SIGNALS
Before you sign

How do you take AI search visibility in-house after an agency has run it?

The short version

Taking AI search visibility in-house after an agency has run it is mostly a handover problem rather than a hiring problem. Four artefacts decide whether the transition costs you a fortnight or a quarter: the prompt set as a file, the full run history with the prompt, engine, date and every cited URL for each run, a page change log by URL and date, and a written list of what was tried and did not work. Ask for all four in writing before the final invoice, because an agency has every incentive to help while it is still being paid and none afterwards. The measurement series is the part people underestimate. Published decay rates mean a single reading describes a page for about a month, so if the history does not transfer, your new in-house baseline starts from zero and the first three months produce numbers nobody can interpret.

How do you take AI search visibility in-house after an agency has run it?

Taking AI search visibility in-house after an agency has run it comes down to transferring four artefacts and one cadence, then proving you can reproduce last month's number without the agency's help. The skills are learnable and the tools are buyable. What is not recoverable after the fact is the record of what has already been measured and changed.

Sequence the exit in that order. Get the artefacts while the relationship is live, run one measurement cycle in parallel with the agency still in place, compare the two readings, and only then end the contract. A handover that starts after the last invoice is an archaeology project, and the agency's obligation to answer the phone has already ended.

Be clear with yourself about which half of the work you are taking on. The page work is visible and feels like the job. The other half is a standing measurement series and a presence programme off your own domain, and both are easy to drop in a reorganisation because neither produces anything to look at in week one.

Our guide on how to switch AEO agencies without losing your progress covers the same handover when the destination is another firm rather than your own team, and the artefact list is identical. What changes in-house is who carries the cadence.

What do you need to get out of the agency before the final invoice?

Get four things out of the agency before the final invoice, in files you hold rather than views inside a platform. The test for each is whether a new person could pick the work up from it in a day without asking the old supplier anything.

What to ask for Why it matters What it costs you if it is missing
The prompt set, as a file Defines what was ever being measured Your next report is not comparable with any earlier one
Full run history: prompt, engine, date, every cited URL Turns readings into a series you can plot A new baseline, and a quarter before any trend is legible
Page change log by URL and date Lets a movement be lined up against a cause Every later change in citations is unattributable
Written list of what was tried and failed Stops you buying the same dead end twice Repeated work, usually in month two or three
Account ownership and billing for any tools Keeps the data when the seat is removed History deleted at the end of the billing period
Drafts and source notes for pages in progress Half-finished pages are the commonest casualty Rewriting work already paid for

Put the list in the contract rather than the exit conversation. The four artefacts above are the same ones worth specifying when you buy, and our guide to what to put in an RFP for AI search visibility covers the exit and portability clause that makes them enforceable instead of negotiable.

One practical warning about platform exports. A tool that reports a monthly summary is not the same as a tool that exports the underlying runs, and the difference only becomes visible on the day you ask. Check the export before the notice period starts.

Why is the run history worth more than the monthly reports?

The run history is worth more than the monthly reports because the reports are conclusions and the history is the evidence, and in this category the evidence decays fast enough that you cannot reconstruct it later. A summary saying share of voice rose four points is unusable once you need to know which prompts moved and which URLs replaced you.

Decay is the reason the series matters. Trakkr found that 73.5% of citation URLs appeared exactly once and did not return, measured on 108,650 URLs from a 200-brand sample, and that the median brand fell to half its peak citation count in 31 days, measured on a brand-level set of 857,138 daily reports across 10,991 brands and eight models, in "The Half-Life of AI Citations" (published 22 April 2026, read 19 November 2026). A single reading describes a brand for about a month, so a year of monthly summaries with no runs behind them is twelve snapshots of a moving thing.

Volume decides how much movement means anything, which is the other thing only the history can tell you. BrightEdge's volatility analysis reported frequently cited domains moving about 0.7% week to week against swings above 50% for sporadically cited ones, a 70-fold gap it found across ChatGPT, Perplexity, Google AI Overviews and AI Mode, with an inflection near 50 citations where weekly volatility drops from 50% to 8%, in "AI Search Engine Citation Volatility: The 70x Stability Gap" (23 October 2025, read 19 November 2026). Below that volume a missing citation is one sample, and only your own history tells you which side of it you are on.

Our guide to measuring AI citation decay works through the two numbers that come off the series, citation half-life and the one-and-done rate, and how to read them without over-reacting to a single run.

Who should own this in-house, and what do they do each month?

One named person should own it in-house, with content capacity behind them and a developer they can reach, and their monthly job is a measurement cycle rather than a content calendar. The owner does not need to write the pages. The owner needs to run the same prompts on schedule, read what changed, and decide what gets rewritten next.

