GEO measurement budgets: allocate runs without losing comparability
A GEO measurement budget can run out in several ways: no runs remain, spending reaches its limit, or the team cannot review more answers. Removing questions until the counter fits may solve the bill while leaving a series that no longer supports the same decision.
This guide helps plan the next measurement wave from an existing prompt bank. The output is an allocation sheet: what runs every time, what rotates, what capacity stays in reserve and which conclusions a reduction can no longer support. It does not prescribe vendor prices or a repetition count that works for everyone.
The prompt bank defines the questions and their version. Here the task is funding and scheduling their execution without presenting a reduced sample as coverage of the entire bank.
Define the decision that must survive a reduction
Write down the comparison the team needs at the end of the period. Identify its market, language, model, mode and time window. Tracking requires retaining the conditions needed to compare with the previous wave; exploration can follow a different schedule, but it must be labeled accordingly.
Separate essential requirements from desirable additions before seeing the answers. If the budget cannot support the essential objective, explicitly narrow the scope or defer the comparison. A design is not sufficient merely because it fits the available allowance.
NIST places the definition and prioritization of objectives before design selection. That principle informs this planning sheet; it does not make the GEO protocol a NIST-validated experiment or a representative estimate of the whole market.
Record who can approve a scope change. A financial constraint and a finding about visibility belong in different records: spending less does not establish that the brand has changed.
Budget in the units you actually consume
Distinguish logical responses, technical attempts and billable units. One response may require a retry; a service may deduct whole analyses rather than individual questions. Check your tool's current contract and retain its unit instead of turning every counter into an assumed per-prompt rate.
Within each configuration cell, planned logical workload can be expressed as:
Planned runs = distinct questions × repetitions per question per wave × waves in the period.
Add the cells without multiplying again by markets or models already separated into rows. For conversations, images or other workflows, first define what counts as a run; do not treat a complete conversation path as one call if it consumes several.
Track money, allowance and hours separately. Monetary cost includes applicable capture, evaluation, storage and human work where you have assigned a monetary value. Review hours remain a constraint even when they are not monetized. A plan within its spending and usage limits may still be impossible to review on time.
Keep estimates, authorized limits and actual consumption in separate columns. If a cost is not observable or you lack a current quote, mark it as pending rather than zero.
Protect a comparable core and schedule rotation
The core contains the questions and conditions that must recur in the waves being compared. Freeze their identifiers, versions and schedule before starting measurement. Rotation covers additional blocks that do not need to run in every wave.
A rotating row retains its planned date, inclusion reason and next review. Do not interpret its absence in another wave as a missing brand mention. That condition was not measured then. Nor should its last result be carried forward to fill the current dashboard.
Publish one view of the comparable core and another of each period's additional coverage. Combining them can make an aggregate change reflect questions entering or leaving the sample rather than different answers. Core selection limits what you can conclude about the rest of the bank.
There is no need to set a universal core-to-rotation percentage. First cost the set required for the agreed comparison, then allocate the remainder with a documented rationale. If an essential condition changes, mark a break or an overlap period rather than retaining a misleading continuity label.
Allocate capacity without confusing it with business weight
A question's weight in an indicator is not the number of times it should run. Weighting by demand and business value addresses the importance of each unit; capacity allocation must also consider observed uncertainty, the pending decision and the cost of obtaining useful evidence.
Use the AI answer variability protocol to justify repetitions within each cell. Do not copy a reference count as a precision guarantee. Without comparable history, reserve a calibration phase and label its conclusions as provisional.
Before expanding a cell, record which uncertainty the expansion addresses, which result would support a decision and what execution limit is approved. Do not keep adding repetitions until the brand appears or an inconvenient difference goes away.
When some cells receive more observations than others, an average across all responses gives those cells more influence. Retain cell-level summaries and the agreed aggregation rule. Increasing the sample must not silently alter the indicator's business priorities.
Keep reserves separate from committed resources
Set aside a reserve for contingencies defined before execution. Record what can activate it, who authorizes its use and which comparison must remain intact. Do not also count it as capacity available to expand the bank.
The NIST experimental planning checklist recommends checking run feasibility, allowing for unexpected events and retaining data and records. This proposal's reserve and change history apply those planning principles; the reserve's size needs justification from your actual conditions.
