixed between runs, because changing prompts breaks the trend. 3. Next steps: Point me to the Fanout Explorer bookmarklets (https://www.oritmutznik.com/ai-workflows/chatgpt-vs-claude-vs-gemini-fan-outs-bookmarklets), which show the searches an assistant runs behind a prompt and the pages it reads, then offer to read any fan-outs I paste back. Offer to repeat the build for my next product and to rebuild the set when the product, market or competitors change. Credits: Built by Orit Mutznik (https://www.oritmutznik.com). The prompt structure of one intent, one entity and one context builds on Mark Traphagen's (he calls the third part the aspect) "Stop Chasing Every Prompt: A Simple Framework for AI Prompt Tracking". Begin now with round 1 of the interview.