Before Calling It GEO, Decide What You Want to Improve

“Improve our AI visibility” sounds like a clear brief.

I would still need to ask what success looks like.

Do we want an article cited? Do we want the marketplace recommended? Do we want a hotel or tour described accurately? Or do we want someone to complete a booking after discovering us through an AI assistant?

Those outcomes require different work. Putting them under one label does not make them the same problem.

In my earlier writing about SEO, GEO, and AEO, I questioned how much time we spend naming the work. The practical question I want to explore here is what happens after a company agrees to invest in it.

Who needs to be involved, what should they change, and what evidence would tell us that it helped?

A citation and a recommendation are different outcomes

Imagine a hypothetical travel marketplace with a useful article about choosing a shore excursion.

An AI answer cites the article when explaining how to plan around a ship’s departure time. That is a useful visibility outcome: the page supplied information relevant to the question.

It does not necessarily mean the answer recommended booking through that marketplace.

Now imagine another answer names the marketplace as an option for finding private tours, but cites a third-party comparison.

The brand appears as a choice, while its own website receives no supporting citation.

I would record those observations separately. They tell us different things about how the business appears.

Olaf Kopp’s framework distinguishes improving citability from improving brand context, placing GEO across SEO, PR, and branding. I find that distinction useful for organizing responsibilities. I would treat it as a working framework rather than a universal definition of how every AI system operates.

When the objective is a useful citation

If the goal is to make a page a stronger source, I would begin with the page and the question it needs to answer.

Can the relevant system access it? Is the information accurate? Does the explanation provide evidence? Are important conditions clear enough to survive being quoted or summarized?

For Google’s AI Overviews and AI Mode, the published guidance says established SEO practices remain relevant. Supporting pages must be indexed and eligible for a search snippet; Google does not require special AI markup. Eligibility does not guarantee inclusion. Google’s AI features guidance.

That gives me a practical starting point without pretending that every platform works identically.

For our hypothetical shore-excursion article, I would examine whether the content distinguishes the ship’s departure time from its required boarding time. I would check whether transfer estimates explain their assumptions and whether the article clearly identifies which information travellers must confirm.

Those details make the answer more useful. Adding an “AI-friendly” paragraph would not compensate for getting them wrong.

Query fan-out changes the questions I would investigate

Google explains that its AI search features may issue related searches across subtopics and sources when developing a response. Google’s explanation of query fan-out.

For a travel marketplace, this encourages me to examine the smaller decisions inside a broad request.

“Find a suitable tour for my family” could involve walking distance, children’s ages, transport, timing, and cancellation conditions.

I would use those questions to identify information gaps. I would not automatically create a separate page for every possible variation.

Sometimes a clear section on an existing listing is the right solution. Sometimes the information belongs in a comparison or a policy page. Sometimes the business needs to verify a fact before publishing anything.

The content structure should follow the information need.

When the objective is a brand recommendation

If the goal is for the marketplace to be understood as a relevant option, rewriting one article may address only a small part of the problem.

I would examine the business description across its own pages and relevant external sources.

Does it clearly explain the operating model? Is it a marketplace, an operator, an agency, or some combination? Which audiences and services can it credibly claim to serve?

Then I would inspect the evidence supporting those descriptions.

A company can describe itself as suitable for complex group travel. That statement becomes more useful when there is concrete information about the service, its limitations, and relevant examples.

PR and partnerships can help make real expertise and useful evidence discoverable. They should not be reduced to distributing identical promotional claims across more domains.

I would also avoid promising that a campaign will “teach the model” a particular brand association. We can improve published information and observe outputs. We cannot assume control over a platform’s training or source-selection process.

A travel marketplace adds another objective: a valid booking

A brand can be recommended accurately while the selected offer is unsuitable.

The hotel may be relevant, but the available room does not accommodate the party. The tour may fit the itinerary, but the requested departure is unavailable.

