How to build the question set your GEO measurement runs on

A GEO question set is the fixed list of real buyer questions, or prompts, that you put to AI assistants every time you measure your brand's visibility. Build it from search data, Google's "People also ask" boxes and your own sales and support conversations, and keep most of it unbranded. Your GEO results can only be as good as the questions behind them. A set full of questions nobody asks will produce a confident number that means nothing.

The question set is the part of GEO measurement most teams rush. It deserves a day of work, because every figure you report for the next year will rest on it.

In short

• Draw questions from three sources: search data, "People also ask", and what customers say to your sales and support teams.
• Do not cut questions for low search volume. In our analysis of 96 Google South Africa searches on 11 October 2026, all 45 searches with fewer than 100 monthly searches, or no measurable volume, showed an AI Overview.
• Keep most questions unbranded: they are where you compete for a place in the answer.
• Size the set so every figure you report rests on at least 30 answers.
• Fix the wording and version the set. Changing a question changes its answer.

Where do real buyer questions come from?

Search data

Start with what people already type into Google. Google's keyword volumes (from the Keyword Planner in Google Ads) show demand for South Africa, and Search Console shows the searches your own site already appears for.

Use volume to find themes, not to cut questions. In our 11 October 2026 analysis, an AI Overview appeared on all 31 searches with fewer than 100 monthly searches and on all 14 with no measurable volume. For the 17 searches with 1,000 or more a month, it appeared on 14 (82.4%). The busy searches without an Overview were tool, booking or place searches, such as "car finance calculator south africa" at 2,900 searches a month. Low-volume questions are exactly where AI answers appear, so a set built only from high-volume keywords misses much of the ground.

Then rewrite each search the way a buyer would put it to an assistant. "best car insurance south africa" becomes "Which car insurer is best value for a first-time driver in South Africa?" Keep a note of the original search alongside it.

"People also ask"

Google's "People also ask" box lists related questions under a search result. In our 11 October 2026 sample it appeared on 94 of the 96 South African searches (97.9%), more often than AI Overviews themselves (92 of 96). Open each of your core searches, expand the box, and copy the questions. They are already phrased as questions, so they need little rewriting, and each one you expand tends to reveal more.

Sales and support conversations

The questions that matter most often never show up in keyword tools. Read recent sales call notes, WhatsApp and email enquiries, chat logs and the objections your sales team hears every week. Look for four kinds:

• Comparisons: "How are you different from the bank I use now?"
• Cost: "What will this cost me a month?"
• Fit: "Do you cover drivers under 25?"
• Worries: "What happens if I miss a payment?"

Remove names, numbers and any other personal details before the questions go into a shared spreadsheet. Then turn each one into the unbranded question the buyer would have asked an assistant before they ever reached you.

How many questions should a GEO question set have?

Work backwards from what you want to report. Every figure needs at least 30 answers under it; how to measure GEO explains why. The table assumes three engines and one run of each question per engine.

You want to report	Answers needed per run	Questions that get you there
One figure per engine	30 per engine, 90 in total	30 questions
Three product lines, per engine	30 per line per engine, 270 in total	30 per product line, 90 in total
Branded and unbranded separately, per engine	30 of each per engine	At least 30 unbranded and 30 branded
A month-on-month change of about 8 points	About 300 per engine	300 questions; repeats of fewer questions count for less

For most brands, 50 to 100 questions is a practical start: large enough to report per engine, small enough that someone can read the answers when a number moves.

What mix of branded and unbranded questions?

Branded questions name you: "Is Kestrel Insurance good at paying claims?" Unbranded questions do not: "Which insurers pay claims quickly in South Africa?" Branded questions check how assistants describe you. Unbranded questions show whether you are in the running at all, which is why they should make up most of the set. About 80% unbranded and 20% branded is a sensible starting split. Report the two groups separately, as our guide to branded vs unbranded questions explains.

Mix the question shapes too. In our 11 October 2026 sample, "best", "cheapest" and similar list searches showed an AI Overview on 58 of 59 searches (98.3%), and how, what and when questions on 24 of 25 (96.0%). Cover at least these:

1. List questions: best, cheapest, top.
2. Comparison questions: one option against another, such as "medical aid vs hospital plan".
3. Cost questions: how much something costs per month or in total.
4. How-to and problem questions: how to claim, switch, apply or fix something.
5. Fit questions: the best option for a specific buyer, place or budget.
6. Branded questions: reputation, facts and comparisons that name you.

A worked example: Kestrel Insurance (fictional)

Kestrel Insurance is a fictional South African short-term insurer. The searches and volumes in the first column are real: they come from our 11 October 2026 analysis, and every one of them showed an AI Overview. The questions built from them are illustrative.

Source	Seed	Question in the set	Type
Search data	best car insurance south africa (1,600 a month)	Which car insurer is best value for a first-time driver in South Africa?	Unbranded, fit
Search data	cheapest car insurance south africa (1,600 a month)	What is the cheapest comprehensive cover for a 2019 hatchback in Johannesburg?	Unbranded, cost
Search data	how much is car insurance per month in south africa (110 a month)	How much should car insurance cost per month in Durban?	Unbranded, cost
Search data	best home insurance south africa (320 a month)	Which home insurers cover solar panels and inverters?	Unbranded, fit
Search data	is gap cover worth it south africa (no measurable volume)	Is gap cover worth it if I already have a hospital plan?	Unbranded, comparison
Sales calls	"Do you cover drivers under 25?"	Which insurers cover drivers under 25 without a very high excess?	Unbranded, fit
Support	"How long does a claim take?"	Which South African insurers pay car claims fastest?	Unbranded, comparison
Branded	None	Is Kestrel Insurance good at paying claims?	Branded

Kestrel settles on 80 questions: 64 unbranded, split across car, home and business cover (about 21 each), and 16 branded. Asked once on each of three engines, that is 240 answers a run, 192 of them unbranded. That gives 64 unbranded answers per engine, enough for a figure per engine. Each product line per engine gets only about 21, below the 30-answer floor. So Kestrel reports product lines with the engines combined (about 64 answers each), or asks each question twice.

How do you keep a question set useful?

• Fix the wording. Store the set with a version number and a date.
• Add, do not edit. New questions go in as a new group and are reported separately until they have a history.
• Retire questions with a note, and recalculate the trend without them so the comparison stays fair.
• Review it each quarter against new sales conversations and the "People also ask" boxes on your core searches.
• Do not write questions to flatter. A question only your brand could answer will always look good and tell you nothing.

Once the set is fixed, the next step is deciding what to count. GEO metrics for your marketing dashboard covers presence, first position, share of voice and citations.