SEO forecasting.

Some learnings.

A thread...

“History doesn’t repeat itself, but it often rhymes.”

This is often the case.

You can learn a lot from history, so the more historical data you’ve got , the better.

And the more you know about that data, the better.

Don’t start your forecasts with a narrow view. Broaden your lens.

Use historical traffic data > estimating keyword improvements.

You’ll get more accurate forecasts.

This is your base.

Historical data is a known known.

Some things you have no control over:

- Pandemics: Rare, significant
- Algo updates: Common, significant

Base your projections on your knowns. Use your unknowns for caveats and communicating risk.

Forecast metrics your client cares about.

This is usually organic sessions/revenue.

Not clicks/impressions from Search Console.

Know when not to forecast.

Know when not to forecast, and explain why.

It’s surprisingly effective in pitches.

Why forecast at all?

Because it can help answer:

1. what can we achieve?
2. have we really made a difference?

Because it’s expected. So you might as well be good at it.

It can set the right expectations. “We need more resource to achieve our targets.”

It can be persuasive. “The SEO team exceeded their targets. Perhaps we should...listen?”

Data extraction is easy, but data is imperfect.

If you’re starting out/pitching, ask questions of the data. Why are there missing values? Why is there a peak every February?

As you get going, annotate/log your activity. It’ll make forecasting easier in the future.

Don’t overcomplicate things. Use models you understand.

I like Holt Winters & Prophet because they’re intuitive for non-analysts like me.

Prophet is well-documented and, critically, I know *why* I’m changing an input to fine-tune the model.

*Not all models work for all data*

Evaluate the accuracy of your forecast.

Use most of your data for training, some of your data for testing.

Compare actual values with forecast values (there’s a lot of ways how to do this).

MAPE is commonly used.

Forecasting isn’t enough.

That’s because there’s usually an intervention point.

- new clients
- new objectives
- new products

Forecast trajectory should change.

Present forecast scenarios on top of your base forecasts;

Ie. What if we achieve 10-20% growth on top of the base forecast.

Subjectiveness based on domain expertise is fine.

But what about revenue?

You can use statistical models to forecast revenue in the same way as traffic.

But what if traffic increases? You need to show a relationship between your SEO strategy and the bottom line.

You have less control with revenue forecasts.

Avg. cvr and avg. order value multipliers increase uncertainty.

I tend to create seasonal indices for them both, predict new annual averages and multiply by traffic forecasts.

It’s v simple but good enough to show opps.

Forecasts are often forced upon us.

“We need a forecast from you for the year”

The issue is, you’ll forecast and forget.

That’s a problem because forecasts show you both what you can achieve AND help you know whether you’re on track.

Create two forecasts.

An initial one for the forecast period and a moving forecast.

Forecasts are improved with more recent observations.

So if you’ve got a forecast that updates over time, it can give you a better idea of what you’ll likely achieve. - “Should we change?”

Finally...

Yes, you’ll need to manage & communicate uncertainty.

Yes, you’ll need to re-forecast if something big happens. It happens less than you think (other than 2020 obvs).

Oh...and all of this is in my course https://futurethought.io/forecasting-for-seo-course/