We Built Our ChatGPT Ad Targeting From Our Own Share-of-Model Data

By Stephen Honight, Founder of Lmo7

We launched a ChatGPT ad account for Lmo7 a few weeks ago. The interesting part was not the ad. It was where the targeting came from. We did not open a keyword planner. We used our own Share-of-Model data, the same data we run for clients, and let it tell us which conversations to pay to appear in.

That sounds like a small process choice. My sense is it is the whole game. If you build a ChatGPT ad account the way you built a Google Ads account, you will get the setup wrong and you will not understand why the numbers look strange. The mechanism is different, so the input has to be different too.

Let me walk through what we actually did, what came back, and the bit that was genuinely humbling for an AI search agency to see about its own brand.

Context hints are not keywords

Start with the mechanism, because it changes everything downstream.

On Google, you buy keywords. You bid on a string, someone types that string, your ad has a chance to show. The mental model is match a query to a term.

ChatGPT ads do not work like that. The targeting input is a context hint, and a context hint is free text. You describe the situation or the person you want to appear for in plain language, not as a keyword you are bidding against. The system reads that description and decides when the live conversation is close enough to serve your ad.

This trips people up. Most performance teams arrive with the Google model in their head, write a tight list of head terms and wonder why the account behaves oddly. The platform is closer to describing an audience and an intent than buying a slot against a phrase. It reads more like a brief than a bid.

So the question stops being “what keywords do I want” and becomes “what conversation is my buyer actually having, and how do I describe it well enough that the model puts me in the room”. That is a different input. And it happens to be exactly what a Share-of-Model analysis produces.

What Share-of-Model gives you that a keyword tool does not

Share-of-Model is our way of measuring how a brand shows up when people ask AI assistants for help. We track the real prompts a buyer would use, across ChatGPT, Gemini and the other engines, and we look at whether the brand gets mentioned, where it sits, what gets said about it and which sources the model leans on.

Run that for your own buyer and you are holding two things at once. You have the prompts and the intents your audience actually uses. And you have the honest picture of whether you show up in them for free.

That second half is the point most teams miss. One Share-of-Model pull tells you what to fix to get recommended organically, which is the long game, and it tells you exactly which conversations you are not winning today and could pay to appear in, which is the quick win. Same data, two roads out. The organic gaps are your paid targeting brief, handed to you.

So we pointed the analysis at ourselves.

The buyer we described

We defined the buyer plainly. An ecommerce director at a £1m to £10m consumer brand, looking for an agency to improve their Amazon performance and their visibility in AI-led shopping. That is our Challenger buyer more or less exactly.

Then we ran the prompts that person would realistically use. Things in the shape of “which agency should I use to improve my Amazon presence”, “who helps brands show up in AI search”, “best AEO agency for a consumer brand”, and the softer research prompts that sit earlier in that journey.

From those results we wrote the ad targeting as three layers of context hints, going from broad to sharp.

The top layer was the wide audience. Ecommerce and marketing leaders at consumer brands who are starting to think about AI and Amazon. Lots of reach, low intent.

The middle layer narrowed it. People actively comparing how to improve Amazon and AI search visibility, weighing whether to hire in or bring in help.

The bottom layer was the sharp slice. Founders and ecommerce leads comparing agencies for Amazon and AI-search work and asking what it costs. That is the ready-to-buy conversation. Small, but the one that matters.

Narrowing in layers matters because a context hint that is too broad wastes budget on people who are curious but nowhere near a decision, and a hint that is too narrow never gets enough conversations to serve into. You want a wide net and a sharp net running at the same time so you can see which one earns its keep.

The creative itself was deliberately simple. The Lmo7 Agency. Learn about agentic commerce today. A link to the site. The work was in the targeting, not the headline.

The humbling part

Here is where it got uncomfortable, and I think it is the most useful thing in this whole piece.

When we ran our own buyer’s prompts and asked the models to recommend an agency for Amazon and AI-search work, we mostly did not show up. The names that came back were Perpetua, Pacvue, Tinuiti and Jungle Scout. Bigger, older, more cited. Where Lmo7 did get named, the word that came back attached to us was “unproven”.

