Who Actually Does This Work, and How You Staff It

Two thirds of agencies have the demand. Almost nobody has posted the job.

Semrush pulled 3,900 US SEO job listings off Indeed on November 25, 2025. Across that entire set, SGE, AEO and AI search appeared in 6.3% of senior listings and 3.7% of everything else. Technical SEO appeared in 5.8% of senior listings. Project management appeared in 31%.

Now set that against demand. AgencyAnalytics surveyed 494 agency professionals between February and April 2026 and found 66% reporting increased client demand for AEO and SEO for AI-driven search, the number one new service category of the year. Roughly half of those agencies have ten or fewer full time employees.

So two thirds of the market is being asked for the service, and the job description for the person who delivers it does not exist. Every owner writes one from scratch, and the ones I have read are wish lists: enterprise SEO, LLM architecture, schema, digital PR, prompt monitoring and attribution, in one body, for one salary.

The field knows it has a gap. Lumar interviewed more than two dozen SEO professionals for its 2026 skills report and 81% named GEO, AEO or AI search as the skill to learn next, ahead of data analysis at 50%. That is a self-selected qualitative sample of about twenty five people, so read it as a mood, not a measurement. The mood is unambiguous.

Staffing matters more than the two decisions that precede it. Scope and pricing are decided once. Staffing decides whether that scope gets delivered honestly every month for two years.

The delivery skill profile, in the order that matters

This is the profile I would hire against, ranked by how often the missing piece is what kills the engagement rather than by how impressive it sounds in an interview.

The bottom three are the ones agencies screen for, because they are the ones a portfolio review can assess. The top three decide whether the work is real.

Notice what is absent. No LLM engineering. No prompt engineering as a discipline. No schema specialism, for reasons that follow. If a candidate opens with model architecture, they are describing a research job, and you are hiring for a delivery job.

The three requirements almost no SEO hire meets

Reading logs unprompted. Crawler access is the highest certainty mechanical lever in this field and it is invisible from every dashboard. Cloudflare's network measurement for 2025 put crawl-to-refer ratios as high as 500,000 to 1 for Anthropic and 3,700 to 1 for GPTBot, against 3:1 to 30:1 for Google. Those are completely different traffic economics, and the only way to know which bots reach a specific client site is to open the log file. Most SEOs have never asked a client for one.

Sampling discipline. SparkToro and Gumshoe.ai ran 12 prompts 2,961 times across ChatGPT, Claude and Google AI with 600 volunteers.

There's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses.

Someone who has not internalised that will screenshot a good answer and send it to the client as proof. Next month they will get a bad one and have nothing to say. The requirement is that they fix the prompt set and the repeat count before they look at a single result. That is a research habit, not an SEO habit, and you cannot bolt it on later.

Reporting a null. Three of the most heavily sold tactics in this category have published null results. Ahrefs ran a matched difference-in-differences test on 1,885 pages that added JSON-LD against 4,000 control pages.

Adding schema produced no major uplift in citations on any platform.

SE Ranking analysed nearly 300,000 domains and found llms.txt had no measurable relationship to citation, and that removing it from their model improved prediction accuracy. Profound ran a genuinely controlled experiment, 189 control pages against 192 treatment pages over 21 days.

We tested the claim that serving Markdown to AI crawlers drives more bot traffic. The data doesn't support it, at least not at a scale that would justify treating it as a priority.

A delivery person who cannot say that sentence will sell the client schema and llms.txt instead, because both are easy to invoice and impossible to disprove inside a monthly report. That is the most expensive hiring mistake available in this category, and it is invisible for about nine months.

What ramp actually takes, and why "30 days" is a marketing claim

The guides currently ranking for this query are confident about training time. One states that teams with strong SEO and content skills can transition to GEO delivery with a structured 30-day training program. No source. No cohort. No definition of competent.

I went looking for real ramp evidence and there is none. No published study of time to productivity exists for this role, because the role is roughly eighteen months old. The generic onboarding numbers that circulate, the eight months to full productivity and the rest, trace back to vendor blogs citing each other. I am not going to launder those into a figure and put my name on it.

So here is a derivation instead, with the assumption on the table.

The unit of learning here is the measurement cycle, not the training module. Because answers are as volatile as SparkToro measured, one month of data teaches a new practitioner almost nothing about whether their intervention worked. They need enough cycles to separate their own effect from the noise floor. Three cycles is the minimum at which a month-over-month change starts to mean something. That threshold is my judgment, not a finding, and if you disagree with it the whole timeline moves with you.

Ninety days to supervised delivery. About two quarters to unsupervised. If that reads slow, compare it with the alternative, which is a confident person reporting noise as progress for six months while the client quietly loses faith.

The cost is not tuition, it is protected non-billable time. Five hours a week for thirteen weeks is sixty five hours. At the $98.90 average agency hourly rate Ahrefs got from a poll of 439 SEO professionals, that is roughly $6,400 of foregone billable capacity per trainee per quarter. That poll is dated August 2024, so treat it as the 2024 benchmark rather than today's rate.

