What is Ad group?

An ad group is the container between a campaign and its ads on Google Ads and Microsoft Advertising. It holds a set of keywords, any negatives specific to them, and the ads that serve when those keywords match.

Because it is where keywords and ads meet, it is the level at which a search account's relevance is decided. The platform judges how well an ad matches the intent behind a query, and an ad can only be written against the queries its group actually serves.

That relationship is the whole of what structure at this level is for.

The useful question is not how many keywords a group should have. It is whether one ad can speak directly to every query the group serves.

A group containing several ways of asking for the same thing passes. An ad can repeat the language of those queries, answer the intent in the headline, and point at a page built for it. Relevance is high, expected click-through is high, rank is better and the clicks cost less.

A group containing thirty loosely related terms fails. The ad has to be generic enough to cover all of them, which means it speaks directly to none, and the relevance diagnostic will report it.

The failure is visible in the search-term report read against the ads: queries spanning intents that one ad cannot address are a group that needs splitting.

For years the standard advice was maximum granularity — one keyword per group, sometimes one keyword per group per match type — and it was correct under the conditions of the time.

Match types were literal, so a keyword defined its queries precisely and an ad could be written against that exact phrase. Bids were manual, so a group per keyword meant a bid per keyword. The fragmentation cost nothing because nothing depended on pooled data.

Both conditions changed. Semantic matching means a keyword no longer defines its queries precisely, so the ad cannot be written against one phrase with any confidence. Automated bidding needs conversion volume, and dividing an account into hundreds of groups leaves each with too little for the model to learn from.

The relevance benefit of tight grouping survives. The extreme version of it now costs more than it returns in most accounts.

Intent, not keyword count. Queries asking the same thing belong together regardless of how many there are.

Landing page. If two sets of queries should go to different pages, they are different groups, because the destination is part of the relevance.

Stage. A query researching a category and a query naming a product are different intents even within the same subject.

Brand and non-brand, though this usually deserves separation at the campaign level rather than here.

What not to split on: match type, device, small volume differences, or a desire to see the numbers separately. The first three fragment for no benefit and the fourth is a reporting need that segmentation can meet without structural cost.

Keywords and their match types, which define eligibility.

Negative keywords specific to the group, which are also how a query is routed away from a group that should not receive it — the only mechanism that actually forces that outcome.

The ads. Under assembled text formats, the asset library from which ads are built.

Audience attachments, where the account applies them here rather than at the campaign.

Bids, under manual bidding. Under an automated strategy these are computed per auction and there is nothing to set.

Budgets and the bid strategy itself are above, at the campaign, which is why a change at this level is far less disruptive than a change at that one.

For most search accounts the shape that holds up is unremarkable.

Campaigns separated by the things the platform cannot see — brand against non-brand, markets with different economics, budgets that are committed separately.

Ad groups inside them separated by intent, each tight enough that its ad repeats the language of its queries and its landing page answers them.

Negative keywords maintained from the search-term report, applied at the level that matches the decision.

Enough conversion volume in each campaign for the bidding to learn.

That is less clever than the highly segmented structures of a decade ago and it performs better, because the thing the segmentation was buying is now done per auction by a model that sees more than any structure can encode.

A search matching keywords in more than one group does not trigger all of them; the platform picks one.

An exact match on the query is generally preferred over looser matches elsewhere in the account, and beyond that the selection leans on relevance and expected performance rather than purely on bid or match type.

Two consequences follow. A group showing very little volume may not be losing auctions at all — it may be losing the internal selection to another group holding a more plausible candidate. And adding a keyword to the group you want a query to land in does not force the outcome; only a negative on the groups that should not receive it does.

That makes negatives the routing mechanism as well as the exclusion mechanism, which is a role they are rarely given credit for.

Inherited accounts built on the old advice are common, and the work of simplifying them has a sensible order.

Group by intent first. List the queries actually being served, from the search-term report, and cluster them by what they are asking for. The clusters are the groups the account should have.

Merge, do not delete. Moving keywords into fewer groups preserves their history where the platform allows it; deleting and recreating does not.

Write one ad per resulting theme, using the language of its queries.

Move negatives up or down to the level that matches the decision each encodes.

Then leave it alone long enough for the bidding to settle, because the consolidation is itself a material change.

They are the same level of the hierarchy under two names: Google Ads calls it an ad group, Meta calls it an ad set, and the object sits between the campaign and the ads. What separates them is what each platform puts at that level. An ad group holds keywords, which makes it a unit of intent: the queries it answers are what define it, and the whole of the craft here is keeping a group tight enough that one ad can repeat the language of every query it serves. That is why relevance, quality diagnostics and the search-term report are the instruments that matter on this side — each of them is a measure of how well the ad matches the intent the group was built around. None of that transfers to an ad set, which has no keywords and no intent to be relevant to. What does transfer is the principle underneath: one theme per container, enough conversion volume in each for the bidding to learn, and no fragmentation that buys control the auction has already taken over.