AI Visibility Seasonality: How to Compare Baselines

Comparing two months can mix changes in AI answers with changes in what you chose to measure. If the sample contains more questions about a promotion, season, or event, the aggregate percentage can move even when brand presence stays unchanged within each group.

Before calling that movement an improvement, keep a reference that separates changes in the question mix from changes in comparable answers. This guide proposes an event register and two reporting views for teams that already have a prompt bank. It does not claim that Mentio has measured a seasonal pattern or that AI recommendations follow a universal calendar.

Identify the series you are comparing

Separate three series that may move at the same time without measuring the same thing:

• Observable demand: searches, internal queries, or aggregated requests that indicate interest in a need.
• Visibility in a fixed sample: answers to the same question families under comparable conditions.
• Visibility under the current mix: results weighted by a distribution of needs that changes between periods.

More demand does not require an assistant to recommend your brand more often for the same prompt. The number of opportunities may change without a change in observed appearance rates. A fixed sample can also remain stable while becoming less representative of seasonal needs.

Label the question each chart answers. Use the prompt bank weighting method to build the weights. The task here is to preserve two time-based views, not to design another business-value formula.

Build an event calendar before looking at the result

Record periods that could affect interpretation before you see the change. Include events relevant to the business and market being studied, rather than a generic holiday list.

Distinguish a recurring event from a campaign you have run only once. Both happening in December does not establish seasonality. Record your own promotions and catalog changes too: they might explain a difference that the report attributes to the calendar.

For events that shift dates between years, align observations by distance from the event. “Two weeks before the start” may be more useful than the same day of the month. Keep the original dates as well so another person can reconstruct the comparison.

Choose three time references and state their limits

Each reference answers a different question. You do not need to publish all three when data is missing, but you should identify which references are unavailable.

A comparable recent period. Match window length, day-of-week composition, event phase, and measurement conditions. This supports operational monitoring; it does not, on its own, separate calendar effects from trends.

The same phase in an earlier cycle. Find the equivalent period in a previous campaign or season. Check whether the market, offer, prompt bank, model, or surface changed. Similar dates do not make incompatible series comparable.

Families less exposed to the event. Retain a group whose need is not directly tied to that calendar. It helps check whether movement is concentrated where expected. Do not call it a causal control unless you can defend that comparison.

Google recommends looking at longer history and equivalent periods when investigating drops in Search traffic. That guidance concerns Google Search; it does not establish that assistant answers follow the same pattern.

Document why you chose the windows before selecting the one with the most favorable result. To evaluate an intervention, use the GEO experiment design.

Keep a fixed-mix view and a current-relevance view

The fixed view retains reference families and weights. The current view retains the same outcome definitions but uses the period's weights, with their source and version. Show both alongside family-level results.

For each family i, retain its observed rate r_i,t, reference weight w_i,0, and current weight w_i,t. Weights in each view must sum to one over the same comparable set.

• Fixed view at t: sum of w_i,0 × r_i,t.
• Current view at t: sum of w_i,t × r_i,t.
• Mix difference at t: current view minus fixed view.

The last difference describes the arithmetic effect of applying different weights to the same rates. It is not a causal estimate of seasonality. To describe change within the sample, compare the current fixed view with the fixed view in the reference period.

Reading the views together prevents two mistakes. If only the current view improves, examine how much comes from giving more weight to families where the brand already appeared. If the fixed view improves, an observed change remains under constant weights, but explanations other than a GEO action are still possible.

Do not assign zero to a family with no valid answers. Show missing coverage and label the aggregate incomplete, or publish an explicitly labeled common subset recalculated for both periods. Silently renormalizing weights could remove the hardest part of the bank.

New seasonal questions can have their own panel. They should not enter the fixed series retroactively as though they had belonged to the original baseline.

What you can say without a year of data

With one episode, you can describe observations during an event, not estimate a repeatable annual pattern. Label the explanation a “calendar hypothesis” and record which future check would allow you to revisit it.

Seek external context without treating it as a replacement for your measurement. Google Trends normalizes search interest on a 0–100 scale; it does not provide assistant conversation counts. GSC records your property's exposure on Google, not total market demand. Retain the source, filters, geography, and period.

Do not divide mention rate by a Trends index to manufacture “seasonally adjusted visibility.” These are different quantities, and that operation requires a model that justifies their relationship. Nor should a source with little data be treated as evidence of no interest.

If you later collect enough cycles to model the series, validate the model on periods not used to fit it. A structure that describes the past well may fail with a new offer or surface. Always retain the observed series alongside adjusted results and explain the assumptions behind the adjustment.

When the baseline needs a break

A change in model, search mode, market, or extraction rule may prevent comparison even when the calendar matches. Mark the break and start a new reference. Link the series only when a documented comparability test supports it.

Also distinguish commercial availability from measurement failure. A retired offer may be a relevant outcome for the question, while a failed run is a missing observation. The decision follows the defined metric, not whichever treatment improves the percentage.

Keep repetitions within each window according to the AI answer variability protocol. Do not extend a window across event phases merely to reach a response count without recording the change.

The record that should accompany the chart

Before sending the report, check that a reader can recover:

1. Event, phase, market, and dates for both windows.
2. Prompt bank version, surface, and measurement rules.
3. Family-level results, valid answers, and missing coverage.
4. Fixed-mix and current views, including their weights.
5. Offer, content, and provider changes occurring at the same time.
6. Available time references and comparisons that could not be made.
7. Supported conclusion, uncertainty, and next check.

In the executive visibility report, state what changed and against which reference. “Coverage increased under the current mix; the fixed view stayed stable” is more informative than crediting the whole increase to optimization.

If you use Mentio, keep this record alongside your analyses and check which data your configuration provides. This method does not announce an automatic seasonal-adjustment feature.

FAQ

Does an increase during a campaign prove GEO improvement?

No. Examine family-level rates, the question mix, and concurrent changes separately. A time comparison, even with fixed weights, does not by itself identify the effect of a GEO action.

Can I measure seasonality with a few weeks of data?

You can describe an event and formulate a calendar hypothesis. A few weeks are not enough to establish a repeatable annual pattern; retain that limitation and plan comparisons with later cycles.

Should I change every prompt when a season starts?

No. Keep a comparable core and add an identified seasonal panel. New questions should not enter the fixed baseline retroactively without recalculating and explaining the series.

Is the difference between fixed and current views the seasonal effect?

No. It is the arithmetic effect of changing weights over the same observed rates. Attributing it to the season requires additional evidence and consideration of other explanations.

What should I do when a family has missing answers?

Record the missing coverage. Do not turn it into zero or redistribute its weight without disclosure. You can label the aggregate incomplete or compare a common subset, recalculated and labeled in both periods.

Want to know if AI mentions your brand?

Discover your visibility in ChatGPT, Claude and Gemini in minutes.

Analyze your brand for free

Related articles