Most agency delivery problems don’t look like delivery problems at first. They look like a good month. More clients, more deliverables, more Slack threads. Then a piece of work ships without the citation audit someone usually remembers to run, a client asks why their report looks different from last month’s, and the whole team discovers that the process everyone assumed existed was actually living inside one person’s head.
That’s the real cost of ad-hoc work: it isn’t slow, it’s inconsistent. And in AI search, inconsistency is expensive. When a large language model decides which sources to pull into an answer, it’s weighing signals that accumulate over time — clarity, structure, corroboration across places people actually talk. A campaign that runs sharply for six weeks and sloppily for the next six doesn’t compound. It resets.
Over a decade in reputation and search, I’ve come to treat systems as the deliverable underneath the deliverable. The work of Cory Maki SaaS AI visibility engagements — the audits, the content briefs, the Reddit and community research, the citation tracking — is only as good as the process that makes it repeatable when I’m not the one doing it.
Why ad-hoc delivery breaks at exactly the wrong moment
Every agency and every SaaS team starts ad-hoc. That’s correct. You don’t know what the process should be until you’ve done the work enough times to see the pattern. The mistake is staying there past the point of learning.
Ad-hoc delivery breaks in three predictable ways:
- Quality becomes personality-dependent. The output is excellent when your best strategist touches it and average when they don’t. Clients feel that gap even if they can’t name it.
- Onboarding takes forever. New team members learn by shadowing, which means your senior people stop producing in order to teach, and the teaching is never quite the same twice.
- Nothing is measurable. If the process changes silently every month, you can’t tell whether a result came from the strategy or the execution. You lose the ability to learn from your own work.
That third one matters most for AI search. Generative Engine Optimization (GEO) — the practice of earning visibility inside AI-generated answers rather than just in the blue links — is a feedback-driven discipline. You publish, you monitor which sources ChatGPT, Perplexity and Google AI Overviews actually cite for a query set, and you adjust. If your execution is a moving target, the feedback tells you nothing.
What a workflow actually is (and isn’t)
A workflow isn’t a 40-page SOP nobody reads. The way I think about this: a workflow is the shortest set of steps that produces the same quality of output regardless of who runs it, plus the checks that catch it when something goes wrong.
Three components, in order of importance:
- Inputs. What has to be true before work starts — access, brand context, approved entity language, a defined query set. Most rework traces back to a missing input, not a bad execution.
- Steps with owners. Each step names a person or a system, not a team. “Research” is not an owner. “Analyst pulls the top 20 cited domains for the query set into the tracker” is.
- Definition of done. A short, concrete checklist. Not “high quality,” but “answers the question in the first 60 words, includes a comparison table, names competitors accurately, cites primary sources.”
Write it at the level of detail a competent new hire needs — not less, not more. If a step requires judgment, say so explicitly and name who makes the call. Pretending judgment can be proceduralized is how teams end up with documentation they quietly ignore.
A concrete example: the GEO content pipeline
Here’s what this looks like in practice for a SaaS AI visibility engagement. The goal is AI citations — getting the client’s name and content surfaced as a source inside AI answers — and the pipeline has five stages.
1. Query set definition. Before anything is written, we lock a list of the prompts real buyers type: category comparisons, “best tool for X,” alternatives queries, integration and pricing questions. This becomes the measurement baseline. It’s an input, and nothing downstream starts until it exists.
2. Citation landscape audit. Run the query set across the major AI surfaces and record which domains, threads and review pages get pulled in. The output is a ranked list of where the answers are actually being formed — often Reddit, review platforms, comparison sites and a handful of publications, rather than the client’s own blog.
3. Gap and asset mapping. Compare what those cited sources say about the category against what the client can credibly claim. This is where reputation work and content work meet. If the third-party record is thin or wrong, publishing more owned content won’t fix it — that’s a online reputation management problem wearing a content costume.
4. Production against a citability standard. Every asset ships against the same structural checklist: a direct answer near the top, clear headings that mirror real questions, defined terms, specific numbers where we have them, and no unsupported claims. Structure is what makes a page easy for a model to extract and attribute. This is the core of the ARC Method and AI-citation frameworks I use to keep production consistent across writers.
5. Re-measurement and review. Rerun the query set on a fixed cadence. Log changes. Feed what moved back into stage three.
Notice that nothing in that pipeline is clever. The value is that it runs the same way every cycle, which means the results are comparable and the team can be trained on it. If you’re earlier in the journey, my walkthrough of where SaaS founders should start with AI visibility covers the first two stages in more detail.
Where fulfillment automation belongs
Once a workflow is stable, automation gets easy — and not before. Automating an undefined process just produces mistakes faster.
In my work with clients, fulfillment automation earns its keep in the unglamorous middle: intake forms that enforce required inputs, scripts that pull citation data into a shared tracker, templated reports that assemble themselves from that tracker, task creation that fires when a stage completes, and AI agents that handle first-pass research summaries or draft outlines against the citability checklist.
What I don’t automate is judgment — which queries matter, whether a claim is defensible, how to handle a reputation issue, whether a piece is actually good. Those stay human, and the workflow should make it obvious where a human has to sign off. I’ve written more about drawing that line in automating fulfillment without losing quality.
A simple test: if a step is the same every time and the failure mode is “someone forgot,” automate it. If the step is different every time and the failure mode is “someone made the wrong call,” systematize the decision criteria instead and leave the call to a person.
How to build your first repeatable workflow this month
- Pick the deliverable you produce most often. Not the most complex one. The most frequent one — that’s where consistency compounds fastest.
- Document the last three times you did it. Actual steps, actual order, including the parts you improvised. The gap between those three versions is your process debt.
- Write the definition of done first. Work backwards from it. Most teams write steps first and never agree on what finished means.
- Run it once with someone else driving. Every place they have to ask you a question is a missing step. Fix it in the doc immediately, not later.
- Automate one input and one handoff. Start with the intake form and the task trigger. Leave the creative work alone for now.
- Set a review date. A workflow you never revise becomes a workflow people route around.
The durable principle
What I’m seeing across AI search is that visibility rewards consistency more than intensity. Models synthesize from an accumulated record — what’s published, what’s structured clearly, what other people corroborate. One brilliant month doesn’t build that. A hundred competent, repeatable cycles do.
That’s why I treat delivery systems as strategy, not admin. Systems and automation are what let quality scale past the capacity of any single person, and quality sustained over time is what earns citations. Reputation is earned that way too — slowly, through a record that holds up under scrutiny, which is the same logic behind AI reputation management work.
Build the workflow. Automate the parts that are the same every time. Protect the judgment. The results take care of themselves at a pace you can actually repeat.