Can I do this without buying a tool?

Weeks one and three, yes. Week two is where manual work breaks down, because a baseline requires repeated sampling across engines and hand-querying does not produce a defensible number. Most teams do the first cycle manually, discover the sampling problem for themselves, and buy tooling for the second.

How many prompts should I start with?

Twenty to thirty for a first cycle. Enough to cover your main buying questions, few enough to sample repeatedly without excessive cost. Expand once you know which ones move.

What if I find nothing wrong technically?

Then you have eliminated the cheapest explanation, which is worth knowing, and your bottleneck is structural or authority-related. Move to passage-level structure on your highest-value pages.

Who needs to be involved?

One person can run it if they have access to server logs and the ability to ship page changes. In most organisations that means a marketer plus a few hours of engineering time in week three.

Running a 30-Day AEO Audit

What a 30-day audit is for

It is not for fixing AEO. Thirty days is not enough for that.

It is for producing two things: a baseline you can defend and one demonstrated improvement. Together those are what turn AEO from a topic someone read about into a funded programme with a number attached.

Structure the four weeks so that each one gates the next.

Week 1: access and instrumentation

The goal is to eliminate the boring explanations before spending money on sophisticated ones.

Days 1–2: crawler access.

Read your live robots.txt and confirm which AI crawlers are permitted. Then fetch your five highest-value pages with an AI user agent and read what comes back:

curl -A "GPTBot" -sL https://example.com/pricing -o pricing.html

Confirm your substantive content — prices, features, specifics — is in that HTML. If it is not, you have found your bottleneck in two days and the rest of the audit is engineering work. See Technical AEO for AI Crawlers.

Days 3–4: log analysis.

Pull thirty days of server or CDN logs filtered to AI user agents. Answer three questions: which crawlers visit, what status codes they receive, and which sections they never reach. A high share of 403s or challenge responses is your answer.

Days 5–7: build the prompt set.

This is the part teams rush and should not. Twenty to thirty questions your buyers actually ask, in their words. Sources worth mining:

Sales call recordings and discovery notes. The highest-value source by a wide margin.
Support tickets and pre-sales questions.
Search Console queries with question intent.
Competitor comparison queries: "X vs Y", "alternatives to X".
Category questions where you would want to appear: "best tool for job", "cheapest way to outcome".

Rules: no branded prompts about your own product name — those tell you nothing about discovery. Cover both insider and plain-language phrasing. Write them as a buyer would type them, not as you would describe your category.

Freeze this set. Changing prompts mid-audit destroys comparability, which is the most common way a first cycle produces unusable data.

Week 2: baseline

Now measure, properly.

Sample. Each prompt, across each engine your buyers use, repeated enough times that variance averages out. Repeated is the operative word. One observation per prompt is an anecdote; ten begins to be a measurement.

This is where doing it by hand collapses. Thirty prompts across four engines at ten samples each is 1,200 queries, and you need to repeat that next month for the comparison to mean anything. This is the specific problem AEO platforms exist to solve, and it is worth understanding why you need one before choosing one.

Record, per prompt and per engine:

Were you cited, with a link?
Were you mentioned without a link?
What position, if a recommendation list?
Which competitors appeared?
Which domains were cited instead of you?
Any factually wrong statement about you.

Compute the six metrics from The AEO Metrics That Matter: citation rate, mention rate, share of voice, average position, sentiment, and crawler serve rate.

Then look at the citation sources. This is the most useful single output of week two and the one most often skipped. When you are not cited, who is? Patterns are diagnostic:

Review aggregators and directories dominating means you have a third-party presence problem, not a content problem.
One competitor recurring means they have a page you need to beat.
Forum and community threads means the model prefers peer testimony to vendor claims.
Your own competitor's comparison page appearing on your brand queries means you have ceded that ground entirely.

Week 3: fix the bottleneck

One bottleneck. Not a list of twenty improvements.

By now the evidence points somewhere specific:

If crawler access is broken — fix that and nothing else. Unblock crawlers, allowlist them at the edge, server-render the pages that matter. This is the highest-return fix available and it needs engineering, so start it early in the week.

If you are retrieved but not quoted — the problem is passage-level. Take your five highest-value pages and apply the revision pass from Structuring Content So AI Can Quote It: answer-first openings, question-phrased headings, self-contained sections, specific checkable claims, visible dates, a real FAQ block.

If false claims are the pattern — this is the fastest win in AEO. Publish an unambiguous, well-structured page stating the correct facts, add Product/Offer and FAQPage schema, and get the correction into the third-party sources the models are citing. Corrections propagate faster than reputation.

If aggregators dominate — your work is off-site. Ensure your listings on the directories being cited are complete, current and accurate. This is unglamorous and effective.

If a single competitor recurs — read their page properly. Usually it is more specific than yours, better structured, and more recently updated. Beat it on those three axes rather than on length.

Ship the changes. Record what you shipped and when, because week four depends on it.

Week 4: re-measure and write up

Re-sample with the identical prompt set and method. Any deviation invalidates the comparison.

Expect modest movement. Four weeks is short, some changes need re-crawling, and some fixes will not surface for another cycle. Directional improvement on the specific thing you fixed is a good result. Do not over-claim: a citation rate moving from nine per cent to twelve per cent on a thirty-prompt set is encouraging, not proven.

Write it up in a form a budget-holder can act on:

Baseline: the six metrics, per engine.
The bottleneck you identified and the evidence for it.
What you shipped.
Post-change measurement, with honest error bars.
The projected cost and shape of running this continuously.
A recommendation on tooling, with the reason.

Point 6 lands better after a manual cycle than before one, because you will have concrete evidence of where the manual approach broke — almost always the sampling in week two, and the attribution in week three.

After the audit

The transition from audit to programme means three things become continuous rather than episodic: sampling on a fixed cadence, crawler serve rate under alert, and a remediation loop with owned capacity.

The third is where most programmes fail. Findings accumulate, nobody has time to act, and the dashboard becomes a monument. Teams that succeed either dedicate real capacity to remediation or use tooling that closes the loop — a prioritised fix queue, scheduled drafts, automatic page variants. Our comparison matrix marks which platforms do this and which stop at reporting; it is the single most consequential difference between them.