Arrow 
Track a fixed panel of questions tied to real buyer decisions, with non-branded discovery questions separate from brand identity checks. Record the origin and intent of each question, freeze its wording for a cycle and version any changes. A prompt panel samples chosen needs; it does not measure all market demand.
Start with a decision a buyer actually needs to make
Collect questions from consented sales notes, support requests, customer interviews, site searches and relevant public discussions. Preserve the original wording privately where necessary, then paraphrase away names and confidential details. For public discussions, retain a source link and capture date.
A Reddit question can reveal vocabulary or a concern, but its votes and comments do not establish search volume or buying frequency. Prioritize questions using fit with the product, specificity of the decision and the quality of the underlying evidence. Label invented or workshop-generated questions as hypotheses.
Keep discovery, evaluation and identity controls separate
A question that already names the company tests a different task from a question asking which solution to choose. Both can be useful, but combining them makes a brand easier to “find” by construction. Similarly, informational definitions should not dilute a commercial recommendation denominator.
Separate intended markets and languages from the outset. A French question aimed at a French buyer and an English question aimed at a US buyer are two contexts, even if translated from the same underlying need. Intended market is not proof of where the collecting device was located.
| Question group | Illustrative question | What it evaluates |
|---|---|---|
| Non-branded commercial | Which GEO tools fit a small B2B team that needs exportable answer evidence? | Selection against a stated buying constraint. |
| Non-branded informational | How can a marketing team tell a citation from a recommendation? | Ability to find and explain a concept or method. |
| Branded identity control | What does Arrow AI provide and which domain is its official website? | Description and identity accuracy, kept outside discovery rates. |
| Specific constraint | Which solution supports our required integration and data-handling limits? | Suitability for a real requirement; populate only with verified buyer constraints. |
Screen out prompts that preselect the desired winner
Use neutral wording that a buyer could plausibly use before knowing the answer. Avoid inserting your own brand into a non-branded question, adding unsupported praise or listing unusually specific features solely because your product has them. Real constraints are legitimate; engineered favoritism is not representative demand.
Give near-duplicates a shared intent label. Ten paraphrases of one use case do not provide the same breadth as ten distinct decisions. If you deliberately study wording sensitivity, keep that experiment separate from the stable baseline and report its purpose.
- Identify the buyer, decision and constraint in one sentence.
- Record whether the question is observed, paraphrased or proposed.
- Map it to one primary intent and a relevant existing source page.
- Check whether the evaluated company type matches the question.
- Review wording without looking at whether the latest answer favors the brand.
Freeze wording and version the panel as a dataset
Create a stable ID for each question and save its exact wording, intent, branded status, language, market and eligible subject type. Archive the panel file used for each cycle. A new question, changed wording or reassigned intent requires a new version; keep the old one available for interpreting its observations.
Do not overwrite the previous panel and compare percentages as if the sample stayed unchanged. Compare the shared, unchanged subset separately, with the reduced coverage disclosed. New questions begin their own baseline; retired questions remain part of the history that was actually collected.
| Change | Treatment | Reason |
|---|---|---|
| Fix internal notes without changing the question or eligibility | Document the metadata correction | Preserves traceability without inventing new observations. |
| Rewrite wording or change intent/market | Issue a new panel version | The measured task or context has changed. |
| Add a new buyer concern | Add it to a new version and establish its baseline | No earlier answer exists for the new question. |
| Change collection interface or session settings | Keep a distinct observation context | An identical question alone does not make results comparable. |
What the existing 70-question Arrow kit contains
The Arrow v2 panel proposes 60 non-branded GEO-solution questions: 40 commercial and 20 informational. It adds ten Arrow identity controls. English and French each have 30 non-branded questions and five identity controls. These are proposed questions, not verified demand or measured results.
Its B2B, yacht and DPE cohorts all evaluate GEO solutions. The yacht questions concern yachting businesses choosing GEO support; they do not ask consumers to choose a yacht broker. The older supplier panel studies different subjects and remains separate. Read the kit instructions before reusing either panel.
Illustrative planning calculation: 12 selected questions, two surfaces and three independent sessions create 72 planned attempts. The actual valid sample may be smaller. Selecting a relevant subset is acceptable if you freeze it, label its scope and do not call it the complete 70-question panel.
Review the panel between cycles, not after every answer
Assign a review cadence that matches how quickly your offer and buyer questions change. Add a newly discovered concern because it matters to buyers, not because testing showed an easy brand win. Keep a short retirement reason when an old question becomes irrelevant.
For collection, record known personalization conditions and session context. OpenAI’s memory documentation explains that personal context can affect responses and searches, so “new chat” alone is not a full environment description. Preserve unknown settings instead of inferring them.
Once the panel is frozen, follow the citation tracking method for evidence and classification. Improve the panel at the next version boundary; use the current cycle to learn what the selected questions actually reveal.
Sources and editorial scope
This is Arrow AI's implementation guidance. Examples are illustrative unless identified as dated observations. Source access and good content do not guarantee a recommendation.
Continue through the GEO evidence library, inspect Arrow GEO's measurement limits, or start an audit.