Agency CRO Audit Prompt Kit (Free CRO Toolkit 04) Edition 2, October 2026 (c) 2026 Anshul (anshuldesigns.com). Free to use with your team and your clients. Please don't sell it or remove this notice. You may send the client summary pages and the client summary template to clients without this notice. This file holds the 12 prompts of the Agency CRO Audit Prompt Kit, as plain text, so nothing breaks when you copy and paste. Copy everything between COPY FROM HERE and TO HERE. Fill the [BRACKETS] with a code name for the client, never the client's real name. Paste your export or text where it says [PASTE ... HERE]. Check every number the AI tool gives you against your own export before you act on it. Ids in these files (L and P are checks, F and A are prompts): L Scorecard 01, Landing Page Scorecard: https://www.anshuldesigns.com/downloads/anshul-landing-page-scorecard.pdf P Scorecard 02, Shopify Product Page Scorecard: https://www.anshuldesigns.com/downloads/anshul-shopify-product-page-scorecard.pdf F Pack 03, Store Data Prompt Pack: https://www.anshuldesigns.com/downloads/anshul-store-data-prompt-pack.pdf A Kit 04, Agency CRO Audit Prompt Kit: https://www.anshuldesigns.com/downloads/anshul-agency-cro-audit-prompt-kit.pdf Prompts in this file: A01 Kickoff and data request A02 Data reconciliation A03 Segment funnel memo A04 Traffic source normaliser A05 Ad Library angle map A06 Message match batch A07 Scorecard to findings A08 Dollarise findings A09 Prioritise and 30/60/90 A10 Test plan A11 Red-team the report A12 Client summary Free CRO audit of your whole store: https://www.anshuldesigns.com/free-cro-audit?from=pdf-agency ============================================================ A01 Kickoff and data request Eight questions and a data request the client can action. About 10 minutes. Use when: Day 1, after the first call and before you ask for any data. Inputs to paste: - Notes from the first call. Your own notes: goals, channels, what they tried, worries. Delete first: Staff names and contact details. - Store web address. The client's store URL, domain only. Delete first: Anything after the domain. - Never paste: Passwords, keys, login links, margins. Pairs with: - F01 Funnel locator (Pack 03) - F03 Landing page league table (Pack 03) Where it goes wrong: A customer-level request, or an untagged Table 2 row. Delete it. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A01 KICKOFF AND DATA REQUEST ROLE You are a CRO strategist at an ecommerce agency. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Platform: [PLATFORM]. Ad channels: [AD CHANNELS]. Audit window: [START DATE] to [END DATE]. INPUTS My notes from the first call (goals, channels, what they have tried, worries): [PASTE CALL NOTES HERE] The store web address, domain only: [PASTE STORE URL HERE] STEPS 1. Write 8 discovery questions: business outcome first (the result, the metric, the page that matters), then constraints (theme access, release rhythm). If my notes answer one, ask a sharper one. Questions ask for goals, dates and decisions, never for figures or spend. 2. Build the data request from this menu only, one date range for all, my audit window (if it crosses 21 to 23 Sep 2026, ask for session metrics from 24 Sep 2026): funnel by device, new vs returning and top 5 sources (A03); sales and orders, all channels and online store (A02); GA4 purchases and revenue (A02); ad platform purchases and value (A02); sessions and orders by source, medium, campaign (A04); ads with destination URLs (A05, A06); sessions and orders by landing page, with AOV (A08). 3. Write Where to find it using only: Shopify Analytics > Reports (Conversion rate breakdown, Sessions by landing page, Sales attributed to marketing, Sessions attributed to marketing campaigns), Shopify Orders, GA4 Reports or Explore, the ad platform reports. For another platform, or if unsure, write: Client to confirm the report name. 4. Write Remove before sending: customer names, emails, addresses and order numbers; for ad reports also audience names, and spend unless the client approved pasting it. 5. List the view-only access and who grants it. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Request nothing outside the menu. Ask for customer-level data only if my notes show a need. 9. End each Why we need it with the tag, for example (A03). 