What Did That GEO Experiment Actually Improve?

A supplier says a content change increased AI visibility. Before deciding whether to repeat it, ask what had to happen for the page to enter the experiment.

Was it discovered by a search engine? Was it selected from a larger collection? Or did the researchers place it directly in front of the model and measure how often the answer cited it?

All three can be useful experiments. They answer different questions.

A critical survey of 45 GEO studies makes that distinction central. It finds evidence that changing already-retrieved material can change how an answer uses or cites it. The stronger claim—that a technique produces durable gains in organic discovery and customer outcomes across platforms—requires a longer chain of evidence.

A page has several chances to drop out

An AI product may answer without searching. If it searches, the relevant page must be available to the retrieval system. The system then selects material, allocates limited space to it and constructs an answer. The answer may cite the page, use a fact without prominently naming the business, or omit it entirely.

Finally, a person may read the answer, click a source, continue researching elsewhere or stop. Citation presence is one event along that path.

This does not mean every product follows one publicly observable sequence. It is a way to identify the claim being tested. A researcher with control over a retrieval system can inspect stages that a business using a consumer interface cannot see.

Suppose a study supplies the same five pages to a model, changes the wording on one, and observes more citations to that page. It has evidence about use after source entry. It has not shown that the edited page becomes easier to find on the live web. That hypothetical design is not defective; the mistake would be expanding its conclusion.

A gain at one stage can meet a loss elsewhere

The survey discusses cases where citation-oriented edits did not survive an end-to-end evaluation. A change can make a passage attractive once supplied while affecting whether the retrieval system selects it in the first place.

Context position is another example. Experiments can move a source earlier or later in a packet and observe what changes. A website owner generally cannot choose where a commercial assistant places that page in its internal context. The result explains a mechanism without necessarily providing a publisher-controlled tactic.

The survey also reports heterogeneous results across tasks, engines and methods. Its discussion of three significantly positive results among 54 method–domain combinations refers to the C-SEO Bench study it reviews. It is not the survey's own new optimization experiment. Such variation is a reason to keep the setting attached to a result, rather than averaging unlike gains into a universal GEO percentage.

Absence of strong evidence is not proof of no effect

The review finds no established, stable, longitudinal, cross-platform causal result covering the full path from optimization to organic discovery and downstream behavior. That is a statement about the evidence it reviewed.

It is not proof that GEO can only work briefly, that content improvements never help, or that all production evidence is worthless. The survey discusses limited suggestive production and referral findings too. Those can justify further investigation without settling a general claim.

The same restraint applies to business outcomes. A citation increase may matter even if it does not send a click. It might change the information a buyer sees. But if the claim is increased traffic or sales, those outcomes need their own measurement; they cannot be inferred from citation count alone.

Put the measured result beside the promised result

For a proposed test or vendor report, write down four things:

• The input: what changed, and what remained fixed.
• The entry condition: how the source reached the system.
• The measured outcome: retrieval, citation, brand representation, visits or another defined event.
• The boundary: product, mode, query set, dates and number of repetitions.

Then compare that record with the recommendation. An edit tested inside a fixed source packet can support a claim about that packet. Applying it to a live site is a separate hypothesis that needs a suitable comparison.

Use the stages you can observe. Save exact questions, complete answers and visible sources; check page accessibility and relevant analytics where available. If internal retrieval or context allocation is hidden, say so. An honest gap in the measurement is more useful than a diagram that quietly treats every stage as known.