SEO for Prophet

This guide is to help users make use of my Prophet Forecasting For SEO App built in python and hosted on Streamlit.

Skip to forecasting tips, if you want some techniques on getting the most out of it.

Using Streamlit Prophet App

1) Upload your data

• Click “Upload your CSV file”.

• File requirements:

  • Column 1: Date (a format like YYYY-MM-DD works best. Excel - saving as csv - can get this wrong, so be aware!).

  • Column 2: Metric to forecast (clicks, sessions, bookings, etc.).

  • Columns 3+: Optional numeric regressors (e.g., Google Trends index, campaign flags, availability).

The app will show you a preview and confirm which columns were detected.

2) Validate your data

• Check that:

  • Dates parsed correctly (no blank or “NaN” values).

  • Metric column contains numbers (commas are auto-removed).

  • Regressors, if present, are numeric.

• If you see errors here, fix your CSV and re-upload.

3) Configure the model

To configure the model, you need to help it understand what to expect in the data:

• Changepoint Prior Scale: How sensitive the model should be to trend changes.

  • Low = smoother, fewer shifts.

  • High = more reactive, risk of overfitting.

• Seasonality Prior Scale: How strong seasonal patterns are.

  • Low = muted seasonality.

  • High = stronger cycles.

• Manual Changepoints: Enter comma-separated dates (e.g., 2024-03-05,2024-06-01) for known events (e.g., migrations, Core Updates).

4) Set frequency

• Choose Infer Automatically (default) or pick Daily / Weekly / Monthly / Yearly.

• Prophet assumes evenly spaced data - if inference fails (it can't detect it properly), you can set it manually.

5) Run the forecast

• Click Run Forecast.

• The app will:

  • Train Prophet on your uploaded data.

  • Produce three outputs:

    1. Forecast Matrix - monthly by year, with % change.

    2. Line chart - Year-on-Year comparison.

    3. Forecast & decomposition plots - show actuals, predictions, trend, and seasonality.

6) Review results

• Use the YoY matrix to compare growth/decline by month.

• Use the decomposition plots to see trend vs. seasonal cycles.

• If results look off/wrong:

  • Adjust changepoint prior scale.

  • Add/remove manual changepoints.

  • Check data frequency.

7) Download everything

• Once complete, click Download Results.

• You’ll get a ZIP file containing:

  • combined_data.csv (actual + forecast values)

  • forecast_matrix.csv (YoY pivot table)

  • PNGs of all plots (forecast, components, YoY line chart)

SEO Forecasting with Prophet Tips

Segment your data before uploading

If you want to ensure you have more accurate forecasts that account to differences in demand OR planned performance increases, create separate files for Brand vs Non-Brand, Desktop vs Mobile, or market-level data.

You will find that cleaner segments give Prophet clearer seasonal signals and improve forecast accuracy.

Use regressors when possible

Regressors (and here) are where the power of Prophet really comes in! This enables you to help the model "understand" events that may impact the model. Here are some types you can work with:

• Add a Google Trends index for your main category or anchor keyword - my trend stitcher app is something I have used for this before too!

• Enrich with as much other data IF available:

  • Changes to planned media spend

  • Planned new product launches

  • Availability/stock rates if known or forecastable.

Sometimes adding more regressors helps, other times not - you will need to experiment.

Mark important SEO events

Use manual changepoints for known SEO events which could impact the data

• Core Updates

• Migrations

• Major redesigns

• Template changes

  
  

This helps to prevent Prophet from “smoothing away” sharp changes you know are real.

Pick the right seasonality mode

This should default to Additive. But If seasonal peaks scale as your site grows (common in retail and travel), switch to Multiplicative.

Set realistic forecast times - not too far into the future!

We all know that you may need to do this for the next year - which isn't super-accurate in Prophet, but it is possible. Just be aware that you need to provide more regressors to the model to help accuracy over 6 months.

Forecast 90 - 180 days ahead for most SEO series and re-forecast if fundamental things changes.

Check the YoY matrix first

The month-by-month comparison is often the fastest way to spot acceleration or decline. Use it to decide where to add regressors or changepoints.

Common pitfalls and fixes

We're getting into the weeds here, but some final tips:

• If a Forecast looks too flat, increase changepoint prior scale.

• Forecast is too “twitchy” - lower changepoint prior scale.

• Seasonality looks unrealistic - reduce seasonality prior or try multiplicative mode.