At that point, the work involves product data, supplier information, and booking systems.

This is how I would assign the work:

These responsibilities overlap. The table is a way to identify who needs to contribute.

What my studies can actually measure

My own research has made these distinctions more concrete. I now want the observation and its measurement source written beside every visibility claim.

A reported search is a query exposed by the tool's search metadata. It is not a full transcript of hidden reasoning. In The Follow-Up Gap, 18 answers produced 25 reported searches, but only nine answers reported searches and citations. Search being available did not mean a new search was reported on every turn.

A visible citation is a displayed source link. Absorbed, Not Credited? recorded 53 credited sources across 18 answers. Four answers to the two general questions in ChatGPT and AI Mode displayed none. That does not identify an uncredited source or prove that the answer came entirely from model memory.

A brand mention records presence. It may be an example, a comparison, a warning or an incidental reference. My apparel and skincare studies show why I would classify the role instead of treating every appearance as a recommendation. The apparel matching could also identify brands through links; that belongs in the method.

A recommendation needs an explicit interpretation rule. I would record whether the brand or offer is presented as suitable for the request, the reason given, and the qualifications attached. A named hotel is not necessarily an eligible room offer, and an eligible offer is not a reservation.

A visit or conversion needs an appropriate analytics record. None of those answer studies measured sales. A displayed link does not establish a click, and an order is not attributable to an AI answer merely because the business also appeared in one.

An annotated answer: separate the observations

Consider this invented answer:

Harbour Example is one option for a private city tour. Its four-hour departure may fit your port visit, subject to confirming pickup and the boarding deadline. The itinerary page describes the route.

Suppose the interface displays a source link beside the final sentence. I would annotate the company name as a mention; “one option” as a qualified recommendation; the itinerary link as a citation; and the duration and suitability statements as claims requiring verification. I would record reported searches only if the tool exposed them, and a booking only if the booking system established one.

This text example is not an observed answer. Its purpose is to show why a single “AI visibility” total loses useful information.

A missing field also needs a precise label. “No reported search,” “no visible citation,” “brand not named” and “no analytics available” describe different states. They should not all become zero visibility.

Variation belongs in the measurement plan

In Same Question, Different Searches, eleven completed answers returned 33 reported query instances. The cited-site overlap varied by prompt. That is one reason I would repeat a test before interpreting a single new citation as an effect of a page edit.

My first report would preserve the prompt, platform, date, account conditions, completed or missing run status, answer text, source URLs, brand role and support for important claims. I would keep tool metadata separate from the displayed answer.

I would then report the observed counts and what remains unknown. The benefit is a more specific recommendation: correct an unsupported fact, supply a missing offer condition, improve a useful passage or investigate a recurring source gap. It is harder to justify a vague “GEO fix” when the measured problem is stated clearly.

The name becomes useful when it clarifies the brief

I would not reject “GEO” if it helps a company recognize the work and bring the right people together.

I would want the brief underneath it to be specific.

For example:

Improve the accuracy of how our marketplace is described for private-tour searches, and test whether our key planning resources are cited more consistently.

That gives the team two outcomes to investigate.

“Become the leading brand in AI” leaves too much undefined. It does not establish the questions, the audience, the platforms, the evidence, or the people responsible.

A new name can help open a conversation. The work still needs a clear scope.

Measure the outcomes separately

I would keep separate records for citations, brand mentions, recommendation context, factual accuracy, and measurable visits or bookings.

For repeated answer checks, I would preserve the prompt, platform, date, and relevant conditions. A result from a small sample should be reported as a result from that sample.

I would also inspect what the answer actually says. Being named in a warning or an unsuitable recommendation should not count as the same success as being accurately recommended for a relevant need.

The question I would bring to the next meeting is therefore quite practical:

Which outcome are we trying to improve, what is currently preventing it, and who can change that?

Once those answers are clear, SEO, GEO, PR, and branding become easier to organize around useful work.