AI is going to touch your BFCM numbers this year whether or not you use any AI tools yourself. The ad platforms buy with it, shoppers research with it, and your team will probably analyze the results with it. Each of those changes what your reports say, and not always in a direction that helps you. Here are the four effects I am planning around.
1. Automated campaigns will chase the easiest sales
Meta’s Advantage+ campaigns and Google’s Performance Max now run a large share of most brands’ budgets. They are good at finding conversions, and during BFCM the easiest conversions are people who already know you and were waiting for the sale. The algorithm will happily spend into them, report a great ROAS, and leave you with fewer new customers than the revenue spike suggests.
What I do about it: split new and returning customer revenue by channel every day of the weekend, and set new-customer goals or exclusions where the platforms allow it. If last year’s BFCM looked amazing in Ads Manager and January was flat, this is usually why. More on reading those numbers in what is new-customer ROAS.
2. AI assistants are a discovery channel you cannot see well
A growing number of shoppers ask ChatGPT, Claude, Perplexity, or Google’s AI answers for “best gifts for a runner under $100” before they ever search for a brand. When those answers mention you, the visit often shows up as referral traffic from the assistant, or as direct traffic when someone copies your name and types it in. Either way, your paid channels may get credit for demand an AI answer created.
Three things help. Track AI assistant referrers as their own channel instead of lumping them into referral. Add an AI assistant option to your post-purchase survey so buyers can tell you. And if you are testing ChatGPT ads, measure them in an attribution tool rather than trusting a brand-new platform’s reporting during its first holiday season. I covered that setup in how to understand the impact of ChatGPT ads before BFCM.
3. More creative, less signal per creative
AI tools make it cheap to produce fifty variants of a BFCM ad. The platforms then spread spend across them, and each variant gets too little data for its numbers to mean much. I have seen teams declare a winner on 12 purchases.
Group creative by concept, offer, or hook, and judge the group. A “30% off sitewide” concept with ten AI-made variants is one test, not ten. Decide the minimum spend or purchases a concept needs before you read it, and write that number down before the weekend starts, when you are less tempted to bend it.
4. AI analysis is fast and confidently wrong
After the weekend, a lot of teams will paste exports into an AI assistant or ask an analytics tool’s built-in AI what happened. That is a genuinely good use of the tools. It is also how you end up with a slick summary built on Meta’s self-reported ROAS, or a year-over-year comparison that silently uses different date ranges.
The fix is the source, not the prompt. Point the AI at deduplicated data rather than platform exports, and make it state the dates, metrics, and attribution model it used. Several attribution tools now connect to AI assistants directly. ThoughtMetric, for example, has AI Connectors for Claude and ChatGPT that read from its multi-touch attribution, and it tracks ChatGPT Ads among its ad integrations. (Disclosure: ThoughtMetric sponsors this site.) Whatever you use, spot-check the first answers against the dashboard before you forward anything to leadership.
What will not change
Shopify’s order count is still the ground truth, MER is still the cleanest single efficiency number for the weekend, and the real test of BFCM is still whether the customers you acquired come back in Q1. AI changes how you get to those numbers. It does not replace them.
Before November 26
- Set up a daily new versus returning split by channel, so automated campaigns cannot hide behind returning buyers.
- Create an AI assistant channel and add it as a survey option.
- Group AI-generated creative by concept and set a minimum sample before judging.
- Decide which data source your post-BFCM AI analysis will read from, and make it a deduplicated one.
Leave a Reply