Can you pilot AEO before signing a retainer?
You can pilot AEO, and the pilot should be judged on the supplier rather than on the citations. Thirty days is enough to prove whether the measurement is real, whether the page work is any good, and whether anybody is making decisions from the data. It is not long enough to prove a citation outcome, for two reasons. New pages start at a disadvantage: Ahrefs analysed about 17 million citations and found the pages AI assistants cite averaging 1,064 days old, against 1,432 days for Google's organic results. ChatGPT leaned newest of the platforms tested, and even there its citations ran about 458 days younger than the organic SERP rather than reaching for last week's post. And the measurement itself is noisy: Ronald Sielinski's 2026 paper "Quantifying Uncertainty in AI Visibility" found that many apparent differences between domains sit inside the noise floor of the sampling. So buy a pilot to test the supplier, and write the retainer so you can leave.
Can you pilot AEO before signing a retainer?
You can pilot AEO, and a good pilot answers a different question from the one most buyers think they are buying. What a short engagement can establish is whether this supplier measures honestly, writes pages that are better than yours, and reaches decisions you can follow. What it cannot establish is how much your citation share will move, because that takes longer than a pilot runs.
Most suppliers will agree to one. An audit plus a first month of page work is a normal way to start, and a supplier who refuses any scoped first phase is telling you something about how they sell. The reverse warning matters as much: a pilot priced as a bargain and delivered as a slide deck costs you a quarter.
Write the success criteria before the work starts, in terms the pilot can actually satisfy. Useful criteria look like a frozen question set with raw answers attached, a named list of pages changed with the before and after visible, and a short document saying what was learned and what would be done next and why. Those are all verifiable in 30 days.
Unverifiable criteria are the trap. A pilot that promises to raise your citation rate by a stated percentage inside a month is promising something no supplier controls, and the way that promise usually gets met is by changing the question set or counting mentions instead of citations. Our guide to telling whether your AEO is working sets out which numbers can be checked and which cannot.
The alternative to a pilot is often better: a month-to-month retainer with a written exit, which gives you the same option to leave without the overhead of scoping a separate project.
How long does an AEO pilot have to run to prove anything?
A pilot needs about eight weeks to say anything about movement and 30 days to say plenty about the supplier. Those are two separate tests, and conflating them is why so many pilots end in an argument.
New pages start behind. Ahrefs' 2026 freshness study of about 17 million citations found the pages AI assistants cite averaging 1,064 days old, against 1,432 days for Google's organic results, so a cited page is typically closer to three years old than to three months. ChatGPT showed the strongest lean towards new pages of any platform in that study, at about 458 days younger than the organic SERP, and Google AI Overviews went the other way, citing pages around 16 days older than organic. A page published in week one of a pilot is competing against pages that have had years to accumulate the signals engines read.
Indexing adds its own delay before any of that applies. A new page has to be crawled and indexed by the engine behind each assistant before it can appear in an answer, and our page on how long until ChatGPT indexes a new page covers what the published tests show. Rewrites of pages the engines already fetch move faster than new pages, which is a good reason to spend a pilot on rewriting rather than on publishing.
The practical shape that works: weeks one and two for two baseline runs a week apart, so you can see how far two identical runs differ before anything changes, weeks three and four for fixes to pages that already get retrieved, weeks five to eight for repeated measurement, and a decision at the end about scope rather than about whether the channel exists. Judge the first month on process, and read our page on how long AEO takes to work before setting an expectation with your own leadership.
What should an AEO pilot include?
An AEO pilot should include a dated baseline, a technical access check, a small number of page rewrites, and a written decision log. Four items, each of which leaves evidence behind that stays useful even if you do not continue.
The baseline comes first and is the part most often done badly. It needs the exact prompts, written down and frozen, the engines they were run on, the date, and the raw answers saved rather than summarised. Run it twice, a week apart, before anything is changed: the second run is what tells you how much your own numbers move on their own. Without the raw answers you cannot check anything later, and a question set that changes between runs makes the second measurement incomparable with the first.
The access check is next, and it is cheap. Can the AI crawlers reach the pages, does the body text arrive in the HTML without JavaScript, are there redirect chains and dead URLs eating the crawl. A pilot that skips this can spend its whole budget on copy that no crawler will ever read, and our page on whether you need a developer for AEO covers which of those fixes need engineering time.
Then the page work, aimed at pages the engines already retrieve. AirOps' 2026 study of retrieval, fan-out and Google SERPs in ChatGPT citations found that 85% of the 548,534 pages ChatGPT retrieved across 15,000 prompts were never cited, so on most sites there is a queue of pages the engine reads and declines to quote. Those are the fastest wins available inside a pilot window.
The decision log is the last item and the one that tells you most about the supplier. Every change should have a reason, and the document should say what is being tested next and what result would change the plan.
Which pilot shape fits which situation?
Four pilot shapes come up repeatedly, and they prove different things at different costs. Pick by what you are uncertain about, not by what is cheapest.
| Pilot shape | What it proves | What it cannot tell you | Best when |
|---|---|---|---|
| Audit only | Where you stand today and which pages are blocked, retrieved or quoted | Whether this supplier can fix anything | You need a business case before any budget exists |
| Measurement only | Whether the reporting is real, repeatable and comparable run to run | Whether page work moves your position | You suspect your current supplier's numbers are a dashboard |
| Audit plus fixes to existing pages | The supplier's page work, against pages the engines already fetch | How the programme performs once new coverage is needed | You have a content library and no idea why it is not cited |
| Month-to-month with an exit | The real engagement, including how they handle a flat month | Nothing, which is why it is usually the better buy | You would otherwise spend a month scoping a pilot |
The last row is the one most buyers overlook. A pilot is a project with its own scoping, reporting and handover, and all of that is overhead you pay for twice if you continue. A monthly engagement with a written notice period and stated exit criteria gives you the same ability to stop after one month, with the first month spent on the work rather than on the framing.
