A retainer is a bet that the thing you're optimizing changes slowly enough for a calendar to keep up.
For twenty years that bet was correct. Google confirmed a few core updates a year. Quarterly content, annual technical audit, monthly report. Human hours could be scheduled against a published rhythm.
In 2026 the buyer doesn't start at Google. They ask ChatGPT for a shortlist, Perplexity for a comparison, Copilot inside the software they already have open. Those systems retrieve, rank and cite differently from each other, and their behavior shifts when models are retrained and grounding rules are tuned.
None of them publish a changelog. That single fact breaks the retainer, and it's also why “just buy the software” isn't the answer. Software alone watches. It doesn't decide, prioritize, or ship the fix. What closes the gap is a program: software that measures continuously, paired with the people who run the strategy and do the work.
“You can't schedule labor against a system that changes without notice. You can only instrument it, then spend the human hours on decisions, not on collecting data.”
Andrew Bethel, COO, Perfectus Labs
Detection lag is the whole argument#
Forget philosophy. There's one number that decides this, and it's the gap between when your visibility changed and when a human noticed.
On a monthly-report cadence, the expected lag is about 15 days, and the realistic lag is closer to 45 once the report is written, read and discussed. Continuous instrumentation closes that gap to the next crawl.
Now price it. If AI-assisted discovery drives even $80k of quarterly pipeline, 45 days of silent decay isn't a reporting inconvenience. It's roughly $40k of pipeline that quietly didn't happen, and nobody on either side of the contract did anything wrong. The cadence just didn't match the system.
That's the argument. Everything below is detail.
What actually arrives in your inbox#
Ask what the deliverable is, then look at what shows up.
A document. Audits, recommendations, roadmaps. Someone on your side then has to implement it, and the implementation queue is where most of this work goes to die. You paid for the diagnosis and inherited the surgery.
A change in production. Schema corrected, render path fixed, entity structure rebuilt, the citation footprint measurably different this month than last.
If what you receive is a PDF with better formatting than last quarter's PDF, you didn't buy a program. You bought a report.
What a program does that scheduled hours can't#
It watches engines that don't report to you. Google Search Console shows you one window. It says nothing about how ChatGPT or Perplexity assembled the shortlist your buyer actually read. Those have to be sampled directly, on a schedule, at a volume nobody is doing by hand.
Its cost per check is effectively zero. Crawl, render, schema validation, entity checks and citation sampling run continuously. Human judgment is then spent on decisions and strategy rather than on collecting data, which is the only part of this work that was ever worth an hourly rate.
It compounds across accounts. Every site the system watches sharpens the pattern library. Capacity models do the opposite: they get slower as they scale, because the same pod absorbs more accounts.
What you keep when you stop#
This is the question almost nobody asks during a renewal.
Stop paying for capacity and the work stops that day. Stop paying for a program and the schema, the corrected technical foundation, the entity structure and the citations already earned stay on your side of the fence, still working. One model rents you activity. The other builds an asset you own.
When a program isn't the right fit#
Straight answer, against our own interest: if you publish four times a year and run a fifteen-page site, you don't need a continuous program. You need one competent person for two weeks. Brand positioning, creative concepting and campaign taste are human work. No system has taste.
What is not human work anymore is measure, detect, correct, repeat. A program runs that loop continuously. Paying an hourly rate for it is the expensive way to be late.
Four questions that end the ambiguity#
Ask any prospective partner, including us:
- How often do you check retrieval behavior, and what triggers a change? “Monthly reporting” is a calendar, not a trigger.
- Do you ship the recommendation or the fix?
- Which engines do you measure besides Google? If the answer is only Google, you're buying 2019.
- If we stop next quarter, what do we keep?
A partner who can't answer all four cleanly is selling hours with program vocabulary bolted on.
The short version#
The question isn't whether a team is good. Plenty of them are. The question is whether the delivery model matches the cadence of the thing you're trying to win, and answer engines move on their own schedule, not on yours.
A managed program pairs the software that detects with the people who decide and ship. That's the whole point: the instrumentation and the work are one contract, run continuously, against a system that changes whenever it wants.
If you want to see what your detection lag currently is, book a call. We'll show you where you're being cited today, where you're not, and what changed while nobody was watching.
