SKAGs for OpenAI Ads: single-intent ad groups

What a SKAG really was, underneath the tactic

A single keyword ad group, or SKAG, did exactly what the name says. Instead of putting a dozen related keywords in one ad group, you gave each keyword its own ad group, usually with the keyword repeated in a couple of match types and an ad written to mirror it word for word. At its peak, a serious Google account might have run thousands of them.

The clever part was never the keyword count. It was isolation. When an ad group holds one keyword, that ad group's spend, clicks, and conversions are that keyword's spend, clicks, and conversions, with nothing else mixed in. You could look at a single line in your account and know, precisely, what one search query was doing for your business. You could write an ad that matched that query so closely it lifted your click-through rate and, through Quality Score, lowered what you paid. And you could set a bid for that one query without dragging twenty others along with it.

Strip away the Google-specific machinery and what is left is a simple, durable idea: isolate one unit of demand per ad group so its numbers tell you the truth about that unit. That idea does not depend on keywords, or on Quality Score, or on Google at all. It depends only on wanting clean attribution and direct control, which is to say it depends on the thing every advertiser wants. Hold onto that, because it is the whole reason SKAGs are about to matter again.

Why SKAGs fell out of favour in Google Ads

If the principle is so sound, why did the industry walk away from it? Three forces, and all three were specific to Google rather than to the idea itself.

The first was close variants. Through 2018 and 2019, Google steadily widened what an exact or phrase match keyword would trigger on, folding in synonyms, reordered words, and queries it judged to share your intent, with no way to opt out. The day exact match stopped meaning exact was the day a single keyword stopped mapping to a single query. The isolation that made SKAGs work leaked away, because the platform was now matching your ad to things you never typed.

The second was automated bidding. Smart Bidding strategies like Target CPA and Target ROAS are hungry for data, and they learn best when a lot of it pools in one place. Spread that same volume across a thousand thin ad groups and each one starves, so the algorithm cannot find a pattern. Google's own guidance shifted toward fewer, broader, intent-based ad groups for exactly this reason, and the industry followed it into single theme ad groups, where you group by intent and let the machine sort the queries.

The third was simply the labour. Building and maintaining thousands of near-identical ad groups by hand, writing the ads, adding the negatives that stopped them stealing each other's traffic, keeping it all tidy, was brutal, ongoing work. Even advocates admitted the overhead was the weakest part of the case. By around 2020 the consensus had settled: SKAGs were a relic of a more manual, more transparent era of search.

Why an opaque platform flips the logic back

Here is the part worth slowing down for. Read those three reasons again with OpenAI Ads in mind, and every one of them either disappears or reverses.

Close variants cannot blur your isolation, because there are no keywords to blur. OpenAI matches ads to the meaning of a conversation, so the leak that undid SKAGs in Google has no equivalent here. There is nothing to opt out of and nothing quietly expanding underneath you.

Automated bidding cannot punish you for fragmenting data, because there is no rich, signal-fed bidding engine to feed. The platform gives you aggregate clicks and little else. Pooling everything into a few broad campaigns does not buy you a smarter algorithm the way it does in Google. It just averages your results together and hides which part of the campaign actually worked. On a platform this opaque, consolidation does not buy intelligence. It buys ignorance.

And the labour, the one objection that was genuinely fair, is the one that technology has quietly removed. The reason nobody wanted to hand-build ten thousand ad groups is that a human had to do it. An assistant connected to your account does not mind. It will create the structure, write an ad for each intent, and run the weekly upkeep without complaint. The single strongest argument against SKAGs was an argument about human effort, and that argument expired the moment an AI could do the work.

So the scorecard inverts. The forces that retired SKAGs in Google are absent or reversed on OpenAI Ads, and the thing they were always good at, clean attribution and direct control on a platform that hides both, is precisely what OpenAI's opacity demands. You cannot isolate a keyword, because there are none. So you isolate the next best unit: intent. One topic or use case per ad group, in one place, with an ad written for it. Call them single-intent ad groups. They are the SKAG, rebuilt for a platform with no keywords, and they may be the single most important structural decision you make on OpenAI Ads.

What the grid looks like in practice

Suppose you sell a project management tool. Your intents are not one fuzzy idea of project management software. They are a set of distinct demands: a gantt chart tool, a task tracker for small teams, a Trello alternative, software for agency resource planning, a simple to-do app for freelancers, and so on. Each of those is a different person with a different problem, and lumping them together would average a high-intent buyer searching for resource planning in with a freelancer who wants a free checklist.

Say you land on forty such intents and you sell in three markets, the United States, the United Kingdom, and Canada. Forty intents times three countries is a hundred and twenty ad groups, each one a clean little experiment. After a few weeks you will not be staring at one blended cost per acquisition for the whole account. You will see that the gantt chart intent pays beautifully in the United States and poorly in Canada, that the freelance to-do intent burns money everywhere, and that agency resource planning is your quiet star. None of that would be visible in a consolidated campaign. The structure is what makes it visible.

A hundred and twenty is a modest example. Add states or regions, or a longer list of intents, and you are quickly into the high hundreds or thousands. That scale is exactly why this was impractical to do by hand and exactly why it is practical now.

Where this goes wrong, and how to keep it honest

The strategy has one real failure mode, and it is the mirror image of its strength. Narrow ad groups split your traffic into small pieces, and a piece that is too small never gathers enough data to judge. An ad group with four clicks in a week is not telling you anything, no matter how cleanly it is isolated. Push the grid too fine and you end up with a thousand ad groups that each say nothing.

So treat granularity as a dial, not a switch. Split as far as your volume can support a readable signal, and no further. When an intent simply never accumulates enough traffic to evaluate, fold it back in with a couple of its closest neighbours and judge the combined group instead. Lengthen your measurement window for the thinner ad groups rather than reacting to a handful of clicks. And hold the line on the one rule that makes the whole thing work: one intent, one place, one message per ad group. The moment you let two intents share a group to save effort, you have given up the only thing the structure was buying you.

It is also worth being honest about what this is and is not. OpenAI Ads matches on conversation context, so single-intent ad groups are not literal keyword targeting and never can be. What they are is structural isolation, an arrangement of your account that forces the platform's aggregate reporting to resolve down to one readable unit at a time. That is a humbler claim than keyword-level control, and it is also the most control the platform allows you to have.