AdCyte

How should you test whether VideoGEO is working?

By AdCyte9 min read
Violet editorial illustration of question cards leading to a checked answer document and an AI answer window with a passing gauge.
Video knowledge, beyond the campaign. Editorial illustration for AdCyte.

The Short Answer

Test VideoGEO by defining the change you want to observe, recording a baseline and comparing compatible evidence after publication. Start with whether useful, accurate content was created and made accessible. Then examine sampled AI answers, relevant website visits and customer behaviour as separate measures. Keep the questions, providers and assessment rules consistent, record failed checks, and allow for other changes that might explain the result. A citation is an observation. Proving that your work caused it requires a stronger test.

Three key takeaways

  • Decide what success means before publishing: content quality, observed visibility and commercial outcomes need different evidence.
  • Compare like with like and retain the answers behind the numbers, including unsuccessful checks and inconvenient results.
  • Use early tests to improve decisions. Small samples can reveal useful problems without supporting sweeping claims about performance.

Start with a decision you are prepared to change

“Does VideoGEO work?” sounds sensible until someone asks what counts as working. An accurate product explanation, a brand mention and a completed purchase are all potentially valuable. They answer different questions, however, and need different tests.

Choose a decision first. Perhaps you want to know whether turning campaign evidence into approved product answers is worth adding to every launch. Perhaps your team needs to establish whether an existing page leaves an important product question unanswered. Perhaps you want to understand how often sampled AI answers cite the brand's own material.

Each decision suggests an observable result. It also tells you what a disappointing result would mean. A test that can only confirm your enthusiasm is a press release wearing a lab coat.

For an early programme, separate three questions: can the team produce useful material reliably, does that material become available for discovery, and do the observations justify the next investment? You can make progress on the first without having settled the third.

Write down the claim before collecting the result

A useful test statement names the intervention, the audience question and the expected observation. For example: adding a reviewed explanation of the racing model's carbon plate may make the manufacturer's page a clearer source for questions about midsole design.

This statement is deliberately narrower than “improve AI visibility”. It identifies a specific information gap and a plausible reason to address it. The work still has value if a human reader finds the answer useful before an AI system cites it.

Record what you will actually measure. You might assess page completeness, the accuracy of sampled answers, citations to the relevant URL, or visits that can be attributed to a supported referral source. Do not quietly substitute an easier metric when the expected one refuses to cooperate.

A good hypothesis leaves room for the answer to be no. It should also distinguish between “no change observed” and “the test could not reliably detect a change”. Those conclusions have very different implications for your budget.

Establish a baseline you can reproduce

Before changing the content, save the current page and the relevant campaign evidence. Record the selected questions, their exact wording, the date, the provider or product surface used, and any settings available to you. Preserve the returned answers and cited URLs where the workflow permits.

Choose questions because they matter to customers. Include questions the existing content answers well, questions it answers poorly and questions outside the product's legitimate scope. Otherwise, you risk assembling a flattering sample that tells you very little.

A direct question about your brand behaves differently from a category recommendation. Keep those groups separate. So do factual product questions and broad requests for advice. A single average can conceal the very differences you need to understand.

Repeat the baseline where practical to get a sense of variation. If the answer changes substantially between compatible checks before you alter anything, a later change deserves especially careful interpretation. The system has already demonstrated its willingness to improvise.

Change something specific enough to inspect

Use AdCyte to analyse the relevant advertising, review the findings against their supporting moments and connect them to approved product facts. Then make a defined content change: an answer block, a comparison, a supporting page or a clearer description.

Record what changed and why. “Updated website” is not much use six weeks later. “Added an explanation of which version includes the carbon plate, supported by the current product record” gives the next reviewer something to work with.

Keep an intervention log that includes the page, publication date, approval owner and source version. Note changes elsewhere too, such as a new product launch, press coverage, distribution expansion or a redesign. You cannot hold the internet still, but you can stop pretending it did.

Run every advert through the analysis process. Decide the publication treatment after review; several assets may contribute to one strong answer rather than requiring several separate pages.

Choose a comparison that fits the decision

Test approachWhat it can help you learnMain limitation
Before-and-after checkWhether observations changed after a defined updateOther events may explain the difference
Repeated compatible checksHow stable an observation is over timeRepetition alone does not establish causation
Comparable untreated questions or pagesWhether a change also appears elsewhereGroups may differ in important ways
Staged rollout across suitable groupsWhether timing and treatment show a consistent relationshipRequires enough comparable material and careful design

A small team can begin with a documented before-and-after comparison. The language of the conclusion should match that design: “we observed” or “the result is consistent with”, followed by the limitations.