An audit-only pilot earns its place in one specific case: you need something to show a finance or leadership gate before any ongoing spend can be approved. Our page on what an AI visibility audit includes describes what that deliverable should contain.
What can a pilot not prove?
A pilot cannot prove that a change in your citations was caused by the work, and pretending otherwise is the most common way these engagements end badly. The measurement is noisy enough that a single before-and-after comparison is not evidence.
Ronald Sielinski's 2026 paper "Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement" makes the statistical case: visibility figures are sample estimates of a response distribution rather than fixed values. Sampling Perplexity, OpenAI SearchGPT and Google Gemini across three consumer product topics, daily over nine days and again at ten minute intervals, he found citation shares varying substantially between repeated samples, with bootstrap intervals putting many apparent differences between domains inside the noise floor of the measurement itself.
What follows for a pilot is practical. One run before and one run after tells you almost nothing, so the question set has to be run repeatedly and the comparison made across runs. A supplier who reports a single-run improvement as a result either does not understand the variance or is counting on you not to ask, and our page on how many prompts you need to track covers how large and how repeated a set has to be.
Attribution is the second thing a pilot cannot settle. Competitors publish during your pilot, engines change their retrieval, and your own team ships a release. None of that is isolated in 30 days, so a pilot result should be read as a direction with a confidence level rather than as a measured effect. Our page on why AI answers change every time explains the underlying variance in more detail.
What should a pilot produce before you sign a retainer?
A pilot should produce four things you keep: the raw runs, the changed pages, a written account of what was learned, and a proposal with exit terms. If any of those is missing at the end, you have bought a presentation.
Raw runs mean the saved answers, prompt by prompt, engine by engine, with dates. Ask for them during the pilot rather than at the end. A supplier who can produce last week's answers within a day is running the measurement they described; one who sends a chart instead has a reporting layer and no archive underneath it.
Changed pages should be readable as a diff. You want to see the heading that was rewritten, the answer that moved to the top of the section, and the sourced number that was added, because that is how you judge whether their page work is better than what your own team would have written.
The proposal matters as much as the work. It should name what happens next, what it depends on from your side, and the conditions under which you would stop, including a notice period and what you keep if you leave. Treat any guarantee of a ranking or a citation as a reason to walk, because no supplier controls either, and our guide to choosing an AEO agency covers the rest of the diligence.
If you want an independent reading of where you stand before any of this, a free visibility assessment runs your buyer questions across ChatGPT, Claude, Perplexity and Google AI Overviews and reports who is named today, from which page, with the raw answers attached.
When is a pilot the wrong move?
A pilot is the wrong move when the blocker is technical, when nobody internally will own the work, or when scoping it costs more than a month of the real thing. In each case the pilot spends your attention on the wrong question.
Technical blockers come first because they invalidate everything else. If your answer content only renders in the browser, or the crawlers are blocked at the CDN, then a pilot measuring citations is measuring a site the engines cannot read. Fix access, then measure, and our page on checking whether AI can read your website gives you the half hour version of that test.
Ownership is the second. Someone on your side has to approve page changes, answer product questions and decide what the pages may claim. Without that person, the pilot stalls in week two and the conclusion you draw is about the supplier when it was about your own capacity. Our page on who should own AI search visibility covers how teams usually resolve it.
The third case is arithmetic. A pilot has its own kickoff, scoping, reporting and handover, and when that overhead approaches the cost of a month of the actual engagement you are paying for packaging. Ask for the monthly scope with a notice period instead, and judge it on the same evidence a pilot would have produced. What drives the price either way is the number of questions tracked, the number of engines, how many pages get rewritten and whether off-site work is included, which is covered in what AEO costs and what moves a quote.
What else do people ask about an AEO pilot?
How long should an AEO pilot be?
An AEO pilot should run about eight weeks if you want to see movement, and 30 days is enough if you are judging the supplier. The first month tells you whether the measurement is real, whether the page work is better than yours and whether decisions are being made from the data. Movement in citations takes longer because pages have to be re-crawled, and because repeated runs are needed before a change can be separated from ordinary variance.
Should an AEO pilot be paid or free?
Paid, in almost every case. A free pilot is a sales exercise, and the work it contains is sized to the sales cycle rather than to your site. A paid, scoped first phase buys you deliverables you keep: the baseline, the raw answers, the changed pages and the write-up. A free assessment is a different thing and a reasonable way to start, because it tells you where you stand before anybody scopes work.
What should you measure during an AEO pilot?
Measure whether your company is named in the answers to a frozen set of buyer questions, which of your pages are cited as the source, and which competitors appear where you do not. Record the raw answers each run. Avoid composite visibility scores as the headline number during a pilot, because they combine retrieval, mentions and citations into one figure that can move without any citation changing.
Can a pilot tell you whether AEO will work for your company?
A pilot tells you whether your buyers' questions produce answers you are absent from, and whether your pages can be fixed. That is most of what you need. What it cannot tell you is the size of the eventual outcome, because the measurement is noisy at short horizons and new pages take time to be read. Treat a pilot as a test of the supplier and of your own capacity to act, not as a forecast.
What happens to the work if you do not continue after the pilot?
Agree that in writing before the pilot starts. The pages, the prompt set, the raw answers and any accounts opened in your name should be yours and should stay accessible, in your CMS and your own storage rather than in a supplier dashboard. Ask specifically about the measurement history, because a baseline you cannot export is a baseline you lose, and rebuilding it costs you the comparability you paid for.