How to know whether your GEO work actually paid off

Incrementality is the result that happened because of what you did, and that wouldn't have happened if you'd done nothing. It's the difference between "my sales went up" and "my sales went up because of this." It sounds obvious, but almost nobody measures it: the usual thing is to see a number rise, take credit for it and move on.

This guide explains how it's really measured, and why in GEO no provider — us included — can give you a perfect proof. We'd rather say so before you hire us.

Why doesn't "it went up" mean "you moved it"?

Because there was almost always something else happening at the same time.

Imagine you publish new content in November and sales rise 20%. It could be your content. It could also be that November is peak season, that a competitor closed, that you were featured in the press, or that the AI model was updated that week for reasons that have nothing to do with you.

The right question isn't "did it go up?" but "would it have gone up anyway?". And to answer it you need something to compare against.

What is a control group?

A control group is the part you deliberately leave untouched, so you can compare.

It's the same thing drug companies do when testing a medication: one group gets the pill, another doesn't, and they compare. Without that second group you wouldn't know whether people got better because of the pill or because the flu was going to pass anyway.

In marketing there are two classic ways to build one:

Holdout test. You split your customers at random: some see the campaign and some don't. If the group that saw it buys more, that difference is your real result.

Matched-market test. When you can't separate people, you separate cities. You pick two markets that have historically behaved the same, turn on the campaign in only one, and compare. If Miami goes up and Tampa doesn't, the difference is yours.

And why can't this be done the same way in GEO?

Because you don't control the channel. Here's the underlying problem, and it's worth saying plainly.

In an advertising campaign you decide who sees the ad, so you can exclude half. In GEO you decide none of that: you don't control who asks ChatGPT, or exactly what they type, or what it answers them. You can't tell the AI "don't mention my business to this user."

In other words: the classic holdout test, as it's taught, is impossible in GEO. Anyone who tells you otherwise is selling something.

What you can do is change what you leave untouched. Instead of separating people, you separate topics, areas or moments.

So how is it measured? Three imperfect ways

1. Leave topics untouched

The most practical one. Instead of splitting customers, you split your own questions.

You take, for example, 40 questions in your category where you don't show up today. You pick 20 at random and work only on those: you write the pages, earn the sources, fix the technical side. You leave the other 20 exactly as they are.

A few months later you compare the two groups. If the worked group improved and the other didn't, that difference is yours. If both improved equally, something external moved them and your work wasn't the cause.

Its limit: topics aren't independent of each other. If you improve the site's overall authority, that can also lift the questions you left untouched — and then you underestimate your own effect.

2. Use time as the control

When you can't split topics, you compare against your own past.

You record how you'd been doing for months, make the change on a specific date, and then compare what happened afterward against what had been happening before.

Its limit: it's the weakest of the three. The world doesn't stand still while you work: if the model was updated, if a competitor came in, if the season changed, all of that gets mixed in with your effect and there's no clean way to separate it.

3. Split by city

If you sell in several markets, you can work on the local content for one and leave another as it is, and compare the business results between the two.

Its limit: it only works if you're present in several comparable cities and have enough volume for the difference to be visible.

Why is it so hard to see the result in traffic?

Because much of the effect doesn't arrive as a click.

When someone sees your business recommended by an assistant, they often don't tap the link: they open another tab and search for your name, or go straight to your site. In your analytics that doesn't show up as "came from AI" — it shows up as a branded search or as direct traffic.

So instead of looking for a single number, you have to look at several signals at once:

• The traffic that does arrive identified. ChatGPT tags the links in its citations, so that part is visible. It's the floor, not the total.
• The increase in searches for your business. If more people start searching for you by name right after the AI started recommending you, that's a strong signal.
• Direct traffic. It rises when people type your address into the browser after seeing you in an answer.

None of the three proves anything on its own. The three together, moving at the same time as your work, are the closest you can get to proof today.

What can be stated honestly, and what can't?

This is the table we care about most in the whole guide.

We call the left column signal. What some would like to sell us in the right column, attribution — and today it doesn't exist with the precision it's credited with.

What does Apoda do with this?

Three concrete things, so you can judge us:

We log every change with its date. Every intervention is recorded and marked on the timeline, so you can see what happened before and after without anyone interpreting it for you.

We report ranges, not a number. Every metric comes with how many runs are behind it. You can read the full answers we keep, one by one.

We say when it can't be attributed. It happens often. When a change coincides with a model update or something external, we say so instead of taking credit for it.

And we also say the uncomfortable part: we can't prove causality to you with a single business and no control group. Nobody can. What we can do is measure with method, show you the raw evidence and be clear about where what we know ends.

If someone offers you certainty in this area, that certainty doesn't exist yet.

Related terms

See also SEO vs. GEO, AI dark traffic and mention and citation.