Black Friday is Nov 27, and at the typical brand it brings about 4.6 times a normal October day’s orders, so every problem gets bigger with it: a promo code leaking to coupon sites, a break on the path to purchase, a channel getting credit for sales your customers say started elsewhere, an influencer your code report undercounts.
Since the MCP launched in August, hundreds of brands have used Fairing’s MCP server to ask Claude about their how did you hear about us survey data. This is the BFCM edition of the checks they run. It covers:
where your direct traffic really comes from
a clean October baseline
where your tracking and your customers disagree
which codes are leaking
which influencers customers actually name
what customers who found you through AI search are worth
where the path to purchase is breaking, and the customer words worth reusing in your sale emails and ads
Each has a starter prompt, follow-ups and, if relevant, a BFCM benchmark. Every prompt runs in Claude through Fairing’s MCP.
The prompts work year-round; the benchmark boxes are for BFCM.
Benchmarks come from three things we’ve published: BFCM statistics and benchmarks, built from order data across 2,200+ brands; how customers hear about you at BFCM, built from survey answers across 800+; and our BFCM CRO checklist, built from answers to an order-issues question. Report figures describe the typical brand; the checklist figure pools all answers. Check them against your own October.
1. Find the word of mouth hiding in your direct traffic
Direct traffic tends to jump at BFCM. A direct order has no UTM tag and no referring site: the customer typed your URL or searched your name. Your survey shows how they heard about you.
A starter prompt:
Prompt
Take my How Did You Hear About Us responses from Oct 1 to 31 and group them into orders with no UTM and orders with any UTM. Show each answer's count and share for both groups side by side, and how many points higher friends and family is in the no-UTM group. Work in pages of 100.
A follow-up to see how much of that word of mouth is gifting:
Prompt
For Nov 27 to Dec 1, 2025 and for Oct 1 to 31, 2025, count the friends-and-family answers and write-ins that mention a gift, a present or someone asking for the product. Give that count and its share of all friends-and-family answers for each window. If I run a separate gift question, use it instead: for each window, show what share of friends-and-family customers said the order was a gift, with counts.
If you run a gift-or-for-yourself question, ask Claude to compare channel mix and average order total for gift orders against self orders.
Direct traffic likely hides word of mouth, and more so at BFCM.
Someone who hears your name from a friend and types it in leaves no tag. At the typical brand, about two-thirds of customers who named friends or family had no UTM tag. Direct orders also named friends or family more often than the brand’s other orders: 7 percentage points more at BFCM, against 4 points more in October. Not all of it is word of mouth, since 29% of untagged orders named a social platform, but a BFCM jump in direct traffic is likely partly word of mouth. Run the starter prompt again on the sale window to see how much of yours is. (friends-and-family findings)
The same lift, three years running.
The median brand’s friends-and-family share rose about 4.6 percentage points over its own October in each of 2023, 2024 and 2025. Gift buyers may be part of it, which is why the gift prompt is worth running.
Change in friends-and-family share of answers, October to BFCM, typical brand: +4.6 points in 2023, +4.7 in 2024 and +4.6 in 2025. Each year uses that year’s brands. Source: how customers hear about you at BFCM, finding 1
2. Measure the sale against October, not “the last 30 days”
Your October channel mix is your survey answer share by channel in a normal month. Every sale-week comparison measures against it. At the typical brand, orders step up about 8% on November 1 and keep climbing, so a rolling “last 30 days” window starts picking up holiday shoppers within days. Pin the window to Oct 1 to 31, and pull it now so you know your normal before the sale starts.
Give me my October channel mix: each channel's share of my How Did You Hear About Us answers for Oct 1 to 31, with response counts and answers per day. Merge answer options that were renamed. Then, for [channel], give one row per week for the 13 weeks ending Oct 31 with its share and response count. Tell me how much that share bounces week to week, and how many points of lift I'd need before calling it more than noise. Show your working.
Keep the answers-per-day column: in sale week you’ll judge each channel on pace, not share.
