Jeff Brokaw · Chief AI and Marketing Officer · Published · Updated · Agency AI series, part 5 of 5
Grand Theft Agency VI is the 2026 edition. 76 agencies. Same job. Same getaway car.
Ninety-six agencies in my review advertise exactly one AI service. Seventy-six advertise the same one. AI search. AEO. GEO. Whatever the page calls it this month. Across all 217 agencies selling AI, 163 sell AI search. That's three out of four, selling the same thing.
That's not focus. That's an SEO shop adding one page to the menu and calling it a practice. (My read, not a measured finding.)
Then I went looking for the loot. Of the 217 agencies in my review that advertise an AI service, 128 had no public client example I could confirm. That's 59%. A steakhouse handing you a picture of a cow.
That doesn't prove they can't cook. But a buyer with nothing to check buys the software and skips you.
You don't want to be in that lineup. Here's the 2027 update.
What should an agency sell with AI in 2027?
Five things. Call them the 2027 patch notes.
- The fix, not the audit. Your own people correcting what the report found.
- The guardrail. A named person watching the AI's output after launch.
- The proof. A case with a starting number, an ending number and dates.
- A price that matches the work. The fee has to pay the people doing it in month three.
- A clear owner. A contract that says who holds what.
That's my judgment, not a measured result. The numbers come from my October 2026 review of 350 selected agency brands. The roster isn't random, and it leans toward digital, SEO and paid-media businesses. Counts describe what agencies advertise and publish, not what they can deliver. The public research summary has the method and the limits.
1. Sell the fix, not the audit
The number: 163 of 350 brands (46.6%) advertise AI search, AEO or GEO. Seventy-six advertise that and nothing else.
When nearly half the roster sells the same report, the report is a commodity. Your client can already buy a tool that shows where their brand appears in AI answers. Nobody can buy your people fixing what it finds. Re:signal lists a fixed AI discoverability audit at $15,000, one time, over two to three weeks. That's a real product with an end date. Then what? The client holds a report and a to-do list. Congratulations, you sold a very expensive grocery list.
I said in the first article that more buyers will ask the agency that finds the problem to help finish the work. I'm holding that.
My prediction: stand-alone AI-search audits get squeezed. The survivors bundle the audit with fixes their own people make. See what agencies advertise and what clients can buy.
2. Sell the guardrail
The gap: only 60 of 350 brands (17.1%) advertise operating, monitoring or governance support. The other 83% advertise nobody to watch the AI after it ships. Only 39 brands (11.1%) advertise both client-facing AI systems and that support.
That's the thinnest category in my data. It's also the one I expect buyers to ask about right after the first AI mistake lands in front of a customer. Somebody has to answer for the output, and a client will pay for that somebody. "The AI did it" is not an incident report.
The work is specific: a named reviewer who can change or stop the next action, a record of what the system did, and a plan for the day a supplier changes something underneath you. My pieces on the human gate and the AI audit log show what that involves. Client AI systems and support covers the 39.
My bet: governance becomes the premium line on an agency's menu.
3. Sell the proof
The receipts: 128 of 217 agencies advertising AI (59%) had no public client example I could confirm. In 84 cases I finished the review and found none. In 44 the evidence stayed unresolved.
The narrow shops were worst. Seven in ten agencies advertising a single AI service gave a buyer nothing to check. The share fell as the menu grew.
Scroll the table sideways to see all columns.
| Categories advertised | Brands | No public client AI example I could confirm | Share |
|---|---|---|---|
| One | 96 | 67 | 70% |
| Two or three | 53 | 31 | 58% |
| Four or five | 47 | 24 | 51% |
| Six | 21 | 6 | 29% |
A "public client example" means a page that connects a client, an AI action and the agency's contribution. It isn't a verified deployment, and a missing example doesn't mean missing work. Plenty of agency work is confidential or never gets written up. Bigger agencies also publish more pages, so a wider menu and a bigger archive probably travel together. Looking only at the 132 brands whose review route finished, the order holds: 74%, 63%, 60% and 42% with no example. The six-service group is only 21 brands, 12 in that subset. Treat it as a pattern worth watching.