If a major investment depends on proving incremental value, involve someone who can design and assess an appropriate experiment. Bigger spreadsheets cannot rescue a comparison that never measured the right thing.

Keep the denominator visible

Decide how a returned answer qualifies for each measure. A brand mention can coexist with a competitor recommendation and a citation to another site. Preserve those distinctions rather than forcing every answer into one triumphant green box.

Here is a hypothetical reporting example. You attempt 30 checks. Three fail, leaving 27 eligible returned answers. The brand appears in nine, is recommended in four and has an owned URL cited in two. Those categories can overlap; adding them together would misrepresent the evidence.

Report the failures separately. State which questions and providers were included and avoid comparing the result with a previous period that used a different mix without explaining the difference. A sample that doubles in size and changes composition has changed the test as well as the numbers.

Distinguish an answer-level citation rate from a provider's separately defined Citation Share metric. Their names may sound related, but definitions and denominators determine whether a comparison makes sense. See the guide to AI visibility measures for the broader distinctions.

Read the answers behind the counts

An increase in mentions can hide worse descriptions. A citation can point to an outdated page. A recommendation can include the brand for a use the product does not support. Read representative answers and every material exception before celebrating a movement in the dashboard.

Use a simple review rubric: is the relevant product fact correct, is the source appropriate, is the answer useful for the question, and does it introduce an unsupported claim? Save enough context for another person to apply the same judgement.

Be precise about where the observation came from. A result obtained through a provider's API is evidence about that sampled response. It does not establish what every customer saw in the provider's consumer interface or in a different search product.

The discipline is less glamorous than announcing a percentage uplift. It is considerably more useful when someone asks you to explain it.

End with the next decision

Agree a review point before launch, allowing enough time for the content and the chosen measurement method to produce meaningful observations. There is no universal waiting period that guarantees a useful answer. Record practical dependencies and review them if access or publication is delayed.

At the review, separate operational learning from visibility observations and commercial evidence. Perhaps the workflow produced accurate answers quickly but the sample showed little stable change. Perhaps visibility improved while referral evidence remained too limited to interpret. Both findings can inform a sensible next step.

Choose whether to improve the source material, extend the test, repeat the workflow elsewhere or stop a low-value activity. Publish the limitations alongside the findings if you share the work. The interesting thought leadership is often the explanation of what you could not conclude, provided it leads to a better question.

Build your first test around a real campaign

Use AdCyte's VideoGEO workflow to turn campaign evidence into reviewed content, then set up a first visibility scan with a defined question set. If several teams need to agree the method, book a demo around one campaign and one concrete decision.

FAQs

Can a before-and-after comparison prove that VideoGEO caused a change?

Usually, it supports an observation rather than a causal conclusion. AI responses, competitor content and other brand activity can change during the same period. A stronger causal claim needs an appropriate comparison or experimental design, sufficient evidence and a clear account of other explanations.

How many questions should our first test include?

Use a manageable set that represents the decision you are testing. Cover important customer needs and preserve distinct groups, such as factual questions and category recommendations. There is no universal question count that makes a test reliable; sample design, variation and the size of the decision all matter.

Should failed checks count as answers without a brand mention?

No. A failed request did not produce an answer you could assess. Record it as a failure and explain how it affects coverage. Calculate answer-based measures from eligible returned answers using a documented rule, so an outage cannot quietly become a supposed decline in brand visibility.

It shows the result from that particular API check. Consumer interfaces and search features can use different retrieval, settings and presentation. Label the provider and surface accurately, and avoid extending a finding from one environment to all of its related products.

What if we see no improvement?

Check publication, source quality, question fit and measurement compatibility before drawing a conclusion. The content may still help visitors even if the sampled visibility does not change. Decide whether the result justifies improving the work, extending the observation period or choosing a more useful question.

Is this article reporting AdCyte test results?

No. It describes a practical testing method and uses a labelled hypothetical example to explain reporting. It does not present customer findings, a benchmark or an uplift claim. Actual results should be published with their source evidence, sample definition, dates, methods and limitations.

Where to go next

Continue reading

AdCyte newsletter

More from the work you already made.

Fresh perspectives on video advertising, useful content and AI discovery. Straight to your inbox.