On a call a few weeks ago, an agency owner I coach was deciding where next quarter’s money should go. Meta had stopped producing for one of his accounts and he was sorting through replacements out loud. What he landed on was this: “If AEO is working and gives us predictability, great, and we could supplement with an SDR.” Then, a beat later, “I just need to have a few irons in the fire at all times.”
That second sentence is the buying question underneath every AI search pitch being made right now, and it has almost nothing to do with AI. He wasn’t asking whether large language models are the future. He was asking whether this line item produces something he can count on.
Your agency should sell AI search optimization when you can measure whether it worked. The demand is arriving much faster than the evidence, so the honest product right now is measurement and content quality, not a promised placement in an answer.
Quick Take
- The demand is real, and it’s showing up in a shape most agencies aren’t set up to catch. On my own site, roughly a quarter of last quarter’s search queries were full sentences rather than keywords, several carrying visible prompt scaffolding.
- The best available research says most of the published tactics don’t work. C-SEO Bench tested fifteen recommended methods and found them “largely ineffective” and often negative, while Google’s own guidance is to do normal SEO and write unique content.
- So the sellable deliverable isn’t a ranking. It’s a baseline, a monthly re-measurement, and content the model has a reason to quote.
- Price it as research, not as a channel, until you have three months of your own data.
In this article
- Is there actual demand for this?
- Does AI search optimization work?
- What you can honestly promise a client
- How to scope and price it
- When to say no
Is there actual demand for AI search optimization?
Yes, and the clearest evidence isn’t a market-size projection, it’s what shows up in your own Search Console. When I pulled ninety days of queries for my site recently, 117 of 474 were seven words or longer. Several arrived with the scaffolding still attached, including one that opened “context: location: united states (not for language). do not include location references in your response. question: i run a small marketing agency and i’m thinking about joining an agency coaching program.” That is not a person typing. That’s an assistant passing a question through, and it landed on my pages.
One query in that set was a buyer describing the problem in their own words: “my competitors are showing up in perplexity answers but i’m not. what steps can i take to start showing up there too?”
The demand also shows up in sales calls, which is where it gets expensive to ignore. A manufacturer’s marketing lead was asked on a discovery call how he’d found the agency he was talking to, and he said, “Google Chat GPT, to be honest.” The agency owner on the other end had suspected it was happening and had no way to confirm it until a stranger said so out loud.
So the interest is real and your clients are going to raise it. The question is what you say next.
Does AI search optimization actually work?
The most useful answer is that nobody selling it can currently prove the mechanism, and the strongest study we have says most of the published playbook underperforms.
C-SEO Bench, published at NeurIPS in 2025, tested fifteen recommended conversational-SEO methods across question-answering and product-recommendation tasks. Its finding was blunt: most of the methods were “not only largely ineffective but also frequently have a negative impact.” What did move results was improving the source’s relevance and ranking. The gains also proved zero-sum as more people adopted them, which is the part that should worry anyone building a productized offer on top.
Google published guidance in May 2026 saying much the same thing, that the work is ordinary SEO plus unique content, and that the tactics circulating in the optimization industry aren’t needed.
Set against that, my own measurement says something more encouraging and more specific. Two of my pages are cited in Google AI Overviews. Both sit at citation rank one inside the answer while ranking 24th and 11th organically. Being cited and ranking well turned out to be different games, and the pages winning citations were the ones explaining something plainly, not the ones built to be optimized.
That gap is the actual opportunity, and it’s also why the offer has to be scoped carefully.
What you can honestly promise a client
You can promise a baseline, a cadence, and content that gives a model something worth lifting. You can’t promise a placement, and any agency promising one is selling a coin flip with a retainer attached.
A defensible scope looks like this. Measure where the client currently appears when a buyer asks the questions that matter, using a fixed set of prompts you re-run identically every month. Fix the things that are broken and verifiable, like a site the model can’t crawl or an entity it has confused with someone else. Then publish content that answers the buyer’s real questions in language the buyer used, and watch what gets quoted.
The confusion problem is more common than people expect and it’s satisfying to fix. For months, one assistant described me as a music publisher in Nashville, which is a different person with my name. Everything else I published was landing against the wrong identity. That’s a fixable, provable piece of work, and it’s worth more to a client than a vague promise about visibility.
How to scope and price it
Price the first engagement as research, because that’s what it honestly is, and because it protects the relationship when month two produces a flat line.
Sell a paid baseline as a standalone piece of work. Run the prompt set, record who gets named, record who gets cited and from which page, and hand the client the actual answers with the sources visible. Most clients have never seen this and the document itself is the value. Then, if they continue, the monthly work is a re-run of the identical prompts plus the content and technical fixes the baseline surfaced.
Two scoping rules keep this from turning into an unbounded retainer. Fix the prompt set in writing before you start, because a moving prompt set produces a moving scoreboard and no comparable data. And write the report so it can say nothing changed, since a monthly report that always finds progress is a report nobody believes by month four.
If you’re an agency owner working out whether this belongs in your lineup, that’s the kind of offer question we work through inside the Dynamic Agency Community.
When to say no
Say no when the client’s actual constraint is somewhere else, which it usually is. An agency I worked with spent months producing content that generated nothing because a noindex tag was still live across the whole site. No amount of AI search work would have touched that, and selling it would have been taking money to work on the wrong layer.
Say no when the client wants a guarantee. You cannot give one and the person who does will get the account, and then lose it in six months with your category’s reputation attached.
And say no when you’d be adding it because competitors list it. That’s the same reasoning that put “we do better work” on a thousand agency homepages, and buyers discount it to zero for the same reason.
The version of this worth selling is narrower and more honest than the pitch going around: find out where the client stands, fix what’s provably broken, publish things worth quoting, and measure it the same way every month. That’s a real service. It just isn’t a magic one, and the agencies who say so out loud are going to keep the clients longer than the ones who don’t.
FAQ
Is AEO the same thing as GEO?
Mostly, yes. Answer engine optimization, generative engine optimization, AI search optimization, LLM SEO: the industry hasn’t settled on a name and the acronyms get used interchangeably. Don’t spend a sales call on the vocabulary. Buyers don’t search for the acronyms anyway, they describe the problem, which is why the query that reached my site read “my competitors are showing up in perplexity answers but i’m not.”
Can I charge a retainer for this yet?
You can, but scope the first month as a paid baseline and let the retainer start after the client has seen real data about their own position. It sets an honest expectation and it gives you something to compare month two against.
What if a client asks for a guarantee?
Tell them the truth, which is that nobody controls what a model cites and that the best available research says most of the tactics being sold don’t reliably move it. Then offer the thing you can guarantee: a fixed prompt set, re-measured monthly, with the results reported whether or not they improved.
Do I need to rank on page one to get cited?
No, and this surprised me when I measured my own site. Both of my AI Overview citations sit at rank one in the answer while the pages behind them rank 24th and 11th in normal search. Citation and ranking move somewhat independently, which means a page that hasn’t cracked the top ten can still be the source the answer is built from.
Should I build this as its own service line or fold it into SEO?
Fold it in first. It shares most of its work with content and technical SEO, and standing it up as a separate service line before you have your own results forces you to market a promise you can’t yet back. Once you’ve got three months of measured client data, it can carry its own page.