An AEO prompt battery is a fixed set of buyer-relevant questions you run against every AI engine on a repeating cadence, giving you a defensible trend line for share of voice, citation rate, and content performance. Designed well, the battery is the foundation of every serious AI answer engine optimization program. Designed poorly, it produces numbers that move for reasons nobody can explain.
Why the battery is the whole game
Everything downstream in AEO depends on the prompt battery. Your share of voice metric is share of voice across the battery. Your attribution model joins mention windows against the battery. Your content briefs come from prompts in the battery where you are losing. If the battery is wrong, everything else is precisely wrong.
This is the same lesson SEO teams learned about keyword lists a decade ago. The keywords you track define the market you think you are in. The prompts you track define the market you actually compete in. Get the battery right, and the rest of the program largely runs itself.
The four prompt types every battery needs
A good battery covers four question types, weighted by how much pipeline each type actually produces.
- Category prompts (about 40 percent of the battery). "What is the best X for Y." "Who are the leading vendors in X." These are the highest-intent, top-of-funnel questions. If you are not present here, you never entered the consideration set.
- Comparison prompts (about 25 percent). "X vs Y." "Alternatives to X." These fire during shortlisting. Weakness here means you get to the demo and lose to the incumbent.
- Feature or capability prompts (about 20 percent). "Which X does Y." "Best X for Z workflow." These fire during evaluation. They convert best if you actually do the thing.
- Job-to-be-done prompts (about 15 percent). "How do I do Y." "What is the right way to Z." Mixed intent. Sometimes surfaces vendors, sometimes not. When it does, intent is high.
The weights are not sacred, but the discipline of writing prompts against a target distribution keeps the battery honest. Without it, teams overweight brand-name prompts because they look better in reports.
How to source the actual prompts
The best prompts do not come from your marketing team. They come from three places.
- Sales call transcripts. What did the prospect say they searched for. What did they ask ChatGPT before the call. Every discovery call is a prompt source.
- Support and CS tickets. The questions your customers ask post-sale are often the same questions the next cohort of buyers is asking pre-sale.
- Search Console and paid search reports. The queries buyers still type into Google are the queries they are also typing into ChatGPT and Perplexity, in slightly longer form.
Gather 200 raw questions from those three sources, cluster them, and pick the 60 to 120 that best cover the four prompt types. That is your battery. Everything else is variance.
Freeze the battery
The single most common failure mode we see is teams tweaking prompts week by week. It feels productive. It destroys the trend line. If you change even ten percent of your prompts, week-over-week share of voice comparisons become meaningless, and you cannot tell whether the change came from your content, the model, or your edits.
The rule is: freeze the battery for a full quarter. Version it. When you add or remove prompts, publish a new version and report both the old and new series side by side for at least four weeks so downstream consumers can see the delta.
Sampling strategy
A single sample of an LLM answer is one draw from a distribution. Run the same prompt five times and you will get five different answers, some of them structurally different. The fix is disciplined multi-sampling.
For most batteries, five samples per prompt per engine per week is the minimum for a stable signal. Ten is better when the category is competitive. Twenty is expensive and rarely justified. Whatever number you pick, keep it constant across weeks. If you sample five one week and ten the next, your confidence intervals move for reasons that have nothing to do with your content.
Engine coverage
Cover at least four engines, ideally five: ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews. Each draws from a different source graph, applies different guardrails, and reaches a different slice of your buyers. Measuring only ChatGPT gives you a large, biased sample. Measuring only Perplexity gives you a smaller, cleaner one. Measuring both gives you triangulation. Our share of voice guide walks through each engine's behavior in detail.
Instrumentation checklist
- Prompt IDs. Every prompt has a stable identifier that survives text edits within a version.
- Sample metadata. Timestamp, engine, model version, temperature, sample number, full raw output.
- Mention parsing. Consistent brand and competitor name normalization. Handles common misspellings and abbreviations.
- Prominence tagging. Primary, secondary, or passing (see our share of voice guide).
- Citation capture. Every URL cited by the model gets stored. This becomes the source graph you optimize against.
- Sentiment tagging. At minimum, neutral or better vs negative. Anything more granular is nice-to-have.
Common mistakes
- Only brand-name prompts. Feels good, does not correlate with pipeline. Category prompts are where the deals live.
- Prompts written by one person. A battery designed by a single PMM misses whole classes of buyer questions. Get sales and CS in the room.
- Silent edits. Editing a prompt without versioning invalidates every trend line downstream.
- Undersampling. Two or three runs per prompt per week are not enough to distinguish signal from noise.
- Ignoring the citation layer. If you do not know which URLs the model cites, you cannot influence what it says.
A 30-day battery build plan
- Week 1. Gather 200 raw prompts from sales calls, support tickets, and search queries. Cluster into the four prompt types.
- Week 2. Draft the first 100 prompts to hit the 40/25/20/15 distribution. Review with sales and PMM.
- Week 3. Stand up multi-sample runs on at least four engines. Baseline the numbers.
- Week 4. Freeze v1.0. Ship the first weekly report to the RevOps and marketing leads.
How the battery feeds the rest of the program
A frozen battery is the foundation for the two loops that turn measurement into revenue. The attribution loop joins mention windows from the battery to CRM first-touches. The action loop turns losing prompts into content briefs and turns competitor-cited URLs into a syndication backlog. Both loops fail without a stable battery underneath.
The order matters. Build the battery first, run it for four weeks to prove stability, then wire attribution, then wire action. Skipping ahead almost always produces a program that has to be rebuilt from scratch within a quarter.
Where to go from here
A prompt battery is not something you can copy from another company. It has to reflect your category, your buyer, and your positioning. If you want a starting point to see what the pattern looks like in your market, our free AEO Grader runs a six-prompt mini-battery across three engines. It is not enough for a real program, but it is enough to see the shape.