A technical retry retains its relationship to the previous attempt and the reason for retrying. It does not automatically create another independent observation. A valid answer with no mention is not a technical failure to replace for a better result.
Define a stop when the spending, allowance, time or execution-window limit is reached. Retain completed and pending work. If continuing requires more resources, return to the authorized decision; do not present a budget increase as an administrative change with no consequences.
Make sacrificed coverage visible
When the plan does not fit, record each reduction's effect before applying it. This matrix is a decision record, not a mandatory order for savings:
| Change | Resource it may release | Limitation to disclose |
|---|---|---|
| Defer a rotating block | That block's runs and review | No current observation of those questions |
| Reduce waves | Capture and review for the period | Lower temporal resolution and gaps between windows |
| Reduce repetitions | Planned attempts per condition | Less information about variability |
| Remove a market or model cell | Work for that configuration | No new conclusion about the removed cell |
| Reduce human review | Specialist time | Narrower scope of the stated quality checks |
Retain out-of-scope rows with their reason and date. Do not turn them into zero visibility or remove them from history to make coverage appear complete.
If a reduction breaks the core, explain which comparison is no longer valid. Closing a period as partial is preferable to combining incompatible results into one trend.
Close the allocation sheet before starting the wave
Each plan row needs a cell and version, block assignment, questions, repetitions, waves, logical workload, billable unit, expected cost and review time. Add an owner, window, execution priority and stopping rule. The reserve has its own row with no visibility result attached.
Check that totals fit the authorized spending, available allowance and team hours simultaneously. Review access and scheduling constraints as well. Do not attempt to recover a deferred wave by concentrating it on another date and calling it equivalent.
Retain the approved plan as a closed version. Add later changes with their date, reason and effect on comparability. This sheet can accompany your Mentio analyses; it does not announce an automatic budget optimizer, new allowances or a commercial plan different from the one contracted.
Reconcile consumption and prepare the next allocation
At closing, compare planned, executed and billed work. Separate valid responses, capture failures, retries and rows never started. Check evaluation work and pending review: completing capture is not the same as completing analysis.
Explain deviations through observable causes. If an invoice groups usage that cannot be assigned to individual cells, retain that limitation instead of allocating it with false precision. Unused allowance does not establish efficiency when the objective remains uncovered.
The closing record should retain the core's result, explored blocks, sacrificed coverage, consumed reserve and unresolved decisions. Use it to propose the next wave before seeing that wave's answers. Savings then remain tied to verifiable scope rather than a more convenient visibility percentage.
The capture failure and attempt history record makes that reconciliation explicit: a planned unit may have several attempts, but only the artifact accepted under the prior rule feeds each metric.
Frequently asked questions
How do I calculate runs for GEO measurement? For each configuration cell, multiply distinct questions by repetitions per wave and waves in the period. Add the cells and record retries and billable units separately; not every service deducts one unit per question.
What share of the budget should stay in reserve? There is no universal share. Justify the reserve using observed incidents, consumption uncertainty and the consequences of interruption. Define activation, ownership and limits before running, without committing the same capacity twice.
Does a higher-value prompt need more repetitions? Not necessarily. Business value informs priority, but repetitions also depend on uncertainty, the decision they must support and available resources. Indicator weight and observation count must be recorded separately.
Can I compare waves with different rotating blocks? You can compare the core that retains equivalent conditions and show rotating blocks separately. Do not attribute an aggregate change to AI behavior when it may result from a different question mix.
What if the budget cannot cover the full core? Stop the planned comparison and disclose the partial scope. Retain pending rows and decide whether to narrow the objective, add resources or schedule a new wave. Do not replace missing observations with zeros or old results.
Want to know if AI mentions your brand?
Discover your visibility in ChatGPT, Claude and Gemini in minutes.
Related articles
How to Build a Prompt Bank for Measuring Your Brand's AI Visibility
Build a representative, stable and actionable prompt bank to measure brand mentions, position and competitors across ChatGPT, Gemini and Perplexity.
GEO AnalyticsHow to Weight an AI Prompt Bank by Demand and Business Value
Assign transparent weights to an AI prompt bank using Search Console, internal search and business value, with a versioned denominator.
Practical GuidesHow to Measure AI Answer Variability Without Biasing Your Benchmark
Learn how to repeat prompts, control changes and separate signal from noise to build a reliable benchmark of brand visibility in AI answers.