Seasonality Adjustment

01What it is and how it works

AI search models surface brands differently based on seasonal query patterns. For example, queries for 'gift ideas' surge in December, and brands that appear in those results see a temporary spike in mentions. Seasonality Adjustment uses historical data — typically at least 12 months — to model a baseline seasonal curve. The curve is built by decomposing the time series into trend, seasonal, and residual components using methods like moving averages or STL (Seasonal and Trend decomposition using Loess). The seasonal component is then subtracted from the raw metric to produce a seasonally adjusted value. This reveals the underlying trend without the noise of calendar effects.

02What to do about it

First, collect at least 12 months of daily or weekly brand appearance scores from your AI search measurement tool. Choose a seasonal decomposition method — for most teams, a 12-month moving average or STL works well. Apply the decomposition to your time series and extract the seasonal component. Subtract it from the raw data to get the adjusted series. Update your dashboards to show both raw and adjusted lines. Re-run the decomposition quarterly to account for shifts in seasonality. If your tool supports it, enable automatic seasonality adjustment in the settings. For brands with multiple product lines, run separate adjustments per category.

03How it is measured or noticed

You notice seasonality when raw metrics show regular peaks and troughs that align with calendar events — holidays, back-to-school, tax season. To measure it, plot the raw time series alongside the seasonally adjusted series. The seasonal component itself can be extracted and plotted as a separate line; its amplitude tells you how much the calendar affects your brand. A common diagnostic is the 'seasonal ratio' — the raw value divided by the adjusted value. If the ratio is consistently above 1 in certain months, seasonality is present. Also compare year-over-year changes: if raw numbers are flat but adjusted numbers are rising, seasonality is masking growth.

04Common mistakes

• Using less than a full year of data — the model cannot identify a yearly cycle without at least 12 months.
• Applying adjustment to short time windows like a week — seasonality is a multi-month pattern.
• Ignoring multiple seasonal cycles, such as weekly patterns within a yearly cycle.
• Assuming the seasonal pattern is identical year after year — it can shift due to changing consumer behavior.
• Confusing seasonality with trend or noise — a sudden spike from a viral post is not seasonality and should not be removed.
• Over-adjusting by removing too much variation, which flattens real signals.

05Limits

Seasonality Adjustment assumes patterns repeat annually. It fails for one-off events (e.g., a product recall, a viral meme) or when the brand fundamentally changes its strategy mid-year (e.g., a rebrand). It is often confused with trend smoothing or anomaly detection — those are separate techniques. The adjustment does not predict future seasonality; it only removes past patterns. For brands with less than 12 months of data, the adjustment is unreliable. Also, if your AI search measurement tool changes its methodology (e.g., a new model version), the historical baseline may become invalid and require recalibration.

06Worked example

In January, Brand A's raw appearance score in AI search was 120, up from 100 in December. But the seasonal component for January is -20 (a typical post-holiday dip). The adjusted score is 140, showing the brand actually gained 40 points after removing the seasonal drop. Without adjustment, the flat line would have been misinterpreted as a loss.

Frequently asked questions

More in Measurement