You can't win prompts you haven't named

Every AEO programme stands or falls on the same thing — and most enterprise teams have never actually built it. The prompt library is The Brief. The 100–200 questions you've explicitly decided your brand belongs inside when buyers turn to AI for guidance. Without it, every downstream metric is hollow, every content investment is a guess, and every dashboard is a polished form of self-deception.

This is the work nobody wants to start with. It's slow, it's editorial, it requires judgement no tool can supply. And it's the single most defensible deliverable in the entire methodology, because every measurement that comes after — citation rate, share of voice, sentiment, source citations — is downstream of which prompts you picked.

Pick badly, and the dashboard tells you nothing. Pick well, and the next twelve months of work have a target.

Why this stage exists at all

AEO has no SERP. That's the bit most teams keep forgetting. In SEO, the keyword universe is fixed, the leaderboard is public, and you can argue about positioning because there's a position to argue about. In AEO, the response varies — by user, by session, by region, by platform. The same prompt asked twice in a row, frankly, can return two different answers. You're not optimising for a slot; you're optimising for inclusion in a generated answer. Which means the question “are we winning?” has no meaning at all until you've defined the prompts.

The Brief is two things at once. It's the list of conversations the brand belongs inside — the editorial decision about which generated answers you intend to be cited in. And it's the measurement frame — the fixed set of inputs you run repeatedly, month over month, to detect movement. Without it, you're producing content into a void and reporting on metrics that don't connect to anything specific.

Of course, you can run any prompt through any visibility tool. Otterly, Frase, HubSpot AEO, Peec — they'll all happily test whatever you feed them. But none of them will decide which prompts are worth feeding. That decision is the human work, and it stays the human work. It's also the work that genuinely cannot be outsourced to a generalist marketing manager — not because they're not smart, but because building a prompt library well requires sitting inside the buyer's vocabulary, not the seller's. Most internal teams, weirdly, are the worst-placed people in their own company to do this.

The Brief is what everything else gets pointed at. If it's right, the rest of the system has a target. If it's wrong, nothing in the system can save it.

The playbook

Five moves, in order. Total time for a serious enterprise build is 15–25 hours over a fortnight — though the better operators stretch it to four weeks to let the buyer-language work breathe.

Run a proper intake before generating a single prompt. The temptation — especially for in-house teams — is to skip straight to “what should we be cited for?” That's the seller's question. The buyer's question is “how would I phrase this if I were the person trying to solve the problem?” Those two questions, weirdly, rarely produce overlapping language.

A useful intake covers:

This sounds basic. It is basic. It's also where almost every prompt library we audit went sideways — they were built off a brand brief, not off buyer evidence.

Go for 150–250 raw prompts in the first pass. Wide first, sharp second. Five sources, in this order:

Now structure the raw list. Every prompt gets columns:

Cull to about 100–150 priority-1 prompts. Fewer than 50 and you don't have statistically meaningful data to work with — citation drift will swamp the signal. More than 300 and the operational overhead eats the engagement.

Distribution rule of thumb: 30% awareness, 40% consideration, 25% decision, 5% post-purchase. Adjust by category — enterprise B2B skews heavier to decision-stage; consumer ecommerce skews heavier to awareness.

This is the manual labour step. There's no shortcut.

Run every priority-1 prompt through at least three platforms — ChatGPT, Perplexity, and Google AI Overviews are non-negotiable. Add Claude if the audience skews technical; add Gemini if the buyer lives in Google Workspace. Use incognito browsing, log the time of day, and take the run seriously — this is the “before” snapshot that justifies the next twelve months of work.

For each prompt × platform combination, capture:

Two-pass it to stay sane: first pass is just Y/N — gives you the “are we on the map?” answer in about 90 minutes for 150 prompts. Second pass is the detailed capture for the prompts that matter most.

Cross-check with the free tools — HubSpot AEO Grader, Ahrefs Brand Radar, Semrush's AI Search Visibility — and screenshot the outputs. They become the visuals in the baseline report and they triangulate the manual numbers, which matters because every tool has its own blind spots.

Last move, and the one most teams skip: explicitly agree, in writing, what success looks like at 90 days, 6 months, and 12 months. Which prompt clusters matter most, ranked. What citation rate target is realistic — enterprise baselines typically come in at 5–15% on priority-1 prompts, and a credible 12-month target is 25–40%. Which engines matter most for this audience, and which ones are being explicitly deprioritised. How citation drift will be handled in monthly reporting — rolling 4-week averages, not raw monthly numbers.

This step is where most engagements quietly fail later. Without written targets, every monthly report becomes a Rorschach test for whatever the loudest stakeholder feels that week — and a programme that's actually working can get cancelled in month four because the wrong stakeholder read a noisy chart on the wrong morning.

What to measure

Five signals come straight off The Brief. None of them mean anything individually — the dashboard is the unit, not any one metric.

Report on rolling averages, never raw monthly numbers. Citation drift — the natural variance in which sources AI systems pull for the same prompt — produces 15–30% month-to-month swings that look like catastrophes in a line graph and like noise in a quarterly view. The team that doesn't understand drift will read every dip as failure. The team that does will absorb it without flinching.

Common failure modes

Five patterns, almost all of them avoidable, almost all of them present in the first audit we run for a new client.