10. Date rule: my range starts on or after 24 Sep 2026, or ends before 21 Sep 2026. Mention these dates only if my rows cross 21 to 23 Sep 2026: then say so and compare orders and sales, not sessions or conversion rates. OUTPUT Table 1 (8 rows) with columns: Question | Why we ask Table 2 with columns: Export | Where to find it | Date range | Why we need it | Remove before sending Table 3 with columns: Access needed | Who grants it TO HERE ============================================================ A02 Data reconciliation Which numbers to quote, with every difference explained. About 10 minutes. Use when: Day 2, when the exports arrive and before any number goes in a report. Inputs to paste: - Shopify totals, attributed orders. Analytics > Reports: orders, net sales, Sales attributed to marketing. - GA4 purchases and revenue. Reports > Monetization > Ecommerce purchases (client to confirm). - Ad platform purchases and value (optional). Ads Manager, with each ad set's attribution setting. Delete first: Spend, audience names. - Never paste: Order-level exports. Pairs with: - F01 Funnel locator (Pack 03) - F02 Device gap (Pack 03) Where it goes wrong: A platform compared with all store orders, or GA4 revenue called net or gross. Check the Compared with column. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A02 DATA RECONCILIATION ROLE You are a data analyst at an ecommerce agency. You only use the numbers I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. Time zone: [TIME ZONE]. INPUTS Totals for the date range and time zone above: Shopify online store orders and net sales (all channels too, if shown), and the orders Shopify attributes to each ad platform: [PASTE SHOPIFY TOTALS HERE] GA4 purchases and purchase revenue: [PASTE GA4 TOTALS HERE] Ad platform purchases and purchase value, with the attribution setting of each ad set: [PASTE AD PLATFORM TOTALS HERE] STEPS 1. List each figure as Metric, Source and Value, in the units I pasted. 2. Reference: for GA4, Shopify online store orders and net sales; for an ad platform, the Shopify orders attributed to it, never all orders. If the reference is missing, write "Not enough data". 3. Difference = (row value - reference) / reference, as a percentage. Flag more than 20% (a rule of thumb). 4. Likely reason, only from: the platform counts only orders it attributes to its own ads, attribution setting (7-day click, 1-day view), consent banners, pixel settings or blocked tracking, duplicate events, refund timing, time zone, currency, all channels vs online store. Label it ASSUMPTION. 5. Quote this? YES for Shopify, YES WITH NOTE where the difference is explained and defined, NO where not. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Shopify is the reference. Never average sources or pick the figure that looks best. Compare orders with orders, revenue with revenue. 9. Never call GA4 revenue net or gross: Google's spec leaves shipping and tax out of purchase value; check the client's tag. OUTPUT Table 1 with columns: Metric | Source | Value | Compared with | Difference | Likely reason | Quote this? Then one line for each flagged row: Tracking flag: , , . Then, only if all ad platforms' purchases added together exceed Shopify online store orders: Over-attribution: / = , with nothing after it. TO HERE ============================================================ A03 Segment funnel memo The segment that hides the defect, and a 5-bullet memo. About 10 minutes. Use when: Day 2, after A02, to see where the funnel leaks and for whom. Inputs to paste: - Conversion rate breakdown. Analytics > Reports. Note open or closed funnel. - The same, by device type. New exploration, dimension Device category (client to confirm). - New vs returning, top 5 sources. If your admin shows them; else GA4 funnel exploration. - Never paste: Order or customer rows. Pairs with: - F01 Funnel locator (Pack 03) - P29 Add to cart confirms, shows checkout (Scorecard 02) Where it goes wrong: A segment with few completed checkouts picked as weakest, or two tools mixed. Check the LOW tags. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A03 SEGMENT FUNNEL MEMO ROLE You are an ecommerce analyst at an agency. You only use the numbers I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS The conversion funnel for the date range above, one table per split: device category, new vs returning visitors, top 5 traffic sources. Per segment: All sessions, Sessions with cart additions, Sessions that reached checkout, Sessions that completed checkout. Name the tool of each table and, for Shopify, open or closed funnel. [PASTE THE TABLES HERE] STEPS 1. Per segment: add-to-cart rate = sessions with cart additions / all sessions; checkout rate = reached checkout / sessions with cart additions; completion rate = completed checkout / reached checkout; conversion rate = completed checkout / all sessions. 