The cadence is the part worth copying wholesale, because it is what makes the numbers mean anything. Our published methodology sets out four kinds of run and what each is for: two baselines a week apart before any change, to size your own noise; a verification run as soon as fixes are live, to confirm the structural change registered; monthly tracking through active work; and a quarterly maintenance run once the work stops, to catch a page that has started to slide.

What our own research looks like in practice gives a sense of scale. For the category study behind our pharmaceutical process skid report we ran the buyer's question across four engines, five times each, for eight skid categories plus two regional markets. A design of that shape produces a readable number rather than an anecdote.

Where the role sits matters less than whether it is one person's named responsibility. Our answer on who should own AI search visibility inside your company covers the trade-offs between SEO, content, PR and product marketing, and what goes wrong when the answer is everybody.

What usually breaks in the first quarter after the agency leaves?

Four things break in the first quarter after an agency leaves, and three of them are invisible until the numbers stop making sense. The off-site work stops first, because nobody in-house has pitching as a job. The prompt set drifts next, as somebody adds the prompts leadership asked about and the series quietly becomes a different series. Nobody re-baselines after the first big site change. And reporting reverts to screenshots of answers, which is where the discipline usually dies.

The off-site loss is the expensive one. The largest study separating page properties from retrieval found that which pages the engine's own search returns explains about 32% of citation variation while page properties explain about 2.8%, with brand-name swaps moving product rankings over identical facts, according to Chen and Koudas in "What Drives AI Citations?" (arXiv 2610.15033, 10 October 2026), across 158,099 answers and 468,037 citations. An in-house programme that only edits pages has kept the smaller half of the mechanism.

Protect the prompt set with a rule rather than good intentions. Freeze the original set as the comparable series, and add new prompts to a second list that is reported separately until it has its own baseline. Teams that mix the two lose a year of comparability in one afternoon, usually for a board meeting.

If you want an outside reading at the handover point, a free visibility assessment run in the same month as the agency's last report gives you two independent readings of the same moment. Nothing cheaper will tell you whether the numbers you inherited describe reality.

How do you tell whether the handover actually worked?

The handover worked if your team can reproduce the agency's last number from the artefacts you hold, and then produce the next one on schedule without asking anybody. Run that test deliberately in the overlap month rather than hoping.

Three checks cover it. Re-run the inherited prompt set yourself and compare your result with the agency's final report, expecting a small difference and investigating a large one. Pick one page the agency changed and trace it through the change log to the run where its citations moved, which proves the log and the history line up. Then write the next monthly report from your own data, in the same format, and see whether anybody can read it.

Hold the same rule the agency should have held for what counts as a real movement: the same direction across three consecutive runs, rather than any week-on-week rise. A new in-house owner under pressure to show value is the most likely person in the company to report noise as progress, and the rule exists to protect them from that.

Expect the first quarter to look flat and plan the reporting narrative accordingly. Structural changes take weeks to show up in citation frequency, the baseline has just been reset, and the honest report for month one says what was measured and what was changed. Our guide on how to tell whether your AEO is actually working covers which signals move first and in what order.

What else do people ask about taking AI search visibility in-house?

Teams taking AI search visibility in-house from an agency ask these five questions most often.

Can an agency refuse to hand over the prompt set and run history?

In practice yes, unless the contract says otherwise, and that is why the portability clause belongs in the agreement rather than the exit email. Where no clause exists, ask while the final invoice is still unpaid and ask for files rather than access. An agency whose platform genuinely cannot export the runs is telling you something useful about what you were buying.

How long should the overlap be between the agency leaving and the team taking over?

One full measurement cycle, which usually means a month, and two if a site change lands in the middle. The purpose of the overlap is a comparison: your team runs the inherited prompt set while the agency runs its own, and the two readings are checked against each other. Without that comparison you cannot tell a handover problem from a real change in the engines.

Do you need to buy an AI visibility tool to run this in-house?

You need somewhere the runs are stored in a comparable form, and that can be a tool or a spreadsheet plus a scheduled script. The tool is worth buying when the prompt set is large enough that manual runs get skipped, and a skipped run is the real failure mode. What matters is that the stored record keeps the prompt, the engine, the date and every cited URL.

Is it cheaper to run AI search visibility in-house?

Cheaper on fees and not always cheaper overall, because the measurement cadence and the off-site work both carry a time cost that shows up as somebody else's deadlines slipping. The honest comparison is against a scope rather than a headline: list what the agency did each month, then decide which of those lines your team will actually keep doing in month six.

What should you do if the agency never kept a change log?

Rebuild what you can from version history and deployment records, then start a clean log from your first day of ownership and say in the first report that attribution before that date is not possible. A log that starts late is still worth keeping. Pretending to attribute an earlier movement to a specific change is worse than admitting the record is missing.

Related guides

The Assessment

Get an independent reading at the handover point.

A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and reads your public pages the way a crawler sees them.

Request a free visibility assessment →
SIGNALS · A BlackSig Systems company