GEO is not a secret door around bad SEO.

When someone proposes a GEO service, ask which pages they would change, what evidence supports the changes and how they would measure them. The name of the service should not substitute for an implementation plan or a clear account of what cannot be guaranteed.

Ask what the recommendation would help a buyer do

Before commissioning another visibility exercise, choose one buying question. Could a correct answer help someone shortlist the business, compare an implementation or understand the scope? If the question has no useful relationship to your offer, appearing in its answer is a weak reason to spend. I would review the answer for accuracy and fit before discussing a visibility score. A favourable mention that describes a service you do not offer can create the wrong enquiries. Correct the underlying page and record the mismatch; do not present it as a win.

The useful version of GEO

The useful distinction is between being easy to explain and being repeatedly recommended. You can improve the first directly: clear eligibility, accurate descriptions and answers to buying questions. The second depends on systems you do not control. A credible GEO programme names that boundary. Its strongest deliverable is often information the business needed anyway, while its weakest deliverable is a score whose definition changes whenever the story gets uncomfortable.

Google’s AI features guidance says there are no additional technical requirements or special optimisations needed for AI Overviews or AI Mode beyond the applicable Search foundations. That does not guarantee inclusion.

Ask what you actually want to be known for

“Appear in ChatGPT” is too vague to manage. A better objective is to be discoverable when an operations leader compares inventory platforms, or when a founder asks how to measure qualified demo acquisition.

Write down the problems you solve, the customers you help and the situations in which your approach makes sense. Turn those into a small set of buyer questions. Include comparisons, implementation concerns and reasons someone might reject your offer. That produces a useful editorial plan even if no AI platform ever cites it.

Separate three kinds of work

This distinction prevents a report full of mentions from being sold as evidence of revenue. It also stops a missing citation from becoming an excuse to rewrite a good page every week.

Build a useful first experiment

Choose one topic where you have real depth. Create a page that states the answer early, explains the conditions and shows how to make a decision. Include a worked example or checklist that readers can apply. Connect it to the relevant service and supporting articles.

Before publishing, record a baseline for a fixed question set. Note the platform, date, question and answer. After publication, repeat the observation consistently. Track referrals and enquiries separately. If you change your questions halfway through, you have changed the test.

My working hypothesis is that clearer, better-supported content is easier for both readers and answer systems to interpret. That is an editorial principle, not a guaranteed citation formula. Validate it through observation instead of dressing it up as a proprietary ranking hack.

Before reporting a visibility lift, establish the AI measurement baseline. For the page itself, the answer-and-evidence guide shows how to turn a vague response into something a buyer can use.

What I would refuse to prioritise

I would not start by adding mysterious files, repeating the brand name dozens of times or buying packages promising a permanent AI ranking. I would first fix unclear descriptions, contradictory claims and missing information about how the work is delivered.

A GEO request I would turn down

Imagine a founder asking for fifty “AI-friendly” articles while the pricing page hides the offer, the product pages use vague language and public descriptions disagree about who the service is for. In that example, I would fix the product explanation before selling a publishing programme.

Start with five questions a prospect asks before contacting you. Write a direct answer using information the business can substantiate. Check that it agrees with the product, sales materials and public company profiles. Only then consider new articles that fill genuine gaps.

The work should remain useful without a citation lift. Once the foundations are strong, test a specific buyer-question gap.

A current reporting update worth using

Bing introduced additional AI visibility views in June 2026, including intent, topic and comparative citation analysis in preview. That offers a better investigation starting point than treating every mention as equally valuable.

My recommendation is to prioritise a repeated gap around a relevant buying question. If the brand appears mainly for general definitions, another definition article may reinforce the wrong kind of visibility. A clearer comparison or eligibility answer may be more useful. The reporting identifies a question to investigate; it does not prove which edit will earn a citation.

Test the claim behind the proposal

Take one proposed GEO deliverable and ask for its mechanism. Which buyer question will it answer? Which page will change? What observation would suggest it helped? A recommendation that cannot survive those questions is not ready for implementation.

Then separate a documented platform requirement from a consultant’s hypothesis. The former needs an appropriate source; the latter needs a bounded test. “Recommended by our framework” does not settle either category.

For the first test, record the existing answer, publish one useful improvement and repeat the same observation method. If the answer changes, inspect its accuracy and destination before celebrating. If it does not change, you still need a useful page to justify the work. That is the commercial discipline a visibility programme should have.

Evaluate an AI visibility recommendation before implementing it

Ask whether the recommendation improves access, information, user experience or measurement. Then ask what evidence supports the claim that it will affect a particular platform. Those are different tests. Rewriting an unclear service description can pass the first test without a controlled experiment proving the second.

When someone presents a study, inspect the sample, date and method. A correlation between established brands and visibility does not prove that buying mentions will reproduce the result. Larger brands may differ in many ways at once. A sensible strategy can use the finding to form a hypothesis while refusing to convert it into a guaranteed tactic.

Ahrefs’ study of AI brand-visibility correlations found relationships between brand signals and observed visibility, and explicitly cautioned that correlation is not causation. The useful lesson is to examine the quality and consistency of your broader footprint. It is not to purchase a prescribed number of links, videos or mentions and expect an automatic recommendation.

A practical first-month work sequence

During the first phase, collect the current information about your business. Compare the homepage, services, author pages, case descriptions and important external profiles. Do they agree on the audience, offer and markets? Record contradictions before editing. A business that calls itself an agency on one page and an independent consultant on another has a clarity problem that can confuse people too.

Next, assemble a buyer-question set. Include questions about fit, alternatives, implementation, cost and risk. Choose enough questions to represent meaningful decisions, not enough to make a dashboard look impressive. Keep informational and buying questions in separate groups so a topic explanation is not scored as though it should naturally list suppliers.

Then select the information gaps you can responsibly fill. Improve one offer page and one substantial supporting resource. Add relevant internal links and inspect the delivered content. Re-run the same observations and note what changed. Treat the process as a monitored information improvement, not as proof that your edit caused every different answer.

Use a buying checklist for outside help

Use this checklist in a proposal meeting. It does not require you to become an answer-engine researcher. It requires the provider to explain the work in terms you can review and the team can continue.

A citation can be the wrong win

Imagine an AI answer recommending an enterprise-only service to a small company because your page says “for businesses of every size.” A mention score may improve; the actual buyer journey has become less accurate.

The first repair is eligibility. Name the company conditions, problem and scope that make the offer suitable. Then check whether observed answers preserve that boundary. A page with fewer mentions and a more accurate description can be the better commercial asset.

Change the priority when the information is already consistent and a relevant comparison repeatedly overlooks a genuine capability. That creates a specific content hypothesis. The evidence should decide which information needs work, rather than the desire to show a rising score in the next presentation.

Sources and further reading

Google Search Central: AI features and your website: Google AI eligibility and reporting boundaries.

Microsoft Bing: new AI visibility insights, June 2026: Preview additions to citation analysis.

References checked 9 October 2026. Sources support the documented facts and tool behaviour; the recommendations and worked examples explain how I would apply them.