2. Write LOW after the segment name if it has fewer than 30 completed checkouts (a rule of thumb). 3. Within each split, a segment's weakest stage is where its rate divided by the best rate in that split is lowest. None if its ratio is 1 at every stage. Label it EARLY (add to cart) or LATE (checkout, completion). 4. Choose the segment and stage that hide the defect: the lowest ratio among segments not marked LOW. 5. Write a client memo of 5 plain-English bullets, each with one number and what it means. Never draw a conclusion from a LOW segment. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Shopify counts sessions; GA4 funnel exploration counts users. If a later stage is larger than an earlier one, say so and do not force it to fit. 9. Compare segments only within one split and one tool. Never mix Shopify and GA4 rows in one comparison. 10. Date rule: my range starts on or after 24 Sep 2026, or ends before 21 Sep 2026. Mention these dates only if my rows cross 21 to 23 Sep 2026: then say so and compare orders and sales, not sessions or conversion rates. OUTPUT Table 1 with columns: Segment | All sessions | Add-to-cart rate | Checkout rate | Completion rate | Conversion rate | Weakest stage, written as stage (EARLY or LATE, ratio a / b = c) Then one line: Open first: . Then the client memo: 5 bullets. TO HERE ============================================================ A04 Traffic source normaliser Clean campaign rows, and the funded traffic that does not convert. About 10 minutes. Use when: Day 2, when source and campaign rows are messy or paid traffic is not converting. Inputs to paste: - Sessions and orders by source, medium and campaign. Analytics > Reports > Sales attributed to marketing (it includes UTM campaign name) with Sessions attributed to marketing campaigns; or GA4 Traffic acquisition, Session source / medium, Session campaign, Ecommerce purchases added. Top 100 rows, one tool. Delete first: Any column with customer names or order numbers. - Never paste: Unreleased product names or prices. Pairs with: - L30 Destination URL is clean (Scorecard 01) Where it goes wrong: A direct or organic row merged into a paid campaign. Open two merged rows and check each source row. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A04 TRAFFIC SOURCE NORMALISER ROLE You are a paid media analyst at an ecommerce agency. You only use the rows I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. Paid media values: [PAID MEDIUM VALUES]. Minimum sessions for a leak: [MINIMUM SESSIONS]. INPUTS Sessions and orders by source, medium and campaign for the date range above, from one tool (the top 100 rows by sessions). Columns: source, medium, campaign, sessions, orders. [PASTE THE ROWS HERE] STEPS 1. Make a merge key for each row: the campaign name in lowercase, with spaces, hyphens and underscores treated as the same character. Merge rows with an identical key only when their media are both in my paid list. Never merge a paid row with an unpaid row. 2. A campaign name that is blank, "(not set)", "(direct)", "(organic)", "(referral)", "(none)", only digits, or an unfilled template tag in curly braces is NAME MISSING. Do not merge it. 3. For each merged campaign add the sessions and the orders, count the rows merged, and calculate conversion rate = orders / sessions. 4. Confidence is LOW if the merged campaign has fewer than 30 orders (a rule of thumb), otherwise OK. 5. Funded but leaking: paid campaigns with at least the minimum sessions and a conversion rate below the combined rate of all rows whose medium is in my paid list. Show at most 5, most sessions first. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Merge only identical keys. Never merge similar names. 9. Direct means unknown. Never call direct traffic loyal or returning customers. OUTPUT Table 1 with columns: Campaign (merged) | Rows merged | Sessions | Orders | Conversion rate | Confidence Table 2, Funded but leaking, with columns: Campaign | Sessions | Orders | Conversion rate Then one line: Paid conversion rate: a / b = c. TO HERE ============================================================ A05 Ad Library angle map Which angles the client keeps paying for, and where the ads land. About 15 minutes to run and 30 to collect. Use when: Day 3, to see what the client has been running and where those ads send people. About 15 minutes to run and 30 to collect the ads. Inputs to paste: - Active ads: text, start date, format, URL. Meta Ad Library (facebook.com/ads/library), by advertiser: copy the ads into a sheet; if your search shows a download option, use it. Exact URLs come from Ads Manager. Delete first: Spend, results and audience names. - Today's date. Type it in the [TODAY DATE] bracket. - Never paste: Spend or results from Ads Manager. Pairs with: - L06 Oldest ad's angle leads; one ad: tick (Scorecard 01) - L30 Destination URL is clean (Scorecard 01) Where it goes wrong: Date arithmetic, and an angle judged from the URL. Recompute one run length; UTM flags from Ad Library say ASSUMPTION. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A05 AD LIBRARY ANGLE MAP ROLE You are a paid social strategist at an ecommerce agency. You only use the ads I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. Today: [TODAY DATE]. INPUTS 10 to 30 ads that are active now, from Meta Ad Library or the client's Ads Manager. For each ad: body text, the date it started running, format, destination URL (from Ads Manager if the client gave access, otherwise the page you reached from the ad's button). [PASTE THE ADS HERE] STEPS 1. Give each ad one angle, judged from its first sentence: PROBLEM (names a pain), IDENTITY (speaks to who the buyer is), USE CASE (a situation or occasion), OFFER (price, discount, bundle, free gift), PROOF (reviews, numbers, press), OTHER. 2. Run length in days = today minus start date. Find the longest-running ad in each angle. 3. List the distinct destination paths used by each angle. 4. Flag each URL that has any of: no utm_source or utm_campaign; ?page=; _pos=; _sid=; a hard-coded fbclid= value; http:// instead of https://; /collections/ followed later by /products/. 5. For each angle with 2 or more ads, name the pasted URL whose path best fits it. Exists? is YES if one fits, NO if every ad in that angle goes to the homepage or a general collection. If no angle has 2 ads, write None. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Ad Library shows the start date but no spend or results (ads shown in the EU or UK also show reach and targeting). Run length is a signal, not proof an ad works. Never say an ad performs. 9. Judge the angle from the ad text only, never from the URL. Label the no-UTM flag ASSUMPTION for any URL that did not come from Ads Manager. OUTPUT Table 1 with columns: Angle | Ads | Longest running (days) | Example hook | Landing pages. Write Longest running as days and the two dates, for example 41 (5 Oct minus 25 Aug). Quote the hook. Table 2 (15 rows at most, most issues first) with columns: URL | Issue Table 3 with columns: Angle | Best page for it | Exists? TO HERE ============================================================ A06 Message match batch The top ads scored against their landing pages, worst first. About 10 minutes for one ad, about 30 for five. Use when: Day 3, for the client's top ads. About 30 minutes for five ads. Inputs to paste: - The ads: text, creative, audience, URL. Ads Manager: pick the top ads by spend yourself and paste the copy. Describe each creative in words. Delete first: Spend and results. - Each landing page: first screen and full text. On a laptop, open the page in Chrome's device mode at 390 x 844, copy the first screen, then select all. Delete first: Reviewer names and customer details. - Never paste: Spend, unless the client approved. Pairs with: - L01 Headline repeats the ad's promise (Scorecard 01) - F10 Ad to page match (Pack 03) Where it goes wrong: A score with no quote, or Product above 1 with no hero image described. Delete any score without a quote. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A06 MESSAGE MATCH BATCH ROLE You are a CRO and paid media specialist at an ecommerce agency. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. Ads: the [NUMBER OF ADS] top ads, chosen by me. INPUTS For each ad: primary text and headline; a description of the creative (what the image or video shows, colours, product and variant); audience COLD or WARM; destination URL. For each page (paste it once if several ads use it): the first-screen text, then the full page text. [PASTE THE ADS AND PAGES HERE] STEPS 1. Score each ad against its page on six points, 0 to 2 each (2 = same promise in the same words, 1 = partly or in other words, 0 = missing or contradicts). Headline: the page headline repeats the ad promise. Offer: the ad offer shows on the first screen with the same terms; no offer in the ad scores 2. Product: the first product shown is the ad product, same variant. Angle: the page leads with the ad angle. Awareness: COLD gets what it is and who it is for first, WARM gets the offer and a button first. Proof: the first screen shows the kind of proof the ad used; no proof in the ad scores 2. 2. Score Product 1 at most if I gave no description of the page hero image. 