Sales increased after the ChatGPT advertising campaign started. That is a useful business observation, but it is not yet an explanation. Customers may have begun planning holidays, payday may have arrived or another team may have emailed a promotion during the same week. A before-and-after comparison first needs a calendar review.
The practical question is which time patterns could change your chosen outcome during the periods being compared. This article provides our recommended internal review for advertisers deciding how much a reported movement actually establishes.
Draw a business timeline alongside the outcome
Place advertising changes, price changes, other promotions, stock events and known operational problems on the same timeline as sales. Record when customers encountered each change, not merely when someone approved it internally. A new offer can influence demand before the advertising campaign formally begins.
Separate recurring patterns from exceptional disruptions. Weekdays, weekends and recurring purchase occasions can form seasonal patterns. A temporary stock shortage or major press mention is a distinct event. Both may undermine a comparison, but they should not receive the same explanation or adjustment.
NIST describes seasonality as periodic variation and discusses graphical detection. For this campaign review, we recommend starting with a daily outcome chart and then grouping observations by relevant calendar categories. A table containing only period totals often hides the variation that needs explaining.
Check the composition as well as the duration
Two equal periods can contain different numbers of weekend days. Consider a hypothetical business that consistently receives 80 orders per weekday and 120 per weekend day. A ten-day period containing eight weekdays and two weekend days produces 880 orders. Another containing six weekdays and four weekend days produces 960.
The apparent increase is 80 divided by 880, approximately 9.1%. No advertising effect exists in this example. Calendar composition explains the entire difference. Dividing both totals by ten does not solve the problem: the daily averages are still 88 and 96.
Compare equivalent weekday mixes where that pattern matters. Complete weeks can address this particular imbalance, but they do not resolve moving holidays, pay cycles or rapidly changing seasonal demand. Review incomplete reporting periods as a separate issue when recent data does not cover the entire day or week.
Inspect rates as well as totals, while keeping their denominators explicit. Orders per website session and orders per calendar day answer different questions. Advertising can change the number and composition of sessions, so dividing by sessions may remove part of the business effect the study was intended to measure.
Same daily outcomes, different period totals
- First ten-day period
8 weekdays × 80 + 2 weekend days × 120 = 880 orders.
- Second ten-day period
6 weekdays × 80 + 4 weekend days × 120 = 960 orders.
- The apparent improvement
80 / 880 = 9.1% more orders is explained entirely by the calendar mix in this example.
Give outcomes equivalent time to mature
An older period may include late conversions and processed returns while a newer one remains incomplete. The difference can look seasonal even though it partly reflects follow-up time. Define when each period is mature enough for comparison and preserve the date on which the dataset was retrieved.
OpenAI distinguishes attributed reporting from incoming conversion events. Our recommendation is to keep report dates, business event dates and observation dates separate in the working dataset. Mixing these concepts can move a late order between periods without any change in customer behavior.
Check whether the metric definition changed too. A new form may register more leads without greater demand. Switching from order date to invoice date can reshape the weekly pattern. Correct or disclose these changes before using seasonality as an explanation for the chart.
Justify the comparison series
The same calendar period last year may be informative, but the business could have changed prices, assortment or customer base. Another region might share seasonal demand while facing different local competition. Organic traffic may itself respond to advertising, making it an unsuitable unaffected reference.
Explain why a proposed comparison should reflect similar demand pressure and why the advertising intervention should not affect it. Examine the relationship over multiple earlier periods. Do not select the series simply because it produces the most attractive answer after campaign launch.
Where operational feasibility permits a predetermined contemporaneous comparison, geographic experiment design may offer another route. It requires separate checks of assignment, targeting controls and spillover. This article assumes no access to native geographic experiments or automatic seasonal controls in ChatGPT Ads.
Make sensitivity analysis understandable
Preserve the raw movement alongside adjusted descriptions. Show, for example, a comparison with matching weekday composition and a separate analysis handling a documented outage under a stated rule. If the decision changes when a handful of days is handled differently, the decision-maker needs to see that dependence.
Avoid trying numerous calendar models until a positive advertising effect remains. If the model was developed after outcomes became visible, label the analysis exploratory. A sophisticated model may support planning, but it cannot replace a credible account of what would have happened without the advertising change.
A narrow interval around a model estimate generally does not cover every incorrect assumption about the comparison series. More observations can reduce sampling uncertainty without eliminating systematic confounding. Report statistical uncertainty and unresolved simultaneous changes as distinct limitations.
The minimum detectable effect belongs to a prospective design with specified assumptions and power. It is not a device for turning an existing before-and-after difference into causal evidence. If the business needs a decision about a small effect, redesign the next measurement exercise rather than implying that extra decimal places answer it.
Match the conclusion to the evidence
Use observed growth when you have a before-and-after comparison. Describe a calendar-adjusted difference as such, with its assumptions visible. Preserve the OpenAI reporting extract alongside the internal event calendar so someone else can reconstruct the comparison.
If the higher outcome appears only around launch, investigate novelty versus sustained response as well. The immediate budget decision may be to maintain a bounded observation period covering more representative trading conditions. That is a concrete decision even when the initial increase cannot yet be causally credited to ChatGPT advertising.
Sources and scope
Assess whether a change after a ChatGPT campaign can be explained by calendar composition, recurring demand or simultaneous business events.
Working methods and examples are editorial suggestions. Check current platform requirements and available features before implementation.