Judge channels on pace, not share.
Answers per day ran 3.8 times October’s pace at the typical brand. Social ran 3.3 times, so its share fell from about 38% to about 34% while its customer count grew. Flag a channel only if its pace falls well behind your all-answers pace. (how answers shift at BFCM)
3. Find where your tracking and your customers disagree
Ad platforms and UTMs report what they can see: clicks and views. The survey records what the customer says at checkout, tied to the order. Most budget questions start where the two disagree.
A starter prompt:
Prompt
Break down my How Did You Hear About Us answers by channel for Oct 1 to 31, first-time customers only. I'll paste our ad spend by channel for the same dates; if you use a spend aggregator like Pipeboard, export it from there. Give one row per channel: responses, share of answers, share of spend and the gap in points, and tell me how you matched spend channels to answers. Flag the two channels with the biggest gaps.
If a channel’s share of spend runs far ahead of its share of answers, or a podcast or YouTube answer is bigger than its spend share, that channel is a good candidate for a holdout or geo test.
If Meta is your biggest answer, see how much of it is ads:
Prompt
Meta is my biggest answer. For Oct 1 to 31, use my Facebook or Instagram follow-up question, if I run one, to show how many of those customers saw an ad, an organic post or an influencer, with counts and shares. If I don't run one, say so.
Then see which Meta campaigns close sales that customers say started elsewhere:
Prompt
For Oct 1 to 31, make a cross-tab of orders with a UTM source of Facebook, Instagram or Meta: one row per UTM campaign, with columns for orders, order total, and the share of orders where the customer named something other than Facebook or Instagram, with word of mouth and podcast as their own columns. Skip campaigns under 20 orders and flag the 3 with the highest share. Work in pages of 100.
Campaigns that show up here are getting last-click credit for customers who say they heard about you elsewhere. That doesn’t mean the ad did nothing, but it’s worth knowing before BFCM ROAS sets next year’s budget.
The UTM shows the last click. The customer names where they first heard.
Both can be true, and reports tend to favor what closed the sale over what started the demand. For paid-social-tagged revenue at the typical brand, the share where the customer named something else rose from about 18% in October to about 27% at BFCM. At BFCM, half of brands fell between 17% and 42%. Use the gap to decide where to test, not what to cut. Run the cross-tab on Oct 1 to 31 to see your own. (paid social finding; what last-click attribution misses)
One more check, for customers whose answer and tag disagree:
Prompt
For Oct 1 to 31, find customers who said they heard about us on TikTok or Instagram but whose order UTM points to search or email (for example google, bing or klaviyo). Give one row per answer and UTM source pair, with the count and its share of all TikTok or Instagram answers. Work in pages of 100.
These are customers who say they found you on TikTok or Instagram, though search or email got the last click. Before you credit branded search or email with a BFCM spike, check how much of it started somewhere else.
Check your email and SMS tags. At the typical brand, about half of customers who credit email or text have no email or SMS tag on their order, in both October and BFCM. If the cause is missing tags, it is fixable before code freeze.
Prompt
For Oct 1 to 31, take responses where the customer said email or text. What share of those orders carry a UTM source or medium for email or SMS? For the rest, list the UTM source and medium they do carry, or "no UTM", with counts, largest first.
If yours looks the same, check every Klaviyo flow, campaign and SMS link for a UTM source and medium, and that your link shortener keeps them.
4. Your codes miss more influencers at BFCM. Find them first.
A promo code says someone typed a code. A UTM says where the last click came from. Neither says where the customer first heard of you. The survey does.
A starter prompt:
Prompt
Take my 10 most-used codes from Oct 1 to 31. I'll paste which influencer or partner owns each code and their commission rate. For each code, show orders, the share of users who named that partner, and separate shares for a coupon or deal site, search and "Other". Flag any partner code where under a quarter of users name that partner, and estimate the commission paid on orders from customers who named someone else. Work in pages of 100.