3. Write each score as n plus a short quote from each side, for example 1 (ad: "Buy 2, get 1 free" / page: "Free gift with 2"). Total of 12 is the sum of the six scores. 4. Rank the ads worst first. For the worst ad, list the changes in the order that gains the most points, then a last row Total: its score now to the score if all are made. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Score only the text I pasted. Never use the URL or what you know about the brand. Write changes in the ad's own words. OUTPUT Table 1, ads worst first, with columns: Ad | Headline | Offer | Product | Angle | Awareness | Proof | Total of 12 Table 2, for the worst ad only, with columns: Order | Match point | Change | Score now to target Then one line: A 2 equals a tick in Scorecard 01; a 0 or 1 is not ticked. TO HERE ============================================================ A07 Scorecard to findings Up to 10 findings in one template, each graded by its sources. About 20 minutes. Use when: Day 4, when the scorecards are done and you have the outputs of A02 to A06. Inputs to paste: - Failed checks with Tier and Effort. Your scoresheets from scorecards 01 and 02, plus A03's Open first line. Delete first: Client names in file names. - Earlier outputs and notes. A02 to A06 outputs and your recording or survey notes. Delete first: Names, emails, order numbers. - Never paste: Raw recordings or customer messages. Pairs with: - Scorecards 01 and 02 Where it goes wrong: Two failed checks from one scorecard counted as two sources. Recount the Sources column. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A07 SCORECARD TO FINDINGS ROLE You are a senior CRO analyst at an ecommerce agency. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS Scorecard results: failed check ids and titles with their Tier and Effort tags, my notes, and A03's Open first line: [PASTE SCORECARD RESULTS HERE] Outputs from earlier steps (A02 to A06), as many as I have, and any anonymised recording notes or survey answers: [PASTE EARLIER OUTPUTS AND NOTES HERE] STEPS 1. Number every observation I pasted: S1, S2 and so on (one failed check, one table row or one note each; the Open first line is not numbered). Keep the numbers the same throughout. 2. Group observations about the same problem into one finding. Write at most 10 findings. Drop observations that do not bear on lost orders. 3. Grade each finding: TRIANGULATED if 2 or more independent sources back it (a different tool or kind of evidence: analytics export, scorecard observation, customer words, ad data), otherwise HYPOTHESIS. Failed checks from one scorecard count as one source. Sources in the summary is that count. 4. Priority: HIGH only if the finding rests on a failed Critical check or its problem is on A03's Open first line; MEDIUM if a failed High check; else LOW. Effort: as the scorecard tags it (Minutes, Days, Project), else ASSUMPTION. 5. Rank: HIGH first, then TRIANGULATED before HYPOTHESIS, then smaller effort first. 6. Why it costs money uses only my numbers, with arithmetic, and no dollar value; the next step adds the dollars. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. One problem per finding. Never split one to add findings or merge two to save space. 9. A source counts toward the grade only if it shows that finding's problem. A number that only sizes the cost is context. OUTPUT For each finding, these lines in this order: N. HIGH: Title (or MEDIUM or LOW) What we found: ... Evidence: S1: ... S2: ... Grade: HYPOTHESIS or TRIANGULATED Why it costs money: ... Fix: ... Priority, effort: ... Then a summary table with columns: # | Finding | Sources | Grade | Effort TO HERE ============================================================ A08 Dollarise findings An upper estimate per month for each finding, with the working. About 10 minutes. Use when: Day 4, after A07, when every finding needs a monthly figure and its working. Inputs to paste: - Findings from A07. The numbered titles and grades. - Page numbers for each finding. Analytics > Reports > Sessions by landing page, with Conversion rate and Orders added. Client to confirm. - The store's own comparable page. Same page type, best rate in the same report. - Never paste: Margins and cost data. Pairs with: - F02 Device gap (Pack 03) - F03 Landing page league table (Pack 03) Where it goes wrong: One gap counted for every finding, which pads the sum. Only the first finding on a page has a value. