GEO Handbook, Part 11: Choosing the right AI platforms and languages

A useful GEO strategy does not start with as many platform logos as possible. It starts with where customers prepare decisions. For a regional SME, a small consistently reviewed measurement scope can be more valuable than a large dashboard with unclear meaning.

1 — Separate platform, product and mode#

One provider can offer several interfaces and functions. An answer in Gemini is therefore not automatically a Google AI Mode result. API output is not an identical copy of a consumer interface either. Record the product, visible mode, model information where available and time. Changes belong in the measurement log.

This overview is not a quality ranking. Features, accounts and availability change. Documentation was checked on 25 October 2026. Your comparison also depends on what was actually visible in your test. Record an unavailable feature as a missing observation, not a negative mention of your business.

2 — Selection needs a business reason#

Start with existing evidence: customer conversations, voluntarily reported enquiry sources, attributable visits and relevant search situations. Each has limits. A referred visit shows a route to your website; a customer’s answer describes their recollection. Both are more useful than an unsupported assumption about the entire Swiss market.

3 — Languages have their own decision contexts#

DE-CH and EN should describe the same real offer, but customer questions need not be word-for-word translations. A Swiss German-language enquiry about a quotation, service area and acceptance may differ from an international property manager’s English question. Meaning, limitations and scope must still agree.

Google documents hreflang for language and regional versions. Related pages should consistently declare their alternatives, including themselves, with reciprocal references. This is a technical association, not proof that every AI system will select a preferred language version. [6]

4 — Calculate review effort before starting#

Planning example, not market data: 12 questions × 2 languages × 3 repetitions produce 72 observations per product and measurement wave. Two products mean 144 answers; four mean 288. These are answers to review, not independent people or search volume. Repetitions reveal variation and increase work; they do not automatically make a purposive sample representative.

5 — Maintain comparability and expose differences#

Use the same questions and scoring rules for comparisons over time. Start new sessions where your protocol requires them, and record sign-in state, location and relevant settings. A fresh chat does not automatically remove every form of personalisation. If conditions cannot be aligned, show separate results rather than one combined winner ranking.

Open source links and check whether they support the claim. A missing link remains “no visible source link”. It establishes neither that no retrieval occurred nor that the answer is correct. Unavailable answers must not silently disappear from the denominator.

6 — When to expand the scope#

Expand once the current wave has been fully reviewed and another platform or language addresses a specific business question. Version the revised protocol. Keep an unchanged core group for trends and report added cases separately. This preserves interpretability as platforms evolve.

Connections within the course#

Sources and working method#

Checked on 25 October 2026. Product descriptions use the linked provider sources. The selection matrix and workload scenario are original planning models, not market shares. No new comparative platform measurement was performed.

Glossary: terms and connections#