Most agencies will not protect that time, and the client-side data says why. The CMO Survey found marketing training budgets have fallen to 3.8% of marketing spending, down from a pre-pandemic 5.8%, and that the most commonly named capability gap was not a skill at all but capability not being resourced with employees, time and money, at 22.3%. The failure mode is not ignorance. It is unfunded intent.

Use this graphic on your site

Free to republish with a link back to this page. Copy the embed code:

<a href="https://josephtimpson.com/insights/how-to-add-geo-services-to-your-agency"><img src="https://josephtimpson.com/assets/infographics/how-to-add-geo-services-to-your-agency.svg" alt="Decision tree with five questions about team capability, learning time, account volume, revenue urgency and null-result tolerance, each leading to train, hire or white label" width="1200" style="max-width:100%;height:auto"></a><p>Graphic by <a href="https://josephtimpson.com/insights/how-to-add-geo-services-to-your-agency">Joseph Timpson</a></p>

The build versus buy break point

Three cost shapes, three different risks. This is the comparison nobody publishes because two of the three answers do not sell anything.

Train internally. First quarter cost is roughly $6,400 in foregone billable time per trainee, plus tooling. No new payroll, and the capability stays in the building. The risk is that the time never gets protected and you arrive at month eight with a half-trained person and a client who has now asked twice.

Hire a specialist. Semrush's listing data puts the median advertised salary at $130,000 for senior SEO roles and $71,630 for everything else. Indeed's own aggregate, drawn from 890 salaries and updated July 17, 2026, puts the average SEO specialist base at $79,063. Call it $18,000 to $33,000 per quarter in base pay before load. That is fixed cost from day one against revenue that does not exist yet.

White label. Fully variable cost, fastest to a first sale, and the only branch that produces revenue in month one. The risk is that you have outsourced the one part of the service you cannot inspect from the outside. Choosing and vetting that partner is its own problem and it is the one buyers get wrong most often.

The break point between hiring and renting is account volume, not enthusiasm. Below roughly eight accounts that would genuinely buy the service, a dedicated hire idles and the fixed cost eats whatever margin the line was supposed to create. That threshold is my own operating number rather than a published benchmark, so run it against your own delivery hours before trusting it.

CMOs are being asked to deliver growth, efficiency and transformation without meaningful budget expansion.

That is your buyer. They are not funding a discovery project. Which argues for the branch that produces a defensible deliverable soonest, even at a worse margin, rather than the branch that produces the best long-run economics on a timeline your pipeline cannot survive.

There is a fourth option nobody names: sell one pilot at cost, deliver it yourself, and decide afterwards. It is slow and it does not scale, and it is the only way to learn what the work costs before you commit payroll to it.

Five interview questions that separate the real ones from the fluent ones

Fluency here is cheap. Everyone has read the same twenty posts. These five are hard to fake.

Question three is the one that matters. If a candidate cannot produce a real example, either they have never tested anything or they have never admitted it, and both disqualify them from this particular role.

Question four separates people who have looked at the data from people who have read about it. Semrush and Growth Memo logged 3,981 domain appearances across 115 prompts in 14 countries and found 61.7% were ghost citations, where a page was used as a source but the brand name never appeared in the answer text. The gap between being cited and being recommended is the whole reporting problem, and a candidate who conflates the two will build you a dashboard that overstates results by roughly two and a half times.

Question five has a documented right answer. Google's own guidance tells businesses that no one can guarantee a number one ranking and to walk away from anyone who promises it. The same discipline transfers directly to AI placement. A candidate who promises citations in a first meeting will promise them to your clients too.

Three starting states, three honest answers

You already have a technical SEO who reads logs. Train them. Ninety days to supervised delivery, about two quarters to unsupervised, and the binding constraint is whether you can defend five non-billable hours a week against the P&L. Cheapest path, and the one most agencies claim to be on while quietly not protecting the time.

You have content and account people, no technical bench, and fewer than eight accounts that would buy. Rent it. A dedicated hire at $18,000 to $33,000 a quarter against three interested accounts is a bet on a pipeline you do not have yet. Keep strategy and the client relationship in house, rent the mechanics, and accept the margin hit for two quarters while you find out whether the demand is real or polite.

You have volume and no capability. Hire, but hire the profile rather than the title. Your best candidate probably does not have GEO anywhere on their resume. They will have a technical search background, a habit of testing things, and at least one story about being wrong in public.

Whatever branch you land on, the deliverable standard is identical, and it is the part clients can actually judge. Get the reporting and the attribution story right and a staffing mistake stays recoverable. Get them wrong and it does not matter who is doing the work.

Jono Alderson framed the underlying point better than I can:

SEO, GEO, relevance engineering and advertising each influence part of that process. They are instruments, not the organising principle.

Which is the argument against hiring a GEO person at all. Hire someone who can run the method end to end and hand them the instruments. The rest of the operator economics series lives here, and if you want a second opinion on which branch you are genuinely on, book a call and bring your account count.