And the long-menu crowd isn't the model. Six of those 21 had no example I could confirm. A long menu with nothing behind it is a longer menu.
Then there's the proof that does exist. Of 24 delivery case studies I reviewed, 13 reported a measured client outcome and only two printed a starting and ending number for the same metric. That's 8%. Here's how I'd structure a case study that clears the bar.
My prediction: the number beats the logo wall. The first agency in a category to print baselines gets the meeting.
4. Sell a price that matches the work
The prices: the same content tool, three prices. I converted these from pounds at about $1.33 to the pound, so read them as rough.
ROASTWriter lists three packages: self-managed access at about $3,500 a month plus about $6,700 setup, then from about $11,700 and from about $16,400 a month as content support and then full-service SEO work get added. Reboot says its ongoing GEO program starts at about $3,300 a month, with the final price depending on goals, complexity and how much of its team is involved.
Those are three different purchases. I won't average them, and neither should you. What I notice is that price climbs with the human work. Nobody pays $16,400 a month for software. They pay for the strategy, the editing and the technical work wrapped around it.
That makes every service you add a staffing decision. Nobody ever renewed a retainer because of a damn menu. They renew because someone does the work every month, and your price has to pay that person.
On the electrical-distribution quoting product where I was Chief AI Officer, a quick product lookup and a heavy 200-line bill of materials cost very different amounts to serve. So I built the pricing around a platform fee, seats, included usage and heavier job classes. The product is prelaunch, so I can't tell you how that pricing performs with customers. I can tell you I wouldn't have set it without knowing what the heavy jobs cost. My cost-per-customer article shows how to meter that, and the case study has the product details.
My bet: retainers priced on hours get renegotiated. When a 40-hour task turns into a 10-hour task, somebody has to decide what replaces the revenue. If you don't decide, your client will. The contract renewal check is where I'd start.
5. Sell a clear owner
The check: I ran seven questions about data, outputs, code, models, control, exit and maintenance across 128 named products and offerings. That's 896 checks. Public material left 568 unanswered, or 63.4%.
Those are gaps in public evidence, not proof that clients lack those rights. Still, a buyer who can't tell what happens to their data, who keeps the code or what happens when the supplier changes the terms will assume the worst. Say it plainly in the proposal: what the client owns, what stays licensed, who holds the account and who fixes it when something upstream breaks.
My prediction: the agency that answers the ownership questions before procurement asks wins the deal. Procurement will ask. Get there first. See who owns the AI for the details.
What would I stop selling?
A wide menu as proof. A page date as a credential. A partner logo as a deliverable.
I checked an archived page to make the point. Imaginuity's services page, saved by the Internet Archive on April 3, 2019, listed an Artificial Intelligence service and named machine learning, natural language understanding and speech recognition. I read it on November 19, 2026. My archive pass covered 208 saved pages from 57 of the 350 brands. A saved page proves what an agency said on a day. It doesn't prove a model shipped, when the work started or whether it was any good. Early isn't the same as good, and a page isn't an adoption date.
Should an agency narrow, broaden or bring in a specialist?
Use the proof you already have, not the preference you started with. These are my rules of thumb. I didn't measure any of the thresholds.
Scroll the table sideways to see all columns.
| Move | I'd do it when | I'd want to see first |
|---|---|---|
| Expand a service you sell | A client has paid for it more than once and can describe the result | A case with starting and ending values, dates and what else changed |
| Test an adjacent job | Existing clients already ask for it and you can name who accepts the work | One paid engagement with an acceptance check and a delivery-cost limit |
| Bring in a specialist | The job needs maintenance or skills you can't carry, such as a client-operated application | A contract that says who holds the account and who fixes it when a supplier changes something |
| Narrow | You can't show a client result or price the support after real delivery | An honest count of hours spent against the fee collected |
The build-versus-buy guide covers the second and third rows.
What I'd run before the next proposal goes out
Spend one meeting on this. For every AI service on your site, write one line each for:
- The client problem it solves
- The person who accepts the work
- The case a buyer can check
- What the price covers after launch
- The outside software it depends on
A service with a blank line comes off the homepage or gets labeled a pilot. I'd rather lose a menu item than sell something nobody owns. And I'd fund a published case before I funded a new service page: record the baseline before the next engagement starts, agree the acceptance check with the client and print the result with its dates.