Read the result two ways. An influencer code used by people who don’t name that influencer is likely collecting credit for sales the influencer didn’t drive. If many of those users name a coupon or deal site, the code has probably spread. An influencer customers name often, with few code redemptions, is likely driving sales your code report misses.
A follow-up that finds influencers your codes undercount:
Prompt
For Sep 1 to Oct 31, list each influencer customers name in write-ins or an influencer follow-up. For each, show how many named them, how many used the code I said they own, and how many used no code. Flag the ones where most buyers used no code.
And to see how codes behaved during last year’s sale:
Prompt
For Thanksgiving through Cyber Monday last year (Nov 27 to Dec 1, 2025), list every code survey respondents used, with order counts. What share of code orders used the single most-used code? For each influencer or affiliate code, show the share of its users who named that influencer, a coupon or deal site, or something else. Work in pages of 100.
If a code looks like it has spread, it may be worth swapping it for a fresh one before the sale. And if customers name an influencer more often than their code shows, consider giving those survey answers some weight when you review the partnership. Influencer attribution without a promo code shows how customers naming the influencer fills the gap.
Codes cover less of your sales at BFCM.
At the typical brand, about 37% of answered orders carried a checkout code in October and about 18% at BFCM. That fits brands switching to sitewide offers that apply without a code. Customers who named an influencer used a code about a third of the time, down from more than half, so code reports likely undercount influencers most at BFCM. Where a brand’s code share rose instead, one code named for the sale usually carried most of it, which likely says more about the offer than about who sent the buyer. If you pay influencers on redemptions, it’s worth agreeing how credit works before the offer goes live. (why code use drops at BFCM, how one sale code takes over)
5. Find the influencers and shows your customers actually name
“Podcast,” “YouTube” and “influencer” are categories until somebody reads the names inside them. Brands often use the MCP to read those names and compare them against what they pay each influencer. Click data misses people who watch on a phone and buy three days later, so a show can look quiet in your UTMs while customers keep naming it.
A starter prompt:
Prompt
Read the "Other" and write-in answers on my How Did You Hear About Us question from Sep 1 to Oct 31. Which influencers, shows or communities do customers name that aren't answer options yet? Group spelling variants, give a count for each, and flag any named 5 or more times as a new answer option. Work in pages of 100.
Then put order value next to each name:
Prompt
For Sep 1 to Oct 31, for everyone who picked YouTube or podcast, list the channels and shows they name in follow-ups or write-ins. Show the top 15 names with response count and average order value, excluding zero-total orders from the average.
Then compare how often each name comes up with what you paid them:
Prompt
I'll paste what I paid each influencer and show from Sep 1 to Oct 31. Match it to the YouTube and podcast names from your last answer, with one row per name: times named, amount paid and paid per mention. Flag the 3 named most for what I paid and the 3 that barely show up, and say so for any name with under 10 mentions.
Use it to decide which partners to renew, cut or renegotiate. Add recurring names as answer options with Auto Suggest, and sort the rest with bulk recategorization. With small counts, ask for the sample size behind each name before a hard cut. For shows, see how podcast attribution works with survey answers.
Influencer and podcast answers grow with the sale.
At the typical brand, their answers per day ran 3.1 to 3.6 times October’s pace at BFCM. An influencer whose mentions jump faster than that in sale week deserves a read by name. Run the sweep on Sep 1 to Oct 31, then on the sale window and through mid-December.
6. See what customers who found you through AI search are worth
Brands compare customers who say they first heard about them from ChatGPT, Claude or another AI assistant with everyone else. AI traffic often arrives with no referrer, so the survey is how you see it.
Prompt
Take responses from Aug 1 to Oct 31 where the customer named ChatGPT, Claude, Perplexity, Gemini or another AI assistant, including write-ins. Compare them with all other respondents on average order total (excluding zero-total orders), share on their first order and share that used a code, with counts for each group. Then list what they wrote in any follow-up. Work in pages of 100.
Surveys see far more AI-driven sales than clicks do.
