Run a manual AEO audit before you buy an AI visibility tool

You can run the core of what a paid AEO tool does with a spreadsheet and one afternoon a month, and you should, because the audit tells you whether a tool is worth buying. Demandbase just made AI-answer visibility an account-level product. That is a useful signal about where B2B marketing is going. It is not proof that you need to buy it today.

On October 20, at its GO conference, Demandbase launched three agents: AEO, Campaign and Audience, reachable by API and MCP. The AEO Agent shows how major AI assistants describe a company to each member of a buying group, identifies the sources behind those answers, flags outdated information and content gaps, and identifies anonymous AI-driven traffic at the account level. No pricing was published.

Why start by hand

AI answers are unstable, and that shapes how to measure them. A SparkToro study, reported by Search Engine Journal, ran each of 12 brand-recommendation prompts 60 to 100 times per platform. ChatGPT and Google's AI returned the same brand list less than 1% of the time. Visibility percentage across many runs was more consistent than ranking position, and the researchers urged buyers to make AI tracking vendors show their math. The study was run with an AI tracking vendor and is not peer reviewed.

Google's own guide, covered in May, says optimizing for generative AI search is “still SEO” and that you do not need llms.txt, special markup or content chunking. So the first dollars should go to fixing what the audit finds, not to reformatting pages for AI.

A useful frame comes from a vendor benchmark of 100 small businesses by Modern Asset Lab. It scored them 80 on Found, 88 on Trusted and just 18 on Recommended. It is vendor-run and directional, but the three stages are a good way to organize your own audit: can assistants read you, verify you, and pick you.

The manual audit, step by step

1. Build the prompt set. Write 15 to 25 unbranded buyer questions, split by buying-group role: the economic buyer, the champion and the end user. Include problem questions (“how do I reduce churn in X”), category questions (“best tools for Y”) and comparison questions (“A vs B”). Never name your own brand in the prompt.
2. Run every prompt several times on each assistant. Because single answers vary, one run proves nothing. Five runs per prompt on each of three assistants is a pragmatic manual floor. That is our suggestion for a hand audit, not a threshold from the research. Use fresh sessions and logged-out or clean accounts where you can.
3. Log five fields per run. Were you mentioned, were you recommended, which competitors appeared, which sources were cited, and whether the claims about you were accurate and current.
4. Compute visibility, not position. For each prompt and assistant, calculate the share of runs where you were mentioned and the share where you were recommended. Ignore “rank,” given how often order changes.
5. Map the sources. Tally which domains the answers cite. These are the pages that shape your story, and many will not be yours: review sites, analyst pages, Reddit threads, competitor comparisons.
6. Sort the gaps into Found, Trusted and Recommended. Missing entirely is a Found problem. Stale or wrong facts are a Trusted problem. Mentioned but never chosen means you lack the specific, comparable claims (pricing, fit, proof) an assistant needs to pick you.
7. Write a ranked fix list and re-run in 30 days. Fix the factual errors first, then the missing claims, then the third-party sources. The re-run is your proof the fixes worked.

When the spend is warranted

Buy only when at least one of these is true, and you can say which in a sentence:

• Volume has outgrown your hours. Time one audit cycle. Multiply prompts by assistants by runs, then multiply the hours by your loaded hourly rate. If a monthly cycle costs more than the license, the tool pays for itself on labor alone.
• You need what a spreadsheet cannot do. Demandbase says its agent identifies anonymous AI-driven traffic by account and maps answers to each buying-group member using its contact data. You cannot reproduce account-level traffic identification by hand.
• The audit found gaps you are already acting on. If you have not fixed the top three, a dashboard will only show you the same problem faster.

If you do buy, ask each vendor how many runs per prompt they average, whether they query consumer interfaces or APIs (the SparkToro coverage flags API-versus-manual parity as an open question), and to show the math behind any score.

How to make the audit agentic, manually

An agentic task is not a smarter prompt. It is a loop with a trigger, a goal, state, tools, a boundary on what it may do alone, and a record. You can build every piece without buying anything.

1. Goal. State a measurable target, such as recommended-share on your category prompts rising from today's baseline over a quarter.
2. State. Put the prompt set and every run's log in a sheet or file the agent reads and writes. This is its memory, and last month's results are what it compares against.
3. Runbook. Turn the seven manual steps above into written instructions: what to run, what to record, how to compute visibility, what counts as a gap.
4. Trigger and tools. Schedule the run monthly in the workflow tool or agent runtime you already use, and give it the assistants or search access it needs. Spot-check API answers against what you see in the consumer apps before you trust them.
5. Boundary. Let it read, run, log and draft the gap report on its own. Require human approval before it changes any page, publishes anything or contacts anyone.
6. Record and escalation. Have it write a month-over-month diff and a ranked fix list, and escalate to you if visibility drops sharply or it cannot reach a source. Add a hard spend cap on the run.

Once that loop has run for a quarter, you will know your baseline, your cost per cycle and which gaps matter. That is the evidence that makes a purchase decision easy in either direction.

What to do about it

• Block one afternoon this month and run the seven-step audit on 20 prompts before you take any vendor demo.
• Report visibility share by assistant, never rank, and record the number of runs behind every figure.
• Fix factual errors and missing proof claims first. Skip llms.txt and page reformatting unless your own data says otherwise.
• Compute your break-even: audit hours times loaded rate against the license. Buy only if the tool wins on hours or does something a sheet cannot.
• Convert the audit into a scheduled loop with a log, an approval boundary and a spend cap, and review it after one quarter.