What I'll recheck in October 2027
I'll rerun the same review on the same 350 brands, with the same six categories and the same rules. Changes in the original roster, unavailable pages and added brands get reported separately, so a bigger sample can't pass for a changed market. Three bets, with thresholds I set before looking:
- Menus widen. At least 19 of the 76 AI-search-only advertisers, one in four, will have added a second category. If fewer do, my read that these shops are starting points and not commitments was wrong.
- Menus outrun proof. More brands will add a category than will move from "no example found" or "unresolved" to a documented client example. I'll recheck the 128 brands in those two states with the same route as the first pass.
- The proof gap holds. The one-service group will still show a higher share with no confirmed public client example than the group advertising four or more. Today it's 70% against 44%. If that gap closes, I'll say the breadth pattern was an artifact of who publishes case pages.
None of this says anything about profit. Margin, retention and demand live in an agency's own books, and I can't see yours. A public-page recheck can tell you what agencies say. It can't tell you what they earn.
The full record-level report stays private. I keep the counts, denominators and dates in each article so you can follow what I counted and what I didn't.
A menu is free to print. Before you add another line to yours, show me who's in the kitchen on a Tuesday in month three.
Frequently asked questions
What should an agency sell with AI in 2027?
The fix, the guardrail, the proof, a price that matches the work and a clear owner. In my review, only 17.1% of brands advertised monitoring or governance support, and 59% of agencies advertising AI had no public client example I could confirm. That's where the gap is.
Should an agency focus on one AI service in 2027?
Not because focus sounds smart. Focus on one service when you can name the client problem, the person who accepts the work, a price that covers the continuing work and a case a buyer can check. Without those four, one service is a narrow menu, and a broad one is a long menu.
What does "44% advertise one AI category" mean?
Of the 217 agency brands in my October 2026 review that advertised at least one of six AI-related categories, 96 advertised exactly one. That's 44.2%. Seventy-six of the 96 advertised AI search and nothing else. It describes public offers from a selected roster, not internal focus, quality or profit.
Does advertising more AI services mean an agency is better?
No. In my data, agencies with wider menus were more likely to show a public client example, but a public example isn't a verified result, and larger agencies publish more pages. Six of the 21 agencies advertising all six categories had no example I could confirm.
Is AI search enough of an offer on its own?
I don't think it will hold up as a stand-alone line. Of the 163 brands advertising AI search, 76 advertised nothing else, and buyers can already buy tools that track AI answers. The review didn't measure demand or margin, so this is my prediction. I expect the durable version to include fixes the agency's own people make.
How should an agency price an AI service?
Price the work you'll still be doing after launch: review, exceptions, supplier changes and support. The public prices I checked ranged from a one-time $15,000 audit to monthly programs from about $3,300 to about $16,400, and they price different work. I wouldn't average them.
When should an agency bring in a specialist instead of building?
When the job needs maintenance or skills your team can't carry, such as a system the client operates. Get the contract to say who holds the account, who pays the supplier and who fixes it when something upstream changes.
Method and sources
The October 15 to 20, 2026 public-evidence review covered 350 selected agency brand and domain records, six overlapping AI-related offer categories and a separate client-work check on the 217 that advertised at least one. The roster isn't random or representative. Counts describe public evidence, not verified delivery, margin or demand. The 128-of-217 count (84 no example found, 44 unresolved) is on the public research page. The cross-tabulation of menu width against public client examples is my own cut of the review records for this article. See the public research method.
Pages checked November 19, 2026. Pound prices were converted to dollars at about $1.33 to the pound (Bank of England rate, October 19, 2026) and rounded. Actual US cost depends on the exchange rate and the buyer's terms. They are not quotes.
- Reboot: GEO services. Published starting price; final price varies.
- Re:signal: AI search approach. Fixed-fee audit; the dollar price is as published.
- ROASTWriter. Three published packages.
- Imaginuity services page, Internet Archive capture of April 3, 2019. Shows what the page said on that date.
Agency AI series, part 5 of 5. Earlier articles cover the AI services agencies advertise, client AI systems and support, who owns the AI and what agency AI case studies prove.