Get four things out of your current agency before you give notice: the tracked question set as a plain list, the raw run logs behind every report they sent you, a record of every page they changed, and the technical fixes they implemented. Those four are your progress. The citations are not portable and never were, because the engines rotate sources on their own: a study by Parse of more than 16,000 prompts, each re-run around 22 times in March and April 2026, found repeat ChatGPT answers to the same prompt shared only about a fifth of their cited sources. What a handover can genuinely destroy is the measurement series that tells you which changes worked, so treat the logs as the asset, overlap the two suppliers by a month, and freeze site changes until the new team has its own baseline.
Switching AEO agencies without losing progress comes down to extracting four artifacts first and changing nothing on the site until the new supplier has measured it. The order matters more than the paperwork: ask for the artifacts while you are still a paying client, because the request is routine then and becomes a negotiation the moment you give notice.
The four artifacts are the question set, the raw run logs, the page change record and the implemented technical fixes. A question set is the list of buyer prompts somebody has been running on your behalf, and it usually took a month of argument to agree. Raw run logs are the actual answers the engines gave, with dates and the sources they named, which is the only evidence that any reported number was ever real. The page change record tells the next team which of your pages have already been rewritten for citation and which are untouched. Technical fixes cover rendering, headings and schema work that may live in a repository your agency has access to.
Site changes are the other half. A handover is exactly when somebody pulls a landing page, reverts a template or removes structured data with a plugin, and then the first month under the new supplier looks like a collapse that nobody can attribute. Freeze publishing and template work for the overlap period, or log every change with a date so the new baseline can be read against it.
What you cannot carry across is the citation position itself, and expecting to is where most of the disappointment in these transitions comes from. Your pages keep whatever standing they have earned, but the week-to-week numbers move for reasons that have nothing to do with which agency is invoicing you.
What you lose in an agency change is the measurement history and the institutional memory, not the visibility. Both are recoverable if you ask for them in the right order, and neither is recoverable once the account manager has moved on and the tool seat has been cancelled.
| Asset | Why it matters to the next team | Who usually holds it | How to get it out |
|---|---|---|---|
| Tracked question set | Defines what "visibility" has been measured against, and takes weeks to agree from scratch | The agency's tracking tool, sometimes only in a report template | Ask for it as a spreadsheet, one prompt per row, with the engine and market noted |
| Raw run logs | The only evidence behind past reports, and the baseline every later comparison needs | A third-party platform on the agency's seat | Request a full export, including the answer text and cited URLs, not a PDF summary |
| Page change record | Separates pages already written for citation from pages nobody has touched | The agency's project tracker, plus your CMS revision history | Ask for a page-by-page list with dates, then verify it against your own CMS |
| Technical fixes | Prevents the new team spending a month rediscovering rendering and schema work | Your repository or CMS, under the agency's logins | Get the ticket list and the commits, then revoke the logins after the overlap |
| Account access | Search Console, analytics and tag manager history that predates the engagement | Accounts the agency created under its own domain | Move ownership to an account on your domain before notice, not after |
Contracts are usually silent on the first two rows. A 2026 review of agency contract terms by Topify notes that most SEO contract templates say nothing about who owns the AI visibility tracking history, and advises asking in writing whether that history transfers or resets to zero with the next provider. Ask before you need the answer.
Accounts created by an agency are the quiet version of the same problem. A property set up under the supplier's own workspace can leave with them, taking the pre-engagement history that makes your before-and-after argument possible.
Your citation count moves the week you switch because AI engines re-select sources constantly, and a month of churn is normal with no agency involved at all. Reading that movement as the new team's fault, or the old team's, is the most common mistake in these transitions.
The size of the churn is now measured. Parse's volatility research, covering more than 16,000 prompts each re-run about 22 times across ChatGPT and Google AI Overviews in March and April 2026, found repeated answers to one prompt shared roughly a fifth of their cited sources on ChatGPT and about 31% on AI Overviews.
Day-to-day the picture is similar. Detailed's study of more than 70,000 answers over 28 days found only about a quarter of the pages ChatGPT cited reappeared the next day, against about half on AI Overviews. A single run before the switch and a single run after it can therefore differ by more than any agency could plausibly cause in four weeks.
Retrieval is not citation either, which widens the noise further. AirOps' 2026 analysis of the ChatGPT fan-out pipeline reports that only about 15% of retrieved pages end up cited, so a page can be read and passed over repeatedly while nothing about it has changed.
The practical defence is sampling. Run the question set at least twice before the handover and twice after it, a week apart, and compare the averages rather than the snapshots. Our guide to why AI engines give a different answer every time explains the mechanism, and how many prompts you need to track covers how large a sample has to be before the comparison means anything.
Before you give notice, ask for seven specific deliverables in writing, with a date attached to each. Vague requests for "our data" get a PDF; named files get files.
Start with the question set as a spreadsheet, one prompt per row, with the engine, the market and the competitors tracked alongside it. Then the raw export of every run behind the reports you have received, with answer text, cited URLs and timestamps. Then the page change record, dated, page by page. Then the technical ticket list with links to commits or CMS revisions.
The fifth is a list of the third-party placements the agency pursued on your behalf, including listicles, directories and review profiles, because off-site work that nobody documents is the easiest thing to pay for twice. Our guide to getting into the best-of lists AI engines quote describes what that work looks like when it is done properly.