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A08 DOLLARISE FINDINGS ROLE You are a CRO analyst at an ecommerce agency. You only use the numbers I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. Overlap percent and my reason: [OVERLAP PERCENT AND REASON]. INPUTS The findings from the previous step: [PASTE THE FINDINGS HERE] For each finding: the page or segment, its sessions, the months in the range, its conversion rate and orders, the store AOV, and the rate and orders of the store's own comparable page (same page type and source if possible): [PASTE THE PAGE NUMBERS HERE] STEPS 1. Monthly sessions = sessions / months. 2. Gap = target rate - current rate, in percentage points; the target is the rate of the store's own comparable page. If there is none, or it is not higher, write "Not enough data" and no value. 3. Monthly value = monthly sessions x gap x AOV. Show rates as decimals: 6,000 x 0.006 x 45 = 1,620. 4. Findings on the same page or segment share one gap: put the value on the first, write "shared with #n" on the others, and count it once. 5. Confidence is LOW if either page has fewer than 30 orders (a rule of thumb), MEDIUM if the grade is HYPOTHESIS, else HIGH. Lowest applies. 6. Raw sum = the counted values. After overlap = raw sum x (1 - overlap percent); the percent applies only to values from different pages that share visitors, and is 0% on one page. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Targets come only from the store's own pages that I pasted. Part of any gap is the product, not the page. If the comparison page is a different product, confidence is at most MEDIUM and Working says why. 9. Call every value an upper estimate, not a forecast. Never promise a lift or show the raw sum to the client. 10. Date rule: my range starts on or after 24 Sep 2026, or ends before 21 Sep 2026. Mention these dates only if my rows cross 21 to 23 Sep 2026: then say so and compare orders and sales, not sessions or conversion rates. OUTPUT Table 1 with columns: Finding | Monthly sessions | Current to target rate | Gap | Monthly value | Working | Confidence Then these three lines: Raw sum (internal only): Overlap percent: After overlap: per month, an upper estimate, not a forecast TO HERE ============================================================ A09 Prioritise and 30/60/90 ICE scores, buckets and a 30, 60 and 90 day plan. About 15 minutes. Use when: Day 4, after A08, when the client needs the work in order and by month. Inputs to paste: - The table from A08. Without the raw sum line. - Effort and FIX or TEST. From the scorecard tags and your judgement. - Team capacity. Who does what, hours a week. Delete first: Names and rates: use roles. - Never paste: Rates, salaries, what the client pays. Pairs with: - F11 Fix or test (Pack 03) - F12 30-day fix plan (Pack 03) Where it goes wrong: Month 1 filled with tests, or a score moved by more than 1 with no reason. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A09 PRIORITISE AND 30/60/90 ROLE You are a CRO programme lead at an ecommerce agency. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS The dollarised table from the previous step: [PASTE THE TABLE HERE] For each finding: effort (Minutes, Days or Project) and FIX or TEST. Then team capacity: who can do what, and hours per week: [PASTE EFFORT AND CAPACITY HERE] STEPS 1. Impact (1 to 10) = Monthly value / highest Monthly value x 10, rounded, minimum 1. Findings that share a value share the Impact. A finding with no value gets 1. 2. Confidence (1 to 10) from my Confidence column: HIGH = 8, MEDIUM = 5, LOW = 2. Ease (1 to 10) from effort: Minutes = 9, Days = 6, Project = 3 (a rule of thumb). Move a score by 1 at most, and give the reason. 3. Write each score as the number, a colon and a reason of up to 8 words. ICE = (Impact + Confidence + Ease) / 3, to one decimal, with the sum shown. 4. Bucket: FIX NOW = FIX with effort Minutes or Days; PROJECT = FIX with effort Project; TEST = TEST. Order each bucket by ICE. 5. 30/60/90: Month 1 holds FIX NOW items only, no tests, within my capacity (label any hours you assume ASSUMPTION). Month 2: PROJECT items and the first TEST. Month 3: the next TEST and a re-score of the changed pages. One test at a time. Owner is a role from my capacity list. Write None for a month with no items. 6. Evidence we'll have names what we can show at the end of the period, such as a dated before and after or a test decision. It is never a forecast. