Not every AI prompt is worth fighting for: how I choose the buyer questions to win
When a team hands me a prompt list, I do not start writing. I sort. Most wasted GEO effort goes into prompts the brand was never going to own, while the few it could win get one tired page. The sort takes an afternoon. Skipping it costs a quarter.
The short answer
Sort buyer prompts by shape before anyone writes a page. Recommendation and comparison prompts are where a specialist brand can actually be named. Definition prompts (“what is…”, “apa itu…”) mostly go to universities and associations. Generic service prompts (“jasa X Jakarta”) reward whoever publishes the most. In Indonesia, track Indonesian prompts first and give English a smaller, deliberate share.
Not every question is a battle you can win
I have sat in rooms where the list ran to eighty prompts and the brief was “own all of them.” That is how you get fifteen thin definition pages and zero recommendation pages that deserve a place on a shortlist. Shape tells you where effort can pay. It does not promise what an engine will do.
Wording changes whether an AI Overview appears at all. In an audit of 55,393 trending Google searches, AI Overviews appeared for 64.7% of question-form searches against 9.5% of other searches. Among non-questions, queries of six or more words triggered them 38.7% of the time; one-word queries, 9.9% (WashU, ACM IMC 2026). And even when an answer appears, it may never name a supplier. Discoverable’s UK study found some AI answers explained a process or asked a follow-up instead of giving a shortlist (Discoverable, Sep 2026). If the answer never names anyone, there is no list to be on. I park those prompts.
The four shapes I sort by
This is the table I put on the wall before any page brief leaves the room.
| Shape | Example | Who usually gets named or cited | Can a specialist win it? | What I do |
|---|---|---|---|---|
| Recommendation | ”rekomendasi software HR untuk perusahaan menengah”, “best payroll software Indonesia” | A named shortlist; listing platforms often among the sources | Yes, the most winnable | Build the page that deserves a place on the list, keep listings accurate, earn one independent mention |
| Comparison | ”GEO vs SEO apa bedanya” | Explainers with clear tables | Often | One honest comparison page with criteria |
| Definition or how-to | ”apa itu employer branding”, “cara mengukur visibilitas AI” | Universities and associations | Rarely as the named answer, sometimes as a cited source | One strong page; judge it by citation, not by being named |
| Generic service | ”jasa GEO Jakarta” | Agencies that publish at volume | Hard without volume | One honest buyer checklist; do not build the programme around it |
Branded prompts (“what is White Wood”) belong in a separate accuracy check. They test what an engine remembers, not whether it recommends you.
The pattern in our own monitoring matches the table. Recommendation prompts are where we have seen specialist brands named at or near the top of the answer. Definition prompts in Indonesian pull university and association pages. Service prompts stay with agencies that publish at volume. That is not a failure of the definition page. It is a mismatch between the question and the prize.
Language is part of the shape
An Indonesian prompt and its English translation are two different prompts. They can return different brands and cite different sources. So I treat Indonesian as the main meter and give English a deliberate minority of slots, for the regional buyers who really ask in English. I do not translate the English list and assume the answers match.
Most Gemini use in Indonesia is Indonesian and mobile, so treating an English list as a mirror of the Indonesian one is a metering error. 84% of Gemini prompts here use the local language and 82% come from mobile (Google internal data, Jan to Mar 2026, The Gemini Report). AI Mode has been available in Indonesia, in Bahasa Indonesia, since 9 September 2025 (Google Indonesia), and it splits one question into subtopics it searches at the same time (Google Search Help). ChatGPT still dominates chatbot share here: 84.91% in September 2026, then Perplexity 8.08% and Gemini 6.13% (Statcounter). Google’s AI Overviews and AI Mode sit inside Search and need their own columns. One blended score hides which surface is empty.
This is also why new White Wood journal posts ship as English and Indonesian twins. A twin carries the same facts, written for readers of each language.
How I sort a prompt list
- Write prompts in buyer words, in the language they would type or say into a phone. Short and informal is normal.
- Tag each prompt by shape and by language.
- Run each one once by hand. If the answer never names a supplier, park it.
- Pick a small set to win (recommendation and comparison) and a small set to be cited for (definitions).
- Freeze the wording. In the monitor we use, editing a tracked prompt resets its history. Add a new prompt instead.
- Re-measure per engine 7 to 14 days after the matching pages go live.
What this is not
This is not abandoning definitions. A cited definition page still builds authority. It is not gaming lists: a page that ranks your own brand first is more likely to help a competitor (see cited is not recommended). And it is not a promise of position. Shape is a filter for effort, not a contract with the engine.
FAQ
How many prompts should we track?
Our measurement guide suggests a frozen set of 20 to 40. The mix of shapes matters more than the count.
Should we track English prompts at all?
Yes, if regional or international buyers ask in English. Keep them as a deliberate minority and report them separately from Indonesian.
Buyers do search “jasa X Jakarta”. Why not chase it?
Because the answer usually goes to whoever publishes most. Keep one honest page for it and put the real effort where a specialist can be named.
Why is my brand never named for “apa itu” prompts?
For definitions, engines tend to cite neutral explainers, often from universities and associations. Aim to be cited there, and win the recommendation prompts instead.
Does phrasing really matter that much?
Yes. In the WashU audit, question-form searches triggered AI Overviews about seven times as often as other searches (64.7% against 9.5%).
Three things I ask for before any page is written
- The prompt list, tagged by shape and language.
- The five to ten prompts we intend to win, with one line each on why we can.
- A per-engine baseline for those prompts, so the re-measure means something.