What to automate and what not in a GEO service: the operations of an agency with 10+ clients
Do the arithmetic before signing the tenth client. A GEO service —Generative Engine Optimization: working on a brand's presence in the answers of AI search engines, like ChatGPT, Gemini or Perplexity— requires checking dozens of questions across several AIs, every week, for each client. With 40 prompts, 4 AIs and weekly frequency, that's 160 checks a week per client. With ten clients, 1,600. By hand, that's not a service: it's a screenshot factory that eats the margin.
The obvious way out is to automate. The equally obvious mistake is automating everything: the client who gets a report no one on your team has read ends up noticing, and the article published without review ends up claiming something the business doesn't offer. This article draws the line between the two, thinking about the real operations of an agency with ten or more active accounts.
The general rule: automate the collection, never the judgement
Everything that's repetitive, voluminous and measurable, a machine does better. Everything that requires knowing the client's business, interpreting a data point or sustaining a relationship, a person does better (and only a person can do). When you're in doubt about a specific task, ask yourself: is the value in doing it or in deciding about it? The former gets automated; the latter doesn't.
What does get automated, no regrets
1. Monitoring. Launching the same dozens of questions to the 4 AIs, week after week, and noting whether they mention the client, what they say about them and which sources they cite. It's machine work by definition: as well as repetitive, it's statistical —AI answers vary between runs, so you need repeated samples over time, not a loose Tuesday screenshot. The detailed arithmetic of what doing this by hand costs is in how to monitor all your clients across 4 AIs without screenshots.
2. Collecting the baseline data. A new client's initial audit —what each AI says today, which competitors appear in their place, which sources feed those answers— is the same mechanics as monitoring, compressed into a week. Automate it and dedicate the human hours to what comes after: reading the results and deciding the plan.
3. The report drafts. Turning the month's data into tables, evolution charts and a PDF for the client is layout, not consultancy. What must never be automatic is the prior reading: the paragraph where you explain what the data means and what you're going to do about it is the one that justifies the retainer.
4. The content drafts. The first version of an article optimised for AIs to cite can be generated from the audit data: which questions have no published answer, what the competition says, which fact is missing. Draft, emphasised. The decision of what gets published and with what claims is your team's, and we'll see why shortly.
What you must never automate
Strategy. Deciding which 40 or 75 questions to measure is the most important decision of the service: it defines what "winning" means for that client. It comes from knowing their business —what they really sell, which client leaves them a margin, what people ask them over the phone—, not from a template. If two of your clients in the same sector have the same prompt list, you've automated it badly.
The conversation with the client. A monthly PDF retains no one; the call where you explain what happened, what you've done and what's next, does. A retainer's churn isn't fought with prettier reports but with someone who shows their face. What to say in that meeting, with volatile data like the AIs', is covered in what to show in the reporting meeting.
Editorial judgement. No draft gets published without someone who knows the client reading it in full: prices, timeframes, services, claims. Picture the worst scenario: an AI citing an incorrect fact about your client because your own content published it wrong. That error isn't fixed by an apology; reviewing 20-30 minutes per piece prevents it.
The mistake of automating everything (and how you pay for it)
There are agencies industrialising GEO in the style of 2012 SEO: mass content with no review and reports no one opens before sending them. It fails on two fronts. The first is commercial: the day the client asks in a meeting about a fact in their own report and no one can answer, the trust doesn't come back. The second is technical, and has its irony: AIs cite pieces with concrete data and specific answers, exactly what filler content generated in bulk doesn't have. Automating editorial judgement produces the one kind of content GEO doesn't need. Which pieces really work is detailed in the monthly content plan for a GEO client.
The typical week with 10 clients, well distributed
For the numbers to work, the machine part is done by a tool: with Surfeo, each client in the portfolio has their 40-75 prompts checked every week across 3-4 AIs, the data accumulating in their dashboard and between 6 and 16 article drafts a month waiting for review, depending on the contracted tier. On that base, the human week looks like this:
- Monday: review of the week's alerts and significant changes (1-2 h for the whole portfolio; you only look at what's moved).
- Tuesday to Thursday: editorial review of drafts and off-site work —reviews, directories, sources— in client blocks (30-45 min per client per week).
- Friday: reports for the clients closing their cycle that month: read the data, write the interpretation, prepare the call.
Total: between 2 and 3 human hours per client per month at cruising speed, plus the meeting. With that, the service's margin holds; with the manual version, it doesn't.
Frequently asked questions
How many team hours does a GEO client need a month if you automate well?
Between 2 and 3 hours of operations (reviewing drafts, off-site, interpreting data) plus the reporting meeting. The first month is the exception: the baseline and the initial plan demand quite a bit more human dedication, as we detail in what to deliver in the first month.
Can I let the content publish itself, without review?
You can; you shouldn't. The risk isn't abstract: it's publishing prices, timeframes or services that aren't the client's and an AI repeating them with your signature underneath. Human review is the part of the service the client can't do themselves with ChatGPT, so as well as protecting you, it sets you apart.
Which part does the client see: the automatic or the human?
They should see both, hiding neither. Automatic monitoring is a sales argument ("we measure your presence every week across 4 AIs, not once a quarter by hand") and the human layer is what makes sense on top. Selling only the machine turns you into a software vendor without being one; selling only the hours makes you incomparable on price.
Where do I start if today I do it all by hand?
With monitoring, which is where the most hours are burned with the least judgement involved. After that, the report drafts. The last thing I'd touch is the content, because there the step from "automatic draft" to "automatic publishing" is slippery and it's worth having the review habit well settled first.
Want to run the numbers with your real portfolio? The agency plan starts at €20/month plus €35-79 per client depending on tier, with up to 10 clients — see the detail in pricing and compare it with what a single weekly hour of manual screenshots costs.