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Do not forecast revenue or lift for any period. OUTPUT Table 1 with columns: Finding | Impact | Confidence | Ease | ICE | Bucket Table 2 (Period is 30, 60 or 90) with columns: Period | Work | Owner | Evidence we'll have TO HERE ============================================================ A10 Test plan A test brief with the sample size and the decision rule. About 10 minutes. Use when: Day 4, for the one finding you will test first. Inputs to paste: - One finding with its evidence. The finding block from your A07 output. - Page sessions and orders, 4 weeks. Analytics > Reports > Sessions by landing page, Orders added. - The variant idea and the smallest lift. Your idea in 1 or 2 sentences; a lift such as 20%. - Never paste: Customer quotes with names. Pairs with: - F11 Fix or test (Pack 03) Where it goes wrong: A variant kept on a better number alone. Re-run n yourself and check the rule names the z-test. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A10 TEST PLAN ROLE You are an experimentation lead at an ecommerce agency. You only use the numbers I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS One finding with its evidence (the block from A07): [PASTE THE FINDING HERE] Page sessions and orders, 4 weeks (see date rule): [PASTE SESSIONS AND ORDERS HERE] The variant idea and the smallest relative lift worth detecting, 20% if none (a rule of thumb): [PASTE THE VARIANT IDEA HERE] STEPS 1. p = orders / sessions for the 4 weeks. Under 30 orders (a rule of thumb), say so in the Sample size row. 2. Visitors per version n = 16 x p x (1 - p) / (p x lift)^2, lift as a decimal (20% = 0.2): a crude planning estimate (80% power, 5% significance, two-sided): a floor. 3. Weeks needed = (n x 2) / (sessions / 4), rounded up. Over 4 weeks (a rule of thumb), the verdict is "Do not A/B test; ship and watch". 4. Hypothesis: Because [evidence], changing [element] for [who] will move [metric]. One primary metric, 2 guardrails from my paste. 5. Pre-launch QA: 5 checks for a person (what to open, on what device, what should show): variant, tracking, original, traffic split. 6. Decision rule, written before launch. A/B test: at the planned sample keep the variant only if the primary metric is better and a two-proportion z-test is significant at 5%, two-sided. A guardrail blocks only if it is significantly worse. Otherwise keep the original. Never stop early. Ship and watch: a before and after cannot prove a lift; report a dated observation on orders, not conversion rate, for the 4 weeks before and after. 7. z-test: z = (p2 - p1) / sqrt(p x (1 - p) x (1/n1 + 1/n2)), p pooled; show as a / b = c. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Test one change. If the idea changes two, say so and ask which to test. 9. Never predict the result. 10. Date rule: my range starts on or after 24 Sep 2026, or ends before 21 Sep 2026. Mention these dates only if my rows cross 21 to 23 Sep 2026: then say so and compare orders and sales, not sessions or conversion rates. OUTPUT Table 1 with columns: Field | Entry, one row each for: Hypothesis, Variant, Primary metric, Guardrails, Sample size (formula and working), Duration, Pre-launch QA, Decision rule Then one line: Verdict: A/B TEST, or Verdict: Do not A/B test; ship and watch. TO HERE ============================================================ A11 Red-team the report The errors a client would catch in your draft, before they do. About 15 minutes. Use when: Day 5, on the draft report, before the client sees it. Inputs to paste: - Draft report text. The report as plain text, from the heading to the last line. Delete first: Logos, names and contact details. - The tables it quotes. The A02, A03 and A08 outputs, and any export a figure came from. - Never paste: Screenshots of the client's admin. Pairs with: - Scorecards 01 and 02 Where it goes wrong: Flags with no quote behind them. Drop any flag that has no quoted claim and no conflicting row. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A11 RED-TEAM THE REPORT ROLE You are a sceptical client-side analyst who checks agency reports. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS The draft report text: [PASTE THE REPORT TEXT HERE] The tables and exports it takes its numbers from: [PASTE THE TABLES HERE] STEPS 1. Number every claim in the report: each figure, percentage, comparison or cause (words such as because, will, caused). 2. Test each claim against these problems, one row for each problem found. ARITHMETIC: recalculate from my tables. NO SOURCE: no pasted row supports it. SMALL SAMPLE: a rate built on fewer than 30 orders (a rule of thumb). DEFINITION MISMATCH: one name for two figures with different sources, ranges, gross vs net, sessions vs people, or session numbers from before and after Shopify's September 2026 change. OVERCLAIM: promises a lift, treats a hypothesis as fact or names a cause the data cannot show. ADDED FINDINGS: values of overlapping findings added together. 