If a client asks how you know your GEO work caused the improvement, a before-and-after chart can’t answer them. During the same period you also updated content, earned mentions, fixed technical debt and did ordinary SEO. So did their PR team. So did Google.

The answer is to hold some pages back, and to say so before you start.

We had the chance to do this last week and did not take it

On 27 September we added in-body internal links to the ten statistics pages on this site. They are our best-performing pages, averaging position 8.2 while the rest of the corpus sits between 21 and 59, and every one of them had zero internal links above the closing block. A reader who stopped two thirds of the way down a 1,200-word page had nowhere to go.

We changed all ten.

That was the mistake. We had a clean cohort of ten near-identical pages and we treated the lot, so in three weeks when the numbers move we won’t be able to say whether it was the links or something else entirely. We should have changed five.

I’m writing this partly so we don’t do it again.

The method

Split the pages you intend to work on into two matched groups. Work on one. Leave the other exactly as it is for an agreed period. Measure both.

The difference between the groups is the part your work is responsible for.

It works because everything else that happens during the engagement happens to both groups. A core update hits both. A burst of brand mentions lifts both. The client’s trade show lifts both. Seasonal demand lifts both. The only systematic difference between the two sets is whether you touched them.

That’s why it survives the objection that kills every other GEO report, which is “you did other things at the same time”. Yes. And those other things applied equally to the pages we did not touch.

Setting one up properly

Match the groups, don’t just split them. Pair pages on the things that predict movement: current position band, impression volume, query intent, page type. Ten commercial pages against ten blog posts gives you two different populations, not a control.

Size it so one page can’t swing it. On a small site that might be five against five. On a large one, thirty against thirty. Be honest that this is judgement rather than a power calculation. A small holdout gives you a direction, not a p-value, and saying so is better than implying a precision you do not have.

Declare it in writing before any work starts. This is the part that matters most. A holdout picked afterwards is just whichever pages made the result look best. If it was not agreed up front, do not present it as evidence.

Agree the window up front too. Structural changes can surface within weeks. Citation authority generally takes three to six months, because it depends on third parties corroborating you. Pick the window to suit the work, and write it down so nobody is renegotiating the finish line once the numbers are in.

Agree what a null result looks like. Decide in advance what size of gap would count as nothing happening. A test you can’t fail isn’t a test, and agreeing this beforehand is the difference between measurement and marketing.

Release the holdout when the window closes. Then do the work on those pages too. A holdout is a temporary measurement cost. It should never become a decision to underserve part of the site, so the engagement letter says when it ends.

What it can’t do

Being straight about the limits, because a method that oversells itself is worse than none.

It can’t measure site-wide work. Organization schema, entity consistency, llms.txt, crawler access and off-site corroboration apply to the whole domain. They lift both groups. So a holdout understates the effect of that work while measuring page-level work cleanly. Say which one you are reporting.

It isn’t a randomised trial. You can’t randomise which pages an assistant chooses to retrieve, and you are working with a handful of pages rather than thousands of patients. This is quasi-experimental. It’s far better than before-and-after, and it isn’t proof in the laboratory sense.

Small sites can’t do it. If the whole site is eight pages, there is nothing to hold back. Be honest and report contribution instead.

Spillover is real. Improving ten pages can lift the domain slightly, which nudges the holdout up too and shrinks the measured gap. That bias runs against you, which is at least the safe direction.

The sentence the client actually wants

Most clients are not asking for causation in the scientific sense. They’re asking for a sentence they can repeat to their own boss without being caught out.

A holdout gives you one that survives scrutiny:

“The twelve pages we rewrote for AI answers moved from X to Y. The thirty comparable pages we left alone moved from X to roughly X. Everything else that happened in that period happened to both.”

Then say plainly what it does not cover, before they ask. That last part is what makes the first part believable.

What we are doing about it

The next engagement that starts here runs with a declared holdout, agreed in the scoping call and written into the engagement letter before any work begins. The method is part of the Synaptic Authority Engine measurement step rather than an optional extra, and we will publish what it shows, including if it shows nothing.

On our own site, the ten statistics pages are already treated and there is no control left to build. That one’s lost. The next change we make here gets split properly, and the prediction goes up before the result does.

If you want to see the kind of movement this would be measuring, the case studies have the citation-share numbers and the timelines attached. If you want to know where your own pages stand before any of this starts, the AI Visibility Score is the free version of that baseline.

Common questions

How do you prove GEO work caused an improvement?

Compare pages you worked on against matched pages you deliberately left alone over the same period. Everything else that happened during the engagement, brand mentions, algorithm changes, the client’s own PR, seasonality, applies to both groups equally. The difference between the two sets is the part your work is responsible for. A before-and-after timeline cannot separate your work from everything else that happened at the same time, which is why it collapses the moment a client asks a hard question.

What is a holdout group in SEO or GEO?

A set of pages you deliberately do not touch, chosen to resemble the pages you are working on, so you have something to measure against. It is borrowed from clinical trials and paid media, where nobody would accept a result without a control. It is standard practice in those fields and almost unheard of in search, which is why search reporting is so easy to argue with.

How many pages should be in the holdout?

Enough that one unusual page cannot swing the result. On a small site that might be five against five, on a large one thirty against thirty. This is a judgement call rather than a calculation: the honest position is that small holdouts give you a directional signal, not a p-value, and you should say so rather than dress it up.

Does a holdout group work for entity and off-site work?

No, and this is its main limit. Organization schema, entity consistency and third-party corroboration apply to the whole domain, so they lift both groups. That means a holdout understates the effect of site-wide work and measures page-level work cleanly. Say which one you are measuring rather than letting the client assume it covers everything.

When should the holdout be agreed?

Before any work starts, in writing. A holdout chosen after the fact is only a selection of whichever pages made the result look best. If it was not declared up front it should not be presented as evidence.

What do you do with the holdout pages afterwards?

Release them. Once the measurement window closes, do the work on them too. A holdout is a temporary measurement cost. It should never become a permanent decision to underserve part of a client’s site, so the engagement letter says when it ends.

What if the client refuses to hold pages back?

Then you cannot claim causation, and the honest move is to stop claiming it. Report contribution instead, show the timeline, and state plainly that other work ran alongside. Most clients accept that once it is explained. The ones who want proof will usually accept a small holdout when they understand it is the only thing that produces it.