That stings a bit when you run an AI search agency. It is also completely fair, and it is a clean demonstration of the thing we tell clients constantly.

The reason we do not show up is not that our content is bad. Our own site scores well on AI readability. The reason is domain authority. We are a young brand with a low domain rating and a thin citation footprint, so the models have little third-party evidence to lean on when they build an answer. Good content with weak authority reads clearly and still gets left out. This is the persistent bottleneck in AI citation, and seeing it happen to our own brand is a sharper lesson than reading it in a slide.

Which is the honest case for paid in the first place. Organic AI recommendation is a slow build. It needs authority, citations, mentions in the places the models trust, and that takes months. Paid lets you appear in the exact conversation today while the long game compounds underneath. You are not choosing between them. You are buying presence now and earning it over time.

The early signal, and why I am not calling it a result

A few days after the account went live, someone found Lmo7 through ChatGPT and reached out to us directly.

I want to be careful here. That is one anecdote. It is not a conversion rate, it is not a return figure, and I am not going to dress it up as proof the account works. Early ad accounts throw off noisy data and it is far too soon to read anything into the numbers. What the enquiry did do was confirm the surface is real. People are asking AI assistants who to hire and acting on what comes back. That behaviour exists now, not in some future state.

The wider pattern I am watching, from our own account and from the client work, is that the sharpest targeting is not brand terms. It is the adjacent conversation. The person is not searching for you by name. They are describing a problem, and the highest-signal play is to be present in the problem, not the brand query. Describe the moment your product solves, not your product. That is where context hints earn their difference from keywords, and it is where most accounts are currently underbuilt.

What this means for your account

If you are setting up ChatGPT ads, or thinking about it, the takeaway is not really about ChatGPT ads. It is about where the targeting should come from.

Do not start from a keyword list. Start from a read of how AI assistants already talk about your category and your brand. That read gives you your paid targeting and your organic to-do list in one pass. Where you already show up for free, you may not need to pay. Where you are absent or described weakly, that is both the ad you should run and the content and authority gap you should fix.

The mistake I would most want a founder to avoid is treating paid as a substitute for the organic work. It is not. If the models call you unproven, ads will put you in the conversation but the moment the buyer digs, the same weak authority is still there. Paid buys the door. The long game decides whether you get invited back in.

So what should you do next

My recommendation depends on where you are, so here is the honest routing.

If you want to run this yourself and you mainly need the data, start with the data layer. Our DaaS tier gives you direct access to the Share-of-Model and channel data, plus an onboarding session to read it, from £250 plus VAT a month. You get the same visibility pull we used on ourselves, and you build the targeting and the fix list from it. This is the right entry if you have the capacity to action it and you want to learn the discipline by doing it.

If you want us to run the ads and the tracking, that is Challenger. The Agentic Stack covers AI search visibility tracking and content optimisation alongside the ad placements in AI-mediated environments, on a modular monthly basis with no lock-in. This is the fit if you would rather we built the account, read the data and ran the test-and-learn loop while your team stays focused elsewhere.

If you are a multi-brand business and the real question is how AI describes your whole portfolio before you spend anything, that is Enterprise. We run multi-brand Share-of-Model tracking and a workshop to align your teams on what the data says and what to prioritise. Bigger brands should diagnose across the portfolio first, then decide where paid and content effort earns the most.

Wherever you land, the first move is the same. Get the visibility read before you write a single line of ad copy. The data you need to target well and the data you need to fix your organic gaps are the same data. You just have to look at your own brand as honestly as we had to look at ours.

Stephen Honight is the Founder of Lmo7, an AI-native agency helping consumer brands win in AI-powered discovery and agentic commerce across Amazon, ChatGPT, Gemini, Google AI and Amazon Rufus. Lmo7 works with brands including Trip Drinks, Veloforte, Brown-Forman, Haleon, Pelotan and Symprove.