In Fairing’s study of AI-driven sales, post-purchase surveys caught about 7.5 times more AI-influenced orders than click-based attribution. If AI search already sends you customers, BFCM’s research-heavy gift shopping may bring more of them. Run the prompt now and again after the sale. (AI is driving more sales than your data shows)
7. Read what customers tell you at checkout
Fix what nearly lost the sale
Two post-purchase questions pay off fastest in sale week. The first is “Did you encounter any issues in placing this order?”, a “Yes”/“No” question where a “Yes” opens “What issue(s) do you want to share?” Add it from the template library if you do not run it yet. Then set your baseline now, before BFCM traffic starts shifting the numbers in November: pull your October “Yes” rate, or start collecting one today, so you have something to compare sale week against.
Prompt
Compare the last 24 hours' "Yes" rate on "Did you encounter any issues in placing this order?" with my pre-sale rate (October, or since I added the question), with the responses behind each. Group the last 24 hours' "Yes" answers by issue, with counts, and tell me which to fix first.
About 1 in 10 customers who finish an order say something went wrong.
Across more than 12 million answers, about 9% said “Yes” and about 90% said “No”. Those are the customers who pushed through anyway. Your own pre-sale “Yes” rate is the line to watch: a jump in sale week likely points to something to fix while the sale is live, and the grouped “Yes” answers show what. Our CRO checklist post covers the setup.
The second is “What almost stopped you from ordering?” Outside attribution, brands get some of the most use from it as a before-and-after check around a price, offer or page change.
Prompt
Read all answers to "What almost stopped you from ordering?" from Sep 1 to Oct 31. Group them into five or six themes, largest first, with counts and a paraphrase of each. For each theme, say whether the fix looks like page copy, an FAQ, a shipping or returns policy, or a price or offer test, then name the one to fix before the sale. Work in pages of 100.
And if you changed a price recently, check whether the concern moved:
Prompt
Compare the share of "What almost stopped you from ordering?" answers that mention price in the 30 days before [date of price change] and the 30 days after, with the number of responses in each window. If fewer than 30 days have passed since the change, use the same length for both windows.
Pull sale-week and October separately and compare themes. Answers about waiting for a sale point to an offer decision. Sizing or trust answers point to a page decision.
Steal the words that won it
Your BFCM emails and ads are being written now, and the best raw material is how customers describe you. Brands read why repeat customers come back, for box inserts, thank-you notes and ads:
Prompt
Read the answers to [why do you keep coming back / why did you reorder] from Aug 1 to Oct 31. Group them into themes with counts, and give me the 10 phrases customers use most, word for word, that I could use on a box insert or in an ad.
They also read why buyers chose you over the alternatives they considered, and swap the customer’s own words in for invented ad copy. The same pull works on any free-text question:
Prompt
Read the free-text answers to [question] from Sep 1 to Oct 31. Pull the 10 phrases customers use most often to describe what they bought and why, word for word, grouped by theme with counts. Then take the top 5 themes in "What almost stopped you from ordering?" answers for the same dates, and for each, name the phrase that best answers it.
Use their phrasing, not their quotes or names.
Make it a habit
Start with a monthly comparison. Use calendar months, so you always know which days each number covers:
Prompt
Break down my How Did You Hear About Us responses by channel for [month, e.g. Oct 1 to 31] and the month before, with counts and shares for both. Merge renamed answer options. List the top 5 channels, and flag any channel that moved more than 2 percentage points and had at least 20 responses in both months.
On Dec 1, run it with explicit windows: Thanksgiving through Cyber Monday (Nov 26 to 30, 2026) against Oct 1 to 31. Those are the windows the benchmarks use, so your numbers line up with ours.
Plenty of teams run a version on a schedule, such as a Monday overlay of the last seven days of survey shares against the previous week’s spend. Set it up once a decision feeds off it. The Monday prompt asks for the date the data was pulled, since reclassifying “Other” answers can shift shares between two runs. Prefer not to prompt at all? The weekly attribution brief does a version for you.
