Keyword Match Types in the AI Matching Era

Keyword match types are the setting that tells Google how far it may roam from your keyword when deciding which searches trigger your ads. There are three — exact, phrase and broad — and the first thing to unlearn is that any of them is literal. Since Google folded language models into matching, all three interpret meaning: exact stopped meaning identical years ago, and broad stopped meaning reckless more recently than its reputation admits. What the types really set is a radius of interpretation.

How the three match types work

The working definitions, as they behave today:

Match type	Notation	What may trigger it
Exact	[running shoes]	Searches with the same meaning or intent as the keyword
Phrase	"running shoes"	Searches that include the meaning of the keyword
Broad	running shoes	Searches related to the keyword, judged on every available signal

"Same meaning" in practice: reorderings, function words added or dropped, synonyms and close paraphrases, implied intent. Phrase's "includes the meaning" allows the query to extend beyond the keyword — best running shoes for flat feet — provided your keyword's meaning stays inside it. Broad reads intent from more than the keyword itself: Google also weighs your landing pages and the other keywords in the ad group, and with Smart Bidding it evaluates per-auction context, reaching related searches that never mention your words at all.

The control spectrum runs from exact to AI Max

Treat matching control as a dial rather than three unrelated products. Exact concedes the least radius. Phrase trades reach for containment. Broad hands the models the widest brief and repays it mainly when paired with conversion data rich enough to steer — the case examined honestly in broad match in 2026. Past broad sits AI Max for Search, the opt-in layer that lets Google expand beyond your keyword list altogether. Each step outward exchanges your specification for Google's inference — a legitimate choice whenever it is made on purpose, with the data to supervise it.

Choose by situation, not ideology. Thin data and tight risk tolerance argue for the narrow end. Strong conversion volume, clean tracking and an active negative-keyword practice make the wide end rational. The one indefensible position is wide matching plus nobody reading what matched.

Auditing match behaviour instead of arguing it

Match types used to be argued in the abstract. With account access, they are audited empirically. The evidence is the search terms report — what you actually paid for, keyword by keyword — and an agent runs that audit continuously, not once a year:

• "For each broad match keyword, split spend between search terms that contain the keyword's meaning and terms that merely relate to it — conversions for both."
• "Which exact match keywords matched queries whose intent differs from the keyword? Worst mismatches first."
• "Draft the negative keywords that would have blocked the non-converting drift, and stage them for approval."

That last step is the loop closing: the same conversation that finds the drift can fence it, as approvable changes with the evidence attached. Match types in 2026 are less a setting you pick once than a boundary you patrol — and now your agent patrols it for you, every change logged.