The sixth is administrative: a written confirmation of which accounts and properties they created, and transfer of ownership to logins on your own domain. The seventh is a short handover note naming the pages they would have worked on next and why, which costs them an hour and saves the next team a fortnight.
Give one date for all seven and keep the tone ordinary. Most agencies deliver without argument, because a client asking for their own prompt list in month nine looks like a client preparing a budget, not one preparing to leave. If the answer is a refusal or a long silence, you have learned something about the engagement that no reference call would have told you.
Overlap the two suppliers by about a month, which is long enough for the incoming team to run its own baseline while the outgoing one is still contractually obliged to answer questions. Agencies that handle transitions regularly recommend a similar window: Shout Digital's guide to switching agencies without losing rankings suggests keeping both engagements live for at least 30 days, and longer for complex setups.
Inside that month, the incoming team should do three things and nothing else. Run the inherited question set unchanged, twice, a week apart. Read your top pages against the questions they are supposed to win. Produce a fault list with the pages named. Changing the prompts in week one is tempting and wrong, because a question set that moves at the same time as the supplier leaves you with no comparable series at all.
Hold the site still for the same month. No template changes, no bulk publishing, no plugin upgrades that touch structured data. Our guide on whether a redesign or migration hurts AI visibility covers what to watch if a release cannot wait.
Use the overlap to compare the two teams on the same evidence, which is a rare opportunity. Hand both the same ten prompts and ask each to explain, in a page, why your name does or does not appear and what they would change first. Two answers on one question set tell you more about judgement than any proposal does, and the answers arrive while you still have leverage with both.
The agency is the problem when the measurement is missing or unverifiable, and the channel is slow when the measurement is solid and the numbers have barely moved. Separating those two is worth doing before you pay anyone to transition, because a switch costs you a quarter either way.
Decay is real and it is not evidence of failure. A study of 3.5 million citation events collected between September 2025 and March 2026 by Scrunch and Stacker, reported by Similarweb's AI search team, put the median time for half a page's AI citations to drop off at about 4.5 weeks, with ChatGPT around 3.4 weeks and Perplexity around 5.8. On those numbers, holding a position at all requires continuing work, so flat-but-stable is a result rather than an absence of one.
Three things point at the supplier instead. Reports with no attached raw answers, because the summary cannot then be checked. A question set that keeps changing, which quietly resets the comparison every month. And a work log that is all new articles with no rewriting of the pages you already have, which is content production invoiced as citation work.
Our guides on how to know if your AEO is actually working and how long AEO takes to work set out the timelines and the warning signs in more detail. If you would rather settle it with an independent reading than an argument, 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.
Put four clauses in the next contract and the handover stops being a negotiation. None of them is unusual, and an agency that objects to all four is telling you how the exit will go.
The first clause names the raw run log as a deliverable, exported monthly to storage you control, in a machine-readable format. Monthly matters more than the format: a log you receive every month cannot be withheld at the end, and it makes every report checkable on the day it arrives.
The second clause states that the question set, the page change record and any prompt library built during the engagement are your property. The third gives you a named handover document at termination: accounts created, work in progress, pages changed, and the next three things they would have done. The fourth requires that accounts and properties be created under your domain, with the agency added as a user rather than an owner.
Add a review point rather than a long term. A twelve month contract with no scheduled point at which either side can stop is the structure that produces reluctant year two renewals and bad-tempered exits. A quarterly review with an agreed exit right costs the supplier nothing if the work is good. Our guide to choosing an AEO agency lists the other terms worth settling before signature, and whether AEO is a project or a retainer explains which parts of the work genuinely recur.
Ask for the handover artifacts before you say anything about leaving, because a request for your own question set and run log is an ordinary request from a client and an awkward one from a client who has just given notice. Nothing in the list is unreasonable to ask for mid-engagement: the prompt list, the raw logs, the page change record and the technical fixes are all things you are already paying for. Once you have them in your own storage, the conversation about renewal is a conversation between equals.
A new agency can rebuild a baseline in a few weeks, but it rebuilds a baseline for today rather than the one you paid for last year. Because the engines rotate sources on their own, a fresh baseline cannot tell you whether a page that stopped being cited in March was overtaken by a competitor or drifted out on its own. The practical consequence is that you lose the ability to argue about cause, which is the main thing a long series is for.
Changing agencies does not by itself cost you citations, because the pages stay where they are and the engines do not know your invoice changed. The risk is in what happens around the switch: pages pulled down, templates reverted, schema removed with a plugin, or a redesign shipped while nobody is measuring. Freeze site changes during the handover and the citation side of the switch takes care of itself.
Ask once in writing, name the specific files, and set a date, because most refusals are a scheduling problem rather than a policy. If the answer is still no, take what you can reconstruct: the prompt list from past reports, the page edits from your CMS revision history, and the technical changes from your repository. Then start a fresh series with the new supplier and treat the old reports as background rather than evidence.
A new agency should hand you two baseline runs and a prioritised page list inside the first month, because that work needs no publishing approval and no engineering time. Movement in citations takes longer and depends on how fast your team can approve changes. What you can judge at week four is whether the measurement is real: same prompts, same engines, raw answers attached, and a list of faults specific enough to argue with.
A free visibility assessment runs your buyer questions across ChatGPT, Claude, Perplexity and Google AI Overviews and reports who is named and from which page, so a handover begins with evidence instead of a handover note.
Request a free visibility assessment →