Prompt
Every Monday, give me the last 7 days of survey shares by channel against the prior 4-week average, with response counts and the date you pulled the data. I'll paste our ad spend by channel; if you use a spend aggregator like Pipeboard, export it from there. Put last week's spend next to each channel and flag any channel where share and spend moved in opposite directions.
A BFCM run order
0 of 10 done
Before the sale
0/5
Check your survey setup before code freeze, so nothing breaks or goes stale once the traffic arrives.
Prompt
List every live question in my post-purchase survey, in order, with its answer options and follow-ups. Flag duplicate questions, any answer option that shows up under two labels in responses from Aug 1 to Oct 31, and any question with no responses in October.
Add the order-issues question from section 7 if you don’t run it yet, so you have a pre-sale “Yes” rate.
Run the top-10 code prompt from section 4 on Oct 1 to 31, and swap any code that looks like it has spread.
Pull your October channel mix (section 2), then compare last year’s sale with last October:
Prompt
Pull my channel mix for Thanksgiving through Cyber Monday last year (Nov 27 to Dec 1, 2025) and for all of October 2025, with answers per day for each channel. Merge renamed answer options, show counts next to percentages, and flag any channel under 20 responses. Then flag channels whose answers per day grew much less than all answers combined.
Check your pace against your own last BFCM. On Monday, Nov 23, compare your Nov 16 to 22 orders with your October daily average, and use the BFCM pacing table to see where you sit next to where you sat last year.
During the sale
0/3
Daily: the order-issues “Yes” rate from section 7.
Every few days: the top-10 code prompt from section 4, with the window set to the sale days so far.
Weekly: compare last week’s survey shares with last week’s ad spend, using the Monday prompt under Make it a habit.
After the sale
0/2
Read through mid-December. At large brands, word of mouth peaks Dec 8 to 10, so a referral or gifting push judged only on the five BFCM days is judged before its peak.
See who the sale brought in:
Prompt
For Thanksgiving through Cyber Monday (Nov 26 to 30, 2026) and for Oct 1 to 31, 2026, group my responses into first orders and second-or-later orders by customer order count. For each group in each window, compare channel mix, share that used a code and average order total excluding zero-total orders, with response counts. If any group has fewer than 20, say so before comparing. Work in pages of 100.
Plan for BFCM buyers to reorder less.
In the 2024 season, first-time buyers from BFCM reordered within 90 days at about 70% of the rate of October’s first-time buyers, and 77% of brands saw the gap. Part of that is timing: October buyers’ first 90 days include BFCM itself, an extra sale to come back for. Looking only at the first 30 days, which mostly avoids that, BFCM buyers still reordered less (3.0% against 3.9%). If your forecasts give BFCM first-time buyers the same repeat rate as October’s, they likely run high. (new customers and reorders)
How to get answers you can trust
Most bad answers come from how the question was asked. Brief Claude up front.
State the window in dates, not “last month.” Use Oct 1 to 31 and Thanksgiving through Cyber Monday, the benchmark windows.
Ask for the count behind every share. A channel at 3% of 150 responses can move several percentage points on noise alone. Widen the window when a group has fewer than about 20 responses.
Exclude zero-total orders when you ask for order value.
Ask it to merge renamed answer options across periods.
Keep free-text pulls focused. One question and 30 to 60 days gives Claude a clean set to theme.
Ask for the date the data was pulled on anything scheduled.
Read an answer as where the customer first heard of you, not proof the channel caused the sale. A gap between survey share and spend share tells you where to test, and the test gives you the budget answer. The MCP is read-only: it changes nothing in your questions, tags or store.
A preamble you can paste before any prompt:
Prompt
Before answering: tell me the exact date range you used, the number of responses behind each figure, and any answer options you merged. Exclude zero-total orders from order value. If any group has fewer than 20 responses, say so instead of ranking it. When reading individual responses, work in pages of 100.