3. In the Where column write the heading or paragraph number. Fix is the corrected sentence or the missing source, using only my paste. 4. Count the claims checked and the claims flagged. RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Flag a claim only if you can quote it and point to the row that conflicts with it or to the missing source. 9. Do not review style. Do not suggest new findings. OUTPUT Table 1 with columns: Claim (quote) | Where | Problem | Fix Then one line: Flagged: n of m claims checked. TO HERE ============================================================ A12 Client summary A one-page summary for your client. About 10 minutes. Use when: Day 5, after A11. Run it again each month with last month's summary. Inputs to paste: - Final findings, plan, two strengths. Your A07 findings after A11, the A09 plan, two things that work. - Values, orders and sessions. Your A08 output and the current online store orders and sessions. - Last month's summary (optional). Your A12 output from last month, or NONE. - Never paste: Internal notes and margins. Pairs with: - Scorecards 01 and 02 Where it goes wrong: A number the client never saw, or the raw sum. Check every figure against A08. Safety: Before you paste: delete names, emails, phone numbers, addresses and order numbers. After: check every number against your export. AI tools can miscalculate and can invent figures. COPY FROM HERE A12 CLIENT SUMMARY ROLE You are a CRO consultant writing for a founder with no analytics background. You only use what I paste. CONTEXT Client: [CLIENT CODE NAME]. Sells [WHAT THEY SELL] to [WHO BUYS]. Date range: [START DATE] to [END DATE]. Currency: [CURRENCY]. INPUTS The final findings (title, grade, fix, owner, timing), the 30/60/90 plan, and 2 things that already work, with their evidence: [PASTE FINDINGS AND PLAN HERE] The dollarised table (raw sum, overlap percent, after-overlap figure) and the current online store orders and sessions: [PASTE VALUES HERE] Last month's summary, or NONE for a first one: [PASTE LAST SUMMARY HERE] STEPS 1. Copy every number from my paste unchanged, except where step 3 says to calculate. Add no figure. 2. Headline: a claim a founder can repeat, under 12 words, from the number 1 finding. 3. Give the current rate as online store orders per 100 visits: orders / sessions x 100 = c. Never call it conversion rate. 4. Use the after-overlap monthly figure only, never the raw sum, as monthly value at stake (upper estimate, overlap removed). 5. Choose 2 strengths with their evidence. If I gave none, write "Not enough data". 6. Top 3 fixes: what, who and when, from my plan. 7. What we cannot tell you yet: each finding graded HYPOTHESIS, worded "not yet confirmed" (print no grade names), and the export that would settle it. 8. If I pasted last month's summary, the headline says what changed; each Value reads "now (last month: x)". RULES 1. Use only what I pasted. If something you need is missing, write "Not enough data" and name the report or export that would answer it. 2. Show the arithmetic for every number you calculate, written as a / b = c. 3. Do not use industry benchmarks, averages or statistics unless I pasted them. 4. Quote the exact row, number or sentence you are judging. 5. Label anything you infer as ASSUMPTION. 6. Never write copy that states a fact I did not give you (prices, reviews, results, certifications, delivery times). 7. Output only the sections listed under OUTPUT, in that order. No introduction and no closing summary. 8. Write for a founder. Do not use CVR, AOV, PDP, ICE, funnel, haircut or hypothesis. Say it in plain words. 9. Never promise a result, a lift or a date for one. 10. Date rule: my range starts on or after 24 Sep 2026, or ends before 21 Sep 2026. Mention these dates only if my rows cross 21 to 23 Sep 2026: then say so and compare orders and sales, not sessions or conversion rates. OUTPUT Headline: Table 1 (3 rows) with columns: Number | Value | Where it comes from. The rows are: online store orders per 100 visits now, monthly value at stake (upper estimate, overlap removed), and the number 1 finding Table 2 (2 rows) with columns: What is already strong | Evidence Table 3 (3 rows) with columns: Top fix | Who | When Table 4 with columns: What we cannot tell you yet | Export that would settle it Then: Next 30 days, 3 bullets. TO HERE ============================================================ Shopify, Google, Google Analytics, GA4, Meta, Facebook, Instagram, Android, Chrome, ChatGPT, Claude and Gemini are trademarks of their owners. This guide is